Cloud Hosting

In this section, we first describe how Ecological Momentary Assessments work and how they differentiate from assessments that are collected within a clinical environment. Second, we present the studies and ML use cases for each dataset. Next, we introduce the non-ML baseline heuristics and explain the ML preprocessing steps. Finally, we describe existing train-test-split approaches (cross-validation) and the splitting approaches at the user- and assessment levels.

Within this context, ecological means within the subjects natural environment", and momentary within this moment" and ideally, in real time16. Assessments collected in research or clinical environments may cause recall bias of the subjects answers and are not primarily designed to track changes in mood or behavior longitudinally. Ecological Momentary Assessments (EMA) thus increase validity and decrease recall bias. They are suitable for asking users in their daily environment about their state of being, which can change over time, by random or interval time sampling. Combining EMAs and mobile crowdsensing sensor measurements allows for multimodal analyses, which can gain new insights in, e.g., chronic diseases8,15. The datasets used within this work have EMA in common and are described in the following subsection.

From ongoing projects of our team, we are constantly collecting mHealth data as well as Ecological Momentary Assessments6,17,18,19. To investigate how the machine learning performance varies based on the splits, we wanted different datasets with different use cases. However, to increase comparability between the use cases, we created multi-class classification tasks.

We train each model using historical assessments, the oldest assessment was collected at time tstart, the latest historical assessment at time tlast. A current assessment is created and collected at time tnow, a future assessment at time tnext. Depending on the study design, the actual point of time tnext may be in some hours or in a few weeks from tnow. For each dataset and for each user, we want to predict a feature (synonym, a question of an assessment) at time tnext using the features at time tnow. This feature at time tnext is then called the target. For each use case, a model is trained using data between tstart and tlast, and given the input data from tnow, it predicts the target at tnext. Figure1 gives a schematic representation of the relevant points of time tstart,tlast,tnow, and tnext.

At time tstart, the first assessment is given; tlast is the last known assessment used for training, whereas tnow is the currently available assessment as input for the classifier and the target is predicted at time ttext.

To increase comparability between the approaches, we used the same model architecture with the same pseudo-random initialisation. The model is a Random Forest classifier with 100 trees and the Gini impurity as the splitting criterion. The whole coding was in Python 3.9, using mostly scikit-learn, pandas and Jupyter Notebooks. Details can be found on GitHub in the supplementary material.

For all datasets that we used in this study, we have ethical approvals (UNITI No. 20-1936-101, TYT No. 15-101-0204, Corona Check No. 71/20-me, and Corona Health No. 130/20-me). The following section provides an overview of the studies, the available datasets with characteristics, and then describes each use case in more detail. An brief overview is given in Table1 with baseline statistics for each dataset in Table2.

To provide some more background info about the studies: The analyses happen with all apps on the so-called EMA questionnaires (synonym: assessment), i.e., the questionnaires that are filled out multiple times in all apps and the respective studies. This can happen several times a day (e.g., for the tinnitus study TrackYourTinnitus (TYT)) or at weekly intervals (e.g., studies in the Corona Health (CH) app). Nevertheless, the analysis happens on the recurring questionnaires, which collect symptoms over time and in the real environment through unforeseen (i.e., random) notifications.

The TrackYourTinnitus (TYT) dataset has the most filled-out assessments with more than 110,000 questionnaires as by 2022-10-24. The Corona Check (CC) study has the most users. This is because each time an assessment is filled out, a new user can optionally be created. Notably, this app has the largest ratio of non-German users and the youngest user group with the largest standard deviation. The Corona Health (CH) app with its studies Mental health for adults, adolescents and physical health for adults has the highest proportion of German users because it was developed in collaboration with the Robert Koch Institute and was primarily promoted in Germany. Unification of treatments and Interventions for Tinnitus patients (UNITI) is a European Union-wide project, which overall aim is to deliver a predictive computational model based on existing and longitudinal data19. The dataset from the UNITI randomized controlled trial is described by Simoes et al.20.

With this app, it is possible to record the individual fluctuations in tinnitus perception. With the help of a mobile device, users can systematically measure the fluctuations of their tinnitus. Via the TYT website or the app, users can also view the progress of their own data and, if necessary, discuss it with their physician.

The ML task at hand is a classification task with target variable Tinnitus distress at time tnow and the questions from the daily questionnaire as the features of the problem. The targets values range in [0,1] on a continuous scale. To make it a classification task, we created bins with step size of 0.2 resulting in 5 classes. The features are perception, loudness, and stressfulness of tinnitus, as well as the current mood, arousal and stress level of a user, the concentration level while filling out the questionnaire, and perception of the worst tinnitus symptom. A detailed description of the features was already done in previous works21. Of note, the time delta of two assessments of one user at tnext and tnow varies between users. Its median value is 11 hours.

The overall goal of UNITI is to treat the heterogeneity of tinnitus patients on an individual basis. This requires understanding more about the patient-specific symptoms that are captured by EMA in real time.

The use case we created at UNITI is like that of TYT. The target variable encumbrance, coded as cumberness, which was also continuously recorded, was divided into an ordinal scale from 0 to 1 in 5 steps. Features also include momentary assessments of the user during completion, such as jawbone, loudness, movement, stress, emotion, and questions about momentary tinnitus. The data was collected using our mobile apps7. Here, of note: on average, the median time gap between two assessment is 24 hours for each user.

At the beginning of the COVID-19 pandemic, it was not easy to get initial feedback about an infection, given the lack of knowledge about the novel virus and the absence of widely available tests. To assist all citizens in this regard, we launched the mobile health app Corona Check together with the Bavarian State Office for Health and Food Safety22.

The Corona Check dataset predicts whether a user has a Covid infection based on a list of given symptoms23. It was developed in the early pandemic back in 2020 and helped people to get quick estimate for an infection without having an antigen test. The target variable has four classes: First, suspected coronavirus (COVID-19) case", second, symptoms, but no known contact with confirmed corona case", third, contact with confirmed corona case, but currently no symptoms", and last, neither symptoms nor contact".

The features are a list of Boolean variables, which were known at this time to be typically related with a Covid infection, such as fever, a sore throat, a runny nose, cough, loss of smell, loss of taste, shortness of breath, headache, muscle pain, diarrhea, and general weakness. Depending on the answers given by a user, the application programming interface returned one of the classes. The median time gap of two assessments for the same user is 8 hours on average with a much larger standard deviation of 24.6 days.

The last four use cases are all derived from a bigger Covid-related mHealth project called Corona Health6,24. The app was developed in collaboration with the Robert Koch-Institute and was primarily promoted in Germany, it includes several studies about the mental or physical health, or the stress level of a user. A user can download the app and then sign up for a study. He or she will then receive a baseline one-time questionnaire, followed by recurring follow-ups with between-study varying time gaps. The follow-up assessment of CHA has a total of 159 questions including a full PHQ9 questionnaire25. We then used the nine questions of PHQ9 as features at tnow to predict the level of depression for this user for tnext. Depression levels are ordinally scaled from None to Severe in a total of 5 classes. The median time gap of two assessments for the same user is 7.5 days. That is, the models predict the future in this time interval.

Similar to the adult cohort, the mental health of adolescents during the pandemic and its lock-downs is also captured by our app using EMA.

A lightweight version of the mental health questionnaire for adults was also offered to adolescents. However, this did not include a full PHQ9 questionnaire, so we created a different use case. The target variable to be classified on a 4-level ordinal scale is perceived dejection coming from the PHQ instruments, features are a subset of quality of live assessments and PHQ questions, such as concernment, tremor, comfort, leisure quality, lethargy, prostration, and irregular sleep. For this study, the median time gap of two follow up assessments is 7.3 days.

Analogous to the mental health of adults, this study aims to track how the physical health of adults changes during the pandemic period.

Adults had the option to sign up for a study with recurring assessments asking for their physical health. The target variable to be classified asks about the constraints in everyday life that arise due to physical pain at tnext. The features for this use case include aspects like sport, nutrition, and pain at tnow. The median time gap of two assessments for the same user is 14.0 days.

This additional study within the Corona Health app asks users about their stress level on a weekly basis. Both features and target are assessed on a five-level ordinal scale from never to very often. The target asks for the ability of stress management, features include the first nine questions of the perceived stress scale instrument26. The median time gap of two assessments for the same user on average is 7.0 days.

We also want to compare the ML approaches with a baseline heuristic (synonym: Baseline model). A baseline heuristic can be a simple ML model like a linear regression or a small Decision Tree, or alternatively, depending on the use case, it could also be a simple statement like The next value equals the last one". The typical approach for improving ML models is to estimate the generalization error of the model on a benchmark data set when compared to a baseline heuristic. However, it is often not clear, which baseline heuristic to consider, i.e.: The same model architecture as the benchmark model, but without tuned hyperparameters? A simple, intrinsically explainable model with or without hyperparameter tuning? A random guess? A naive guess, in which the majority class is predicted? Since we have approaches on a user-level (i.e., we consider users when splitting) and on an assessment-level (i.e., we ignore users when splitting), we also should create baseline heuristics on both levels. We additionally account for within-user variance in Ecological Momentary Assessments by averaging a users previously known assessments. Previously known here means that we calculate the mode or median of all assessments of a user that are older than the given timestamp. In total, this leads to four baseline heuristics (user-level latest, user-level average, assessment-level latest, assessment-level average) that do not use any machine learning but simple heuristics. On the assessment-level, the latest known target or the mean of all known targets so far is taken to predict the next target, no matter of the user-id of this assessment. On the user-level, either the last known, or median, or mode value of this user is taken to predict the target. This, in turn, leads to a cold-start problem for users that appear for the first time in a dataset. In this case, either the last known, or mode, or median of all assessments that are known so far are taken to predict the target.

Before the data and approaches could be compared, it was necessary to homogenize them. In order for all approaches to work on all data sets, at least the following information is necessary: Assessment_id, user_id, timestamp, features, and the target. Any other information such as GPS data, or additional answers to questions of the assessment, we did not include into the ML pipeline. Additionally, targets that were collected on a continuous scale, had to be binned into an ordinal scale of five classes. For an easier interpretation and readability of the outputs, we also created label encodings for each target. To ensure consistency of the pre-processing, we created helper utilities within Python to ensure that the same function was applied on each dataset. For missing values, we created a user-wise missing value treatment. More precisely, if a user skipped a question in an assessment, we filled the missing value with the mean or mode (mode = most common value) of all other answers of this user for this assessment. If a user had only one assessment, we filled it with the overall mean for this question.

For each dataset and for each script, we set random states and seeds to enhance reproducibility. For the outer validation set, we assigned the first 80 % of all users that signed up for a study to the train set, the latest 20% to the test set. To ensure comparability, the test users were the same for all approaches. We did not shuffle the users to simulate a deployment scenario where new users join the study. This would also add potential concept drift from the train to the test set and thus improve the simulation quality.

For the cross-validation within the training set, which we call internal validation, we chose a total of 5 folds with 1 validation fold. We then applied the four baseline heuristics (on user level and assessment level with either latest target or average target as prediction) to calculate the within-train-set performance standard deviation and the mean of the weighted F1 scores for each train fold. The mean and standard deviation of the weighted F1 score are then the estimator of the performance of our model in the test set.

We call one approach superior to another if the final score is higher. The final score to evaluate an approach is calculated as:

$${f}_{1}^{final}={f}_{1}^{test}-alpha {sigma }left({f}_{1}^{train}right)$$

(1)

If the standard deviation between the folds during training is large, the final score is lower. The test set must not contain any selection bias against the underlying population. The pre-factor of the standard deviation is another hyperparameter. The more important model robustness for the use case, the higher should be set.

Within cross-validation, there exist several approaches on how to split up the data into folds and validate them, such as the k-fold approach with k as the number of folds in the training set. Here, k1 folds form the training folds and one fold is the validation fold27. One can then calculate k performance scores and their standard deviation to get an estimator for the performance of the model in the test set, which itself is an estimator for the models performance after deployment (see also Fig.2).

Schematic visualisation of the steps required to perform a k-fold cross-validation, here with k=5.

In addition, there exist the following strategies: First, (repeated) stratified k-fold, in which the target distribution is retained in each fold, which can also be seen in Fig.3. After shuffling the samples, the stratified split can be repeated3. Second, leave-one-out cross-validation28, in which the validation fold contains only one sample while the model has been trained on all other samples. And third, leave-p-out cross-validation, in which (left(begin{array}{c}n\ pend{array}right)) train-test-pairs are created with n equals number of assessments (synonym sample)29.

While this approach retains the class distribution in each fold, it still ignores user groups. Each color represents a different class or user id.

These approaches, however, do not always focus on samples that might belong to our mHealth data peculiarities. To be more specific, they do not account for users (syn. groups, subjects) that generate daily assessments (syn. samples) with a high variance.

To precisely explain the splitting approaches, we would like to differentiate between the terms folds and sets. We call a chunk of samples (synonym: assessments, filled-out questionnaires) a set on the outer split of the data, for which we cut-off the final test set. However, within the training set, we then split further to create training and validation folds. That is, using the term fold, we are in the context of cross validation. When we use the term set, then we are in the outer split of the ML pipeline. Figure4 visualizes this approach. Following this, we define 4 different approaches to split the data. For one of them we ignore the fact that there are users, for the other three we do not. We call these approaches user-cut, average-user, user-wise and time-cut. All approaches have in common that the first 80 % of all users are always in the training set and the remaining 20 % are in the test set. A schematic visualization of the splitting approaches is shown in Fig.5. Within the training set, we then split on user-level for the approaches user-cut, average-user and user-wise, and on assessment-level for the approach time-cut.

In the second step, users are ordered by their study registration time, with the initial 80 % designated as training users and the remaining 20 % as test users. Subsequently, assessments by training users are allocated to the training set, and those by test users to the test set. Within the training set, user grouping dictates the validation approach: group-cross-validation is applied if users are declared as a group, otherwise, standard cross-validation is utilized. We compute the average f1 score, ({f}_{1}^{train}), from training folds and the f1 score on the test set, ({f}_{1}^{test}). The standard deviation of ({f}_{1}^{train},sigma ({f}_{1}^{train})), indicates model robustness. The hyperparameter adjusts the emphasis on robustness, with higher values prioritizing it. Ultimately, ({f}_{1}^{final}), which is a more precise estimate if group-cross-validation is applied, offers a refined measure of model performance in real-world scenarios.

Yellow means that this sample is part of the validation fold, green means it is part of a training fold. Crossed out means that the sample has been dropped in that approach because it does not meet the requirements. Users can be sorted by time to accommodate any concept drift.

In the following section, we will explain the splitting approaches in more detail. The time-cut approach ignores the fact of given groups in the dataset and simply creates validation folds based on the time the assessments arrive in the database. In this example, the month, in which a sample was collected, is known. More precisely, all samples from January until April are in the training set while May is in the test set. The user-cut approach shuffles all user ids and creates five data folds with distinct user-groups. It ignores the time dimension of the data, but provides user-distinct training and validation folds, which is like the GroupKFold cross-validation approach as implemented in scikit-learn30. The average-user approach is very similar to the user-cut approach. However, each answer of a user is replaced by the median or mode answer of this user up to the point in question to reduce within-user-variance. While all the above-mentioned approaches require only one single model to be trained, the user-wise approach requires as many models as distinct users are given in the dataset. Therefore, for each user, 80 % of his or her assessments are used to train a user-specific model, and the remaining 20% of the time-sorted assessments are used to test the model. This means that for this approach, we can directly evaluate on the test set as each model is user specific and we solved the cold-start problem by training the model on the first assessments of this user. If a user has less than 10 assessments, he or she is not evaluated on that approach.

Approval for the UNITI randomized controlled trial and the UNITI app was obtained by the Ethics Committee of the University Clinic of Regensburg (ethical approval No. 20-1936-101). All users read and approved the informed consent before participating in the study. The study was carried out in accordance with relevant guidelines and regulations. The procedures used in this study adhere to the tenets of the Declaration of Helsinki. The Track Your Tinnitus (TYT) study was approved by the Ethics Committee of the University Clinic of Regensburg (ethical approval No. 15-101-0204). The Corona Check (CH) study was approved by the Ethics Committee of the University of Wrzburg (ethical approval no. 71/20-me) and the universitys data protection officer and was carried out in accordance with the General Data Protection Regulations of the European Union. The procedures used in the Corona Health (CH) study were in accordance with the 1964 Helsinki declaration and its later amendments and was approved by the ethics committee of the University of Wrzburg, Germany (No. 130/20-me). Ethical approvals include secondary use. The data from this study are available on request from the corresponding author. The data are not publicly available, as the informed consent of the participants did not provide for public publication of the data.

Further information on research design is available in theNature Portfolio Reporting Summary linked to this article.

See the article here:
Practical approaches in evaluating validation and biases of machine learning applied to mobile health studies ... - Nature.com

April 22, 2024 -Artificial intelligence tools (AI) are proliferating in healthcare at breakneck speed, amplifying calls from healthcare leaders and policymakers for greater alignment on the responsible use of these tools.

The FDA authorized692 artificial intelligence and machine learning devices in 2023, 171 devices more than its 2022 list. In response to this abundance of tools, various organizations have released their own definitions of what it means to use AI responsibly, including organizations like Pfizer, Kaiser Permanente, Optum, the White House, and others. However, the industry lacks an overarching set of principles to improve alignment. Brian Anderson, MD, CEOand co-founder of the Coalition for Health AI (CHAI), discusses the need for guidelines and standards for responsible AI in healthcare. He highlights key principles of responsible AI and emphasizes the importance of having various stakeholders--including patients--involved in developingthese standards and guidelines.

Brian Anderson, MD:

Is there an agreed-upon consensus around what trustworthy, responsible AI looks like in health? The answer we quickly found is no. You have a lot of very innovative companies and health systems building AI according to their own definitions of what that means and what that looks like.

Kelsey Waddill:

Welcome to Season 2 of Industry Perspectives, coming to you from HIMSS 2024 in Orlando, Florida. I'm Kelsey Waddill, multimedia manager and managing editor at Xtelligent Healthcare.

From 2024 to 2030, healthcare AI in the US market is expected to experience a 35.8 percent compound annual growth rate. As the tool proliferates and in view of the risks inherent to the healthcare industry, guidelines for AI utilization are essential and many are asking: what does it mean to use artificial intelligence responsibly in the healthcare context? Brian Anderson, CEO, chief digital health physician, and co-founder of the Coalition for Health AI, or CHAI, founded CHAI with this question in mind. He's going to break it down for us on today's episode of Industry Perspectives.

Brian, it's great to have you on Healthcare Strategies today. Thank you for making the time in a busy HIMSS schedule, I know, so we're glad that this could work out. I wanted to start out by asking about CHAI, the organization I know you co-founded in 2021, and so I wanted to get some more background on just how it started, your story.

Brian Anderson, MD:

Yeah, really its roots came out of the pandemic. So in the pandemic there were a lot of organizations that were non-traditional partners that were inherently competitive. You had pharma companies working together, you had technology giants working together trying to address this pandemic that was in front of us. And so coming out of it, there was a lot of goodwill in that space between a Microsoft, and a Google, and an Amazon, or various competing health systems wanting to still try to find a way to do good together.

And so one of the questions a number of us asked was: is there an agreed upon consensus around what trustworthy, responsible AI looks like in health? The answer we quickly found is no. You have a lot of very innovative companies and health systems building AI according to their own definitions of what that means and what that looks like. And inherently that means that there's a level of opaqueness to how we think about trust and how we think about performance of AI in these siloed companies. In a consequential space like health, that can be a problem.

And so we agreed that it would be a worthwhile cause and mission for our merry band of eight or ten health systems and technology companies to come together to really begin trying to build that consensus definition of what responsible, trustworthy, healthy AI looks like. And so that's been our north star since we launched in 2021, and it quickly took off.

It started with those eight or ten. It quickly grew to like a hundred. The US government got involved right away. Office of the National Coordinator and the Food and Drug Administration--which are the two main regulating bodies for AI--quickly joined our effort. White House, NIH, all the HHS agencies. And then, excitingly, a large number of health systems and other tech companies joined, to the point today where it got to the point where we had 1,300 organizations, thousands of people, part of this coalition of the willing.

That became a challenge. How do you meaningfully engage 1,300 organizations, 2,000 to 3000-plus individuals with a volunteer group? Not very well. And so we made the decision to form a nonprofit, started that in January of this year. The board was convened. They asked me to be CEO. I'm very honored and humbled by that. And so I took that role on. Today is day two, I think, technically, in my new role.

Kelsey Waddill:

Congratulations!

Brian Anderson, MD:

Thanks. So I'm really excited to be here at HIMSS, take part of the vibrant conversation that everyone is having about AI here.

Kelsey Waddill:

Yeah, well, I mean it's definitely one of the major themes of the conference this year. And for good reason because, in the last year alone, there's been so much change in this space in AI specifically and its use in healthcare.

I did want to zoom in for one second, you mentioned this phrase, I know it's throughout your website and your language, and it's "responsible AI in healthcare." And I feel like that's the question right now, is: what does that look like? What does that mean? And so I know that's a big part of why CHAI convened to begin with. So I was wondering if you could shed some light on what you found so far and what that means.

Brian Anderson, MD:

Yeah. It's an important thing to be grounded in. So it starts with, I think, in the basic context that health is a very consequential space. All of us are patients or caregivers at one point in our life. And so the tools that are used on us as patients or caregivers need to have a set of aligned principles that are aligned to the values of our society, that allow us to ensure that the values that we have as humans in our society align with those tools.

And so some of those very basic principles in responsible AI are things like fairness. So is the AI that's created fair? Does it treat people equally? Does it treat people fairly? When there's this concept in AI that I remind people, all algorithms are at the end of the day, is they're programs that are trained on our histories. And so it's a really critical question that we ask ourselves is: are those histories fair? Are they equitable? Right? And you smile, obviously the answer is probably not.

Kelsey Waddill:

Probably no.

Brian Anderson, MD:

And so then it takes an intentional effort when we think about fairness and building AI to do that in a fair way.

Another concept, there are concepts like privacy and security. So the kinds of AI tools that we build, we don't want them to leak out training data that might be, especially in health, personal identifiable, protected health information. And so how we build models--particularly, there's been some news in the LLM space that if you do the right kind of prompting or hacking of these models, you can get it to reveal what its training data is. So how, again, we build models in a privacy-preserving, secure way that doesn't allow for that is really important.

There are other concepts like transparency, which is really important. When you use a tool, you want to know how that tool was made. What were the materials it was made out of? Is it safe for me to get into this car and drive somewhere? Does it meet certain safety standards? For many of the common day things that we use, from microwaves, to toaster ovens, to cars, there's places where you can go and you can read the report, the safety reports on those tools.

In AI, we don't have that yet. And so there's a real transparency problem when it comes to understanding very basic things like: what was this model trained on? What was it made of? How did it do when you tested it and evaluated it? We have all of the car safety tests, the famous car crash videos that we are all familiar with. We don't have that kind of testing information that is developed and maintained by independent entities. We have organizations that sell models, technology companies that make certain claims, but the ability to independently validate that, very hard.

And so this idea of transparency in terms of how models we're trained, what their testing evaluation scores were, what their indications and limitations are, and a whole slew of other things go into this concept of transparency. So principles like that.

Other ones like usability, reliability, robustness--these are all principles of responsible AI. I won't bother detailing them all for you, but those are the principles that we're basing CHAI around. And so when we talk about building the standards or technical specifications, it means going from a 50,000-foot level and saying fairness is important to saying, "okay, in the generative AI space, what does bias mean? What does fairness mean in the output of an LLM?" We don't have a shared agreed upon common definition of what that looks like, and it's really important to have that.

So I'll give you an example. A healthcare organization wants to deploy an LLM as a chat bot out on their website. That LLM, if you give it the same prompt five or six times might have different outputs. A patient trying to navigate beneficiary enrollment might be sent in five different directions if they were to give the same prompting. So that brings up concepts like reliability, fairness, and how we measure accuracy. These are all things that are principles in responsible AI that for generative AI, for chatbots, we don't have a common agreed upon framework about what "good" looks like and how to measure alignment to that good. And so that's what we're focusing on in CHAI because it's an important question. We want that individual going to that hypothetical website with that chatbot to have a fair, reliable experience getting to the right place.

Kelsey Waddill:

Yeah, I think that captures pretty well the complexity of the role that AI plays in healthcare right now. The questions that are being asked right now in each of those points we could dive into for a while. But I am curious: we want to do this well, we want to build out these guidelines well, but there's also a bit of time pressure it seems like from my perspective, in terms of there's people who are, as you kind of alluded to, the privacy and security piece, there's those who want to use this and use any holes that we haven't quite figured out how to cover yet for malicious intent. There's that time pressure. There's also just the time pressure of: people are generating uses of AI at a very rapid pace, and we don't have a structure for this yet that's set in stone.

So I was curious what you would recommend in terms of prioritization in these conversations. Obviously that list you just mentioned I'm sure is part of that, but is there anything else that you'd say about how to pinpoint what we need to be working on right now?

Brian Anderson, MD:

It's a good question. In the health space, it's really hard because it's so consequential. A patient's life could be on the line, and yet you have innovators that are innovating in fantastic, amazing ways that could potentially save people's lives, developing new models that have emerging capabilities, identifying potential diagnoses, identifying new molecules that have never been thought of by a human before. And yet, because it's so consequential, we want to have the right guardrails and guidelines in place.

And so one of the approaches that I think we are recommending, two-part. One is when we formed CHAI, we wanted it to have innovators and regulators at the same table. And so these working groups are going to be focusing on developing these standards, developing the metrics and the methodology for evaluation, with having innovators, and regulators, and patients all at the same table. Because you can't have innovation working at the speed of light developing it without an understanding of what safe and effective, and what the guardrails are. You can't have regulators developing those guardrails without understanding what the standards are and the consensus perspective of what good responsible AI looks like coming from the innovators. And so one part of the answer to your question is bringing all the right stakeholders to the table and having patients at the center of what we're doing.

The second part is--so, because health is so consequential, there's risk involved. I would argue that there's layers or different levels of risk. An ICU, one might agree there's a level of risk that's pretty significant, a patient's life. They're on that edge in terms of life and death. AI that might be used in helping to move beds around in a hospital, not so consequential. So a patient's life might not necessarily be on the line determining efficiencies about where beds are moving.

And so from that perspective, we are looking to work with the innovation community to identify where they can accelerate and innovate in safe spaces, in those bed management efficiency, back office administration, drafting of emails, variety of different use cases that aren't as consequential or aren't as risky. Whereas the more risky ones, the ones like in the ICU where life and death is a matter of potentially what an AI tool is going to recommend or not recommend, those ones require going much more slowly and thinking through the consequences with more rigor and stronger guidelines and guardrails.

And so that's how we're approaching it, is: identifying where the less risky places are, looking to support innovation, building out guidelines around what responsible, trustworthy AI looks like, while slowly building up to some of those more risky places.

Kelsey Waddill:

That makes sense. And in our last question here, I just wanted to hear what you're excited about for this next year in this space, and specifically what CHAI's doing.

Brian Anderson, MD:

Yeah. I would say we're at a very unique moment in AI. I had shared earlier that all algorithms are are programs trained on our histories. We have an opportunity to address that set of inequities in our past. And that could take the form of a variety of different responses.

One of them, the one I hope, and the one we're driving to in CHAI is: how do we meaningfully engage communities that haven't traditionally had the opportunity to participate in so many of the digital health revolutions that have come before? As an example, models require data for training. To create a model that performs well, you need to have robust training data for it to be trained and tuned on whatever that population is. And so how can we work with marginalized communities to enable them to tell their digital story so that we can then partner with the model vendors to then train models on data from those communities, so that those models can be used in meaningful ways to help those communities and the health of those communities? That's one of the exciting things I'm looking forward to for the next year.

Kelsey Waddill:

Great. Well, I'm excited too. And thank you so much for this conversation and for making time for it today.

Brian Anderson, MD:

Absolutely. Thanks, Kelsey.

Kelsey Waddill:

Thank you.

Listeners, thank you for joining us on Healthcare Strategies's Industry Perspectives. When you get a chance, subscribe to our channels on Spotify and Apple. And leave us a review to let us know what you think of this new series. More Industry Perspectives are on the way, so stay tuned.

Excerpt from:
What principles should guide artificial intelligence innovation in healthcare? - HealthCareExecIntelligence

Experimental results Microstructure evolution

The ZK30s AA and ECAPed conditions of the inverse pole figures (IPF) coloring patterns and associated band contrast maps (BC) are shown in Fig.2. High-angle grain boundaries (HAGBs) were colored black, while Low-angle grain boundaries (LAGBs) were colored white for AA condition, and it was colored red for 1P and Bc, as shown in Fig.2. The grain size distribution and misorientation angle distribution of the AA and ECAPed ZK30 samples is shown in Fig.3. From Fig.2a, it was clear that the AA condition revealed a bimodal structure where almost equiaxed refined grains coexist with coarse grains and the grain size was ranged between 3.4 up to 76.7m (Fig.3a) with an average grain size of 26.69m. On the other hand, low fraction of LAGBs as depicted in Fig.3b. Accordingly, the GB map (Fig.2b) showed minimal LAGBs due to the recrystallization process resulting from the annealing process. ECAP processing through 1P exhibited an elongated grain alongside refined grains and the grain size was ranged between 1.13 and 38.1m with an average grain size of 3.24m which indicated that 1P resulted in a partial recrystallization, as shown in Fig.2c,d. As indicated in Fig.2b 1P processing experienced a refinement in the average grain size of 87.8% as compared with the AA condition. In addition, from Fig.2b it was clear that ECAP processing via 1P resulted a significant increase in the grain aspect ratio due to the uncomplete recrystallization process. In terms of the LAGBs distribution, the GB maps of 1P condition revealed a significant increase in the LAGBs fraction (Fig.2d). A significant increase in the LAGBs density of 225% after processing via 1P was depicted compared to the AA sample (Fig.2c). Accordingly, the UFG structure resulted from ECAP processing through 1P led to increase the fraction of LAGBs which agreed with previous study35,36. Shana et al.35 reported that during the early passes of ECAP a generation and multiplication of dislocation is occur which is followed by entanglement of the dislocation forming the LAGBs and hence, the density of LAGBs was increased after processing through 1P. The accumulation of the plastic strain up to 4Bc revealed an almost UFG, which indicated that 4Bc led to a complete dynamic recrystallization (DRX) process (Fig.2e). The grain size was ranged between 0.23 up to 11.7m with average grain size of 1.94m (the average grain size was decreased by 92.7% compared to the AA condition). On the other hand, 4Bc revealed a decrease in the LAGBs density by 25.4% compared to 1P condition due to the dynamic recovery process. The decrease in the LAGBs density after processing through 4Bc was coupled with an increase in the HAGBs by 4.4% compared to 1P condition (Figs.2f, 3b). Accordingly, the rise of the HAGBs after multiple passes can be referred to the transfer of LAGBs into HAGBs during the DRX process.

IPF coloring maps and their corresponding BC maps, superimposed for the ZK30 billets in its AA condition (a,b), and ECAP processed through (c,d) 1P, (e,f) 4Bc (with HAGBs in black lines and LAGBs in white lines (AA) and red lines (1P, 4Bc).

Relative frequency of (a) grain size and (b) misorientation angle of all ZK30 samples.

Similar findings were reported in previous studies. Dumitru et al.36 reported that ECAP processing resulted in the accumulation and re-arrangement of dislocations which resulted in forming a subgrains and equiaxed grains with an UFG structure and a fully homogenous and equiaxed grain structure for ZK30 alloy was attained after the third pass. Furthermore, they reported that the LAGBs is transferred into HAGBs during the multiple passes which leads to the decrease in the LAGBs density. Figueiredo et al.37 reported that the grains evolved during the early passes of ECAP into a bimodal structure while further processing passes resulted in the achievement of a homogenous UFG structure. Zhou et al.38 reported that by increasing the processing passes resulted in generation of new grain boundaries which resulted in increasing the misorientation to accommodate the deformation and the Geometrically Necessary Dislocations (GNDs) generated a part of the total dislocations with a HAGBs, thus develop misorientations between the neighbor grains. Tong et al.39 reported that the fraction of LAGBs is decreased during multiple passes for MgZnCa alloy.

Figure4a displays X-ray diffraction (XRD) patterns of the AA-ZK30 alloy, 1P, and 4Bc extruded samples, revealing peaks corresponding to primary -Mg Phase, Mg7Zn3, and MgZn2 phases in all extruded alloys, with an absence of diffraction peaks corresponding to oxide inclusions. Following 1P-ECAP, the -Mg peak intensity exhibits an initial increase, succeeded by a decrease and fluctuations, signaling texture alterations in the alternative Bc route. The identification of the MgZn2 phase is supported by the equilibrium MgZn binary phase diagram40. However, the weakened peak intensity detected for the MgZn2 phase after the 4BcECAP process indicates that a significant portion of the MgZn2 dissolved into the Mg matrix, attributed to their poor thermal stability. Furthermore, the atomic ratio of Mg/Zn for this phase is approximately 2.33, leading to the deduction that the second phase is the Mg7Zn3 compound. This finding aligns with recent research on MgZn alloys41. Additionally, diffraction patterns of ECAP-processed samples exhibit peak broadening and shifting, indicative of microstructural adjustments during plastic deformation. These alterations undergo analysis for crystallite size and micro-strain using the modified Williamson and Hall (WH) method42, as illustrated in Fig.4b. After a single pass of ECAP, there is a reduction in crystallite size and an escalation in induced micro-strain. Subsequent to four passes-Bc, further reductions in crystallite size and heightened micro-strain (36nm and 1.94103, respectively) are observed. Divergent shearing patterns among the four processing routes, stemming from disparities in sample rotation, result in distinct evolutions of subgrain boundaries. Route BC, characterized by the most extensive angular range of slip, generates subgrain bands on two shearing directions, expediting the transition of subgrain boundaries into high-angle grain boundaries43,44. Consequently, dislocation density and induced micro-strains reach their top in route BC, potentially influenced by texture modifications linked to orientation differences in processing routes. Hence, as the number of ECAP passes increases, an intensive level of deformation is observed, leading to the existence of dynamic recrystallization and grain refinement, particularly in the ECAP 4-pass. This enhanced deformation effectively impedes grain growth. Consequently, the number of passes in the ECAP process is intricately linked to the equivalent strain, inducing grain boundary pinning, and resulting in the formation of finer grains. The grain refinement process can be conceptualized as a repetitive sequence of dynamic recovery and recrystallization in each pass. In the case of the 4Bc ECAP process, dynamic recrystallization dominates, leading to a highly uniform grain reduction and, causing the grain boundaries to become less distinct45. Figure4b indicates that microstructural features vary with ECAP processing routes, aligning well with grain size and mechanical properties.

(a) XRD patterns for the AA ZK30 alloy and after 1P and 4Bc ECAP processing, (b) variations of crystallite size and lattice strain as a function of processing condition using the WilliamsonHall method.

Figure5 shows the volume loss (VL) and average coefficient of friction (COF) for the AA and ECAPed ZK30 alloy. The AA billets exhibited the highest VL at all wear parameters compared to the ECAPed billets as shown in Fig.5. From Fig.5a it revealed that performing the wear test at applied load of 1N exhibited the higher VL compared to the other applied forces. In addition, increasing the applied force up to 3 N revealed lower VL compared to 1 N counterpart at all wear speeds. Further increase in the applied load up to 5 N revealed a notable decrease in the VL. Similar behavior was attained for the ECAP-processed billets through 1P (Fig.5c) and 4Bc (Fig.5e). The VL was improved by increasing the applied load for all samples as shown in Fig.5 which indicated an enhancement in the wear resistance. Increasing the applied load increases the strain hardening of ZK30 alloy that are in contact as reported by Yasmin et al.46 and Kori et al.47. Accordingly, increasing the applied load resulted in increasing the friction force, which in turn hinder the dislocation motion and resulted in higher deformation, so that ZK30 experienced strain hardening and hence, the resistance to abrasion is increased, leading to improving the wear resistance48. Furthermore, increasing the applied load leads to increase the surface in contact with wear ball and hence, increases gripping action of asperities, which help to reduces the wear rate of ZK30 alloy as reported by Thuong et al.48. Out of contrary, increasing the wear speed revealed increasing the VL of the AA billets at all wear loads. For the ECAPed billet processed through 1P, the wear speed of 125mm/s revealed the lowest VL while the wear speed of 250mm/s showed the highest VL (Fig.5c). Similar behaviour was recorded for the 4Bc condition. In addition, from Fig.5c, it was clear that 1P condition showed higher VL compared to 4Bc (Fig.5e) at all wear parameters, indicating that processing via multiple passes resulted in significant grain size refinement (Fig.2). Hence, higher hardness and better wear behavior were attained which agreed with previous study7. In addition, from Fig.5, it was clear that increasing the wear speed increased the VL. For the AA billets tested at 1N load the VL was 1.52106 m3. ECAP processing via 1P significantly improved the wear behavior as the VL was reduced by 85% compared to the AA condition. While compared to the AA condition, the VL improved by 99.8% while straining through 4Bc, which is accounted for by the considerable refinement that 4Bc provides. A similar trend was observed for the ECAPed ZK30 samples tested at a load of 3 and 5 N (Fig.5). Accordingly, the significant grain refinement after ECAP processing (Fig.2) increased the grain boundaries area; hence, a thicker oxide protective layer can be formed, leading to improve the wear resistance of the ECAPed samples. It is worth to mentioning here that, the grain refinement coupled with refining the secondary phase particle and redistribution resulted from processing through ECAP processing through multiple passes resulted in improving the hardness, wear behavior and mechanical properties according to HallPetch equation7,13,49. Similar findings were noted for the ZK30 billets tested at 3 N load, processing through 1P and 4Bc exhibited decreasing the VL by 85%, 99.85%, respectively compared to the AA counterpart. Similar finding was recorded for the findings of ZK30 billets which tested at 5 N load.

Volume loss of ZK30 alloy (a,c,e) and the average coefficient of friction (b,d,f) in its (a,b) AA, (c,d) 1P and (e,f) 4Bc conditions as a function of different wear parameters.

From Fig.5, it can be noticed that the COF curves revealed a notable fluctuation with implementing least square method to smoothing the data, confirming that the friction during the testing of ECAPed ZK30 alloy was not steady for such a time. The remarkable change in the COF can be attributed to the smaller applied load on the surface of the ZK30 samples. Furthermore, the results of Fig.5 revealed that ECAP processing reduced the COF, and hence, better wear behavior was attained. Furthermore, for all ZK30 samples, it was observed that the highest applied load (5 N) coupled with the lowest wear time (110s) exhibited better COF and better wear behavior was displayed. These findings agreed with Farhat et al.50, they reported that decreasing the grain size led to improve the COF and hence improve the wear behavior. Furthermore, they reported that a plastic deformation occurs due to the friction between contacted surface which resisted by the grain boundaries and fine secondary phases. In addition, the strain hardening resulted from ECAP processing leads to decrease the COF and improving the VL50. Sankuru et al.43 reported that ECAP processing foe pure Mg resulted in substantial grain refinement which was reflected in improving both microhardness and wear rate of the ECAPed billets. Furthermore, they found that increasing the number of passes up to 4Bc reduced the wear rate by 50% compared to the AA condition. Based on the applied load and wear velocity and distance, wear mechanism can be classified into mild wear and severe wear regimes49. Wear test parameters in the present study (load up to 5 N and speed up to 250mm/s) falls in the mild wear regime where the delamination wear and oxidation wear mechanisms would predominantly take place43,51.

The worn surface morphologies of the ZK30-AA billet and ECAPed billet processed through 4Bc are shown in Fig.6. From Fig.6 it can revealed that scores of wear grooves which aligned parallel to the wear direction have been degenerated on the worn surface in both AA (Fig.6a) and 4Bc (Fig.6b) conditions. Accordingly, the worn surface was included a combination of adhesion regions and a plastic deformation bands along the wear direction. Furthermore, it can be observed that the wear debris were adhered to the ZK30 worn surface which indicated that the abrasion wear mechanism had occur52. Lim et al.53 reported that hard particle between contacting surfaces scratches samples and resulted in removing small fragments and hence, wear process was occurred. In addition, from Fig.6a,b it can depicted that the wear grooves on the AA billet were much wider than the counterpart of the 4Bc sample and which confirmed the effectiveness of ECAP processing in improving the wear behavior of the ZK30 alloy. Based on the aforementioned findings it can be concluded that ECAP-processed billets exhibited enhanced wear behavior which can be attributed to the obtained UFG structure52.

SEM micrograph of the worn surface after the wear test: (ac) AA alloy; (b) ECAP-processed through 4Bc.

Several regression transformations approach and associations among variables that are independent have been investigated in order to model the wear output responses. The association between the supplied parameters and the resulting responses was modeled using quadratic regression. The models created in the course of the experiment are considered statistically significant and can be used to forecast the response parameters in relation to the input control parameters when the highest possible coefficient of regression of prediction (R2) is closer to 1. The regression Eqs.(9)(14) represent the predicted non-linear model of volume loss (VL) and coefficient and friction (COF) at different passes as a function of velocity (V) and applied load (P), with their associated determination and adjusted coefficients. The current studys adjusted R2 and correlation coefficient R2 values fluctuated between 95.67 and 99.97%, which is extremely near to unity.

$${text{AA }}left{ {begin{array}{*{20}l} {VL = + 1.52067 times 10^{ - 6} - 1.89340 times 10^{ - 9} P - 4.81212 times 10^{ - 11} V + 8.37361 times 10^{ - 12} P * V} hfill & {} hfill \ { - 2.91667E - 10 {text{P}}^{2} - 2.39989E - 14 {text{V}}^{2} } hfill & {(9)} hfill \ {frac{1}{{{text{COF}}}} = + 2.72098 + 0.278289P - 0.029873V - 0.000208 P times V + 0.047980 {text{P}}^{2} } hfill & {} hfill \ { + 0.000111 {text{V}}^{2} - 0.000622 {text{P}}^{2} times V + 6.39031 times 10^{ - 6} P times {text{V}}^{2} } hfill & {(10)} hfill \ end{array} } right.$$

$$1{text{ Pass }}left{ {begin{array}{*{20}l} {VL = + 2.27635 times 10^{ - 7} + 7.22884 times 10^{ - 10} P - 2.46145 times 10^{ - 11} V - 1.03868 times 10^{ - 11} P times V} hfill & {} hfill \ { - 1.82621 times 10^{ - 10} {text{P}}^{2} + 6.10694 times 10^{ - 14} {text{V}}^{2} } hfill & {} hfill \ { + 8.76819 times 10^{ - 13} P^{2} times V + 2.48691 times 10^{ - 14} P times V^{2} } hfill & {(11)} hfill \ {frac{1}{{{text{COF}}}} = - 0.383965 + 1.53600P + 0.013973V - 0.002899 P times V} hfill & {} hfill \ { - 0.104246 P^{2} - 0.000028 V^{2} } hfill & {(12)} hfill \ end{array} } right.$$

$$4{text{ Pass}}left{ {begin{array}{*{20}l} {VL = + 2.29909 times 10^{ - 8} - 2.29012 times 10^{ - 10} P + 2.46146 times 10^{ - 11} V - 6.98269 times 10^{ - 12} P times V } hfill & {} hfill \ { - 1.98249 times 10^{ - 11} {text{P}}^{2} - 7.08320 times 10^{ - 14} {text{V}}^{2} } hfill & {} hfill \ { + 3.23037 times 10^{ - 13} P^{2} * V + 1.70252 times 10^{ - 14} P times V^{2} } hfill & {(13)} hfill \ {frac{1}{{{text{COF}}}} = + 2.77408 - 0.010065P - 0.020097V - 0.003659 P times V} hfill & {} hfill \ { + 0.146561 P^{2} + 0.000099 V^{2} } hfill & {(14)} hfill \ end{array} } right.$$

The experimental data are plotted in Fig.7 as a function of the corresponding predicted values for VL and COF for zero pass, one pass, and four passes. The minimal output value is indicated by blue dots, which gradually change to the maximum output value indicated by red points. The effectiveness of the produced regression models was supported by the analysis of these maps, which showed that the practical and projected values matched remarkably well and that the majority of their intersection locations were rather close to the median line.

Comparison between VL and COF of experimental and predicted values of ZK30 at AA, 1P, and 4Bc.

As a consequence of wear characteristics (P and V), Fig.8 displays 3D response plots created using regression models to assess changes in VL and COF at various ECAP passes. For VL, the volume loss and applied load exhibit an inverse proportionality at various ECAP passes, which is apparent in Fig.8ac. It was observed that increasing the applied load in the wear process will minimize VL. So, the optimal amount of VL was obtained at an applied load of 5N. There is an inverse relation between V of the wear process and VL at different ECAP passes. There is a clear need to change wear speeds for bullets with varying numbers of passes. As a result, the increased number of passes will need a lower wear speed to minimize VL. The minimal VL at zero pass is 1.50085E06 m3 obtained at 5N and 250mm/s. Also, at a single pass, the optimal VL is 2.2266028E07 m3 obtained at 5 N and 148mm/s. Finally, the minimum VL at four passes is 2.07783E08 m3 at 5N and 64.5mm/s.

Three-dimensional plot of VL (ac) and COF (df) of ZK30 at AA, 1P, and 4Bc.

Figure8df presents the effect of wear parameters P and V on the COF for ECAPed ZK30 billets at zero, one, and four passes. There is an inverse proportionate between the applied load in the wear process and the coefficient of friction. As a result, the minimum optimum value of COF of the ZK30 billet at different process passes was obtained at 5 N. On the other hand, the speed used in the wear process decreased with the number of billet passes. The wear test rates for billets at zero, one, and four passes are 250, 64.5, and 64.5mm/s, respectively. The minimum COF at zero pass is 0.380134639, obtained at 5N and 250mm/s. At 5N and 64.5mm/s, the lowest COF at one pass is 0.220277466. Finally, the minimum COF at four passes is 0.23130154 at 5N and 64.5mm/s.

The previously mentioned modern ML algorithms have been used here to provide a solid foundation for analyzing the obtained data and gaining significant insights. The following section will give the results acquired by employing these approaches and thoroughly discuss the findings.

The correlation plots and correlation coefficients (Fig.9) between the input variables, force, and speed, and the six output variables (VL_P0, VL_P1, VL_P4, COF_P0, COF_P1, and COF_P4) for data preprocessing of ML models give valuable insights into the interactions between these variables. Correlation charts help to investigate the strength and direction of a linear relationship between model input and output variables. We can initially observe if there is a positive, negative, or no correlation between each two variables by inspecting the scatterplots. This knowledge aids in comprehending how changes in one variable effect changes in the other. In contrast, the correlation coefficient offers a numerical assessment of the strength and direction of the linear relationship. It ranges from 1 to 1, with near 1 indicating a strong negative correlation, close to 1 indicating a strong positive correlation, and close to 0 indicating no or weak association. It is critical to examine the size and importance of the correlation coefficients when examining the correlation between the force and speed input variables and the six output variables (VL_P0, VL_P1, VL_P4, COF_P0, COF_P1, and COF_P4). A high positive correlation coefficient implies that a rise in one variable is connected with an increase in the other. In contrast, a high negative correlation coefficient indicates that an increase in one variable is associated with an increase in the other. From Fig.9 it was clear that for all ZK30 billets, the both VL and COP were reversely proportional with the applied (in the range of 1-up to- 5N). Regarding the wear speed, the VL of both the AA and 1P conditions exhibited an inversed proportional with the wear speed while 4Bc exhibited a direct proportional with the wear speed (in the range of 64.5- up to- 250mm/s) despite of the COP for all samples revealed an inversed proportional with the wear speed. The VL of AA condition (P0) revealed strong negative correlation coefficient of 0.82 with the applied load while it displayed intermediate negative coefficient of 0.49 with the wear speed. For 1P condition, VL showed a strong negative correlation of 0.74 with the applied load whereas it showed a very weak negative correlation coefficient of 0.13 with the speed. Furthermore, the VL of 4Bc condition displayed a strong negative correlation of 0.99 with the applied load while it displayed a wear positive correlation coefficient of 0.08 with the speed. Similar trend was observed for the COF, the AA, 1P and 4Bc samples displayed intermediate negative coefficient of 0.047, 0.65 and 0.61, respectively with the applied load while it showed a weak negative coefficient of 0.4, 0.05 and 0.22, respectively with wear speed.

Correlation plots of input and output variables showcasing the strength and direction of relationships between each inputoutput variable using correlation coefficients.

Figure10 shows the predicted train and test VL values compared to the original data, indicating that the VL prediction model performed well utilizing the LR (Linear Regression) technique. The R2-score is a popular statistic for assessing the goodness of fit of a regression model. It runs from 0 to 1, with higher values indicating better performance. In this scenario, the R2-scores for both the training and test datasets range from 0.55 to 0.99, indicating that the ML model has established a significant correlation between the projected VL values and the actual data. This shows that the model can account for a considerable percentage of the variability in VL values.

Predicted train and predicted test VL versus actual data computed for different applied loads and number of passes of (a) 0P (AA), (b) 1P, and (c) 4Bc: evaluating the performance of the VL prediction best model achieved using LR algorithm.

The R2-scores for training and testing three distinct ML models for the output variables VL_P0, VL_P1, and VL_P4 are summarized in Fig.11. The R2-score, also known as the coefficient of determination, is a number ranging from 0 to 1 that indicates how well the model fits the data. For VL_P0, R2 for testing is 0.69, and that for training is 0.96, indicating that the ML model predicts the VL_P0 variable with reasonable accuracy on unknown data. On the other hand, the R2 value of 0.96 for training suggests that the model fits the training data rather well. In summary, the performance of the ML models changes depending on the output variables. With R2 values of 0.98 for both training and testing, the model predicts 'VL_P4' with great accuracy. However, the models performance for 'VL_P0' is reasonable, with an R2 score of 0.69 for testing and a high R2 score of 0.96 for training. The models performance for 'VL_P1' is relatively poor, with R2 values of 0.55 for testing and 0.57 for training. Additional assessment measures must be considered to understand the models' prediction capabilities well. Therefore, as presented in the following section, we did no-linear polynomial fitting with extracted equations that accurately link the output and input variables.

Result summary of ML train and test sets displaying R2-score for each model.

Furthermore, the data was subjected to polynomial fitting with first- and second-degree models (Fig.12). The fitting accuracy of the data was assessed using the R2-score, which ranged from 0.92 to 0.98, indicating a good fit. The following equations (Eqs.15 to 17) were extracted from fitting the experimental dataset of the volume loss at different conditions of applied load (P) and the speed (V) as follows:

$${text{VL}}_{text{P}}0 = 1.519e - 06{ } + { } - 2.417e - 09{text{ * P }} + { } - 3.077e - 11{ * }V$$

(15)

$$VL_{text{P}}1 = 2.299e - 07 - 5.446e - 10 * {text{P}} - 5.431e - 11 * V - 5.417e - 11 * {text{P}}^{2} + 2.921e - 12 * {text{P}} V + 1.357e - 13 * V^{2}$$

(16)

$$VL_{text{P}}4 = 2.433e - 08 - 6.200e - 10 * {text{P}} + 1.042e - 12 * V$$

(17)

Predicted versus actual (a) VL_P0 fitted to Eq.15 with R2-score of 0.92, (b) VL_P1 fitted to Eq.16 with R2-score of 0.96, (c) VL_P4 fitted to Eq.17 with R2-score of 0.98.

Figure13 depicts the predicted train and test coefficients of friction (COF) values placed against the actual data. The figure seeks to assess the performance of the best models obtained using the SVM (Support Vector Machine) and GPR (Gaussian Process Regression) algorithms for various applied loads and number of passes (0, 1P, and 4P). The figure assesses the accuracy and efficacy of the COF prediction models by showing the predicted train and test COF values alongside the actual data. By comparing projected and actual data points, we may see how closely the models match the true values. The ML models trained and evaluated on the output variables 'COF_P0', 'COF_P1', and 'COF_P4' using SVM and GPR algorithms show great accuracy and performance, as summarized in Fig.13. The R2 ratings for testing vary from 0.97 to 0.99, showing that the models efficiently capture the predicted variables' variability efficiently. Furthermore, the training R2 scores are consistently high at 0.99, demonstrating a solid fit to the training data. These findings imply that the ML models can accurately predict the values of 'COF_P0', 'COF_P1', and 'COF_P4' and generalize well to new unseen data.

Predicted train and predicted test COF versus actual data computed for different applied loads and number of passes of (a) 0P (AA), (b) 1P, and (c) 4Bc: evaluating the performance of the COF prediction best model achieved using SVM and GPR algorithms.

Figure14 presents a summary of the results obtained through machine learning modeling. The R2 values achieved for COF modeling using SVM and GPR are 0.99 for the training set and range from 0.97 to 0.99 for the testing dataset. These values indicate that the models have successfully captured and accurately represented the trends in the dataset.

Result summary of ML train and test sets displaying R2-score for each model.

The results of the RSM optimization carried out on the volume loss and coefficient of friction at zero pass (AA), along with the relevant variables, are shown in Appendix A-1. The red and blue dots represented the wear circumstance (P and V) and responses (VL and COF) for each of the ensuing optimization findings. The volume loss and coefficient of friction optimization objective were formed to in range, using minimize as the solution target, and the expected result of the desirability function was in the format of smaller-is-better attributes. The values of (A) P=5 N and (B) V=250mm/s were the optimal conditions for volume loss. Appendix A-1(a) shows that this resulted in the lowest volume loss value attainable of 1.50127E-6 m3. Also, the optimal friction coefficient conditions were (A) P=2.911 N and (B) V=250mm/s. This led to the lowest coefficient of friction value possible, which was 0.324575, as shown in Appendix A-1(b).

Appendix A-2 displays the outcomes of the RSM optimization performed on the volume loss and coefficient of friction at one pass, together with the appropriate variables. The volume loss and coefficient of friction optimization objectives were designed to be "in range," with "minimize" as the solution objective. It was anticipated that the intended function would provide "smaller-is-better" traits. The ideal conditions for volume loss were (A) P=4.95 N and (B) V=136.381mm/s. This yielded the lowest volume loss value feasible of 2.22725E-7 m3, as seen in Appendix A-2 (a). The optimal P and V values for the coefficient of friction were found to be (A) P=5 N and (B) V=64.5mm/s. As demonstrated in Appendix A-2 (b), this resulted in the lowest coefficient of friction value achievable, which was 0.220198.

Similarly, Appendix A-3 displays the outcomes of the RSM optimization performed on the volume loss and coefficient of friction at four passes, together with the appropriate variables. The volume loss and coefficient of friction optimization objectives were designed to be "in range," with "minimize" as the solution objective. The desired functions expected result would provide of "smaller-is-better" characteristics. The optimal conditions for volume loss were (A) P=5 N and (B) V=77.6915mm/s. This yielded the lowest volume loss value feasible of 2.12638E-8 m3, as seen in Appendix A-1 (a). The optimal P and V values for the coefficient of friction were found to be (A) P=4.95612 N and (B) V=64.9861mm/s. As seen in Appendix A-1(b), this resulted in the lowest coefficient of friction value achievable, which was 0.235109.

The most appropriate combination of wear-independent factors that contribute to the minimal feasible volume loss and coefficient of friction was determined using a genetic algorithm (GA). Based on genetic algorithm technique, the goal function for each response was determined by taking Eqs.(9)(14) and subjecting them to the wear boundary conditions, P and V. The following expression applies to the recommended functions for objective: Minimize (VL, COF), subjected to ranges of wear conditions: 1P5 (N), 64.5V250 (mm/s).

Figures15 and 16 show the GA optimization techniques performance in terms of fitness value and the running solver view, which were derived from MATLAB, together with the related wear requirements for the lowest VL and COF at zero pass. VL and COF were suggested to be minimized by Eqs.(9) and (10), which were then used as the function of fitness and exposed to the wear boundary limit. According to Fig.15a, the lowest value of VL that GA could find was 1.50085E6 m3 at P=5N and V=249.993mm/s. Furthermore, the GA yielded a minimum COF value of 0.322531 at P=2.91 N and V=250mm/s (Fig.15b).

Optimum VL (a) and COF (b) by GA at AA condition.

Optimum VL (a) and COF (b) by hybrid DOE-GA at AA condition.

The DOEGA hybrid analysis was carried out to enhance the GA outcomes. Wear optimal conditions of VL and COF at zero pass are used to determine the initial populations of hybrid DOEGA. The hybrid DOEGA yielded a minimum VL value of 1.50085E-6 m3 at a speed of 249.993mm/s and a load of 5N (Fig.16a). Similarly, at a 2.91 N and 250mm/s speed load, the hybrid DOEGA yielded a minimum COF (Fig.16b) of 0.322531.

The fitness function, as defined by Eqs.11 and 12, was the depreciation of VL and COF at a 1P, subject to the wear boundary condition. Figure17a,b display the optimal values of VL and COF by GA, which were 2.2266E7 m3 and 0.220278, respectively. The lowest VL measured at 147.313mm/s and 5 N. In comparison, 5 N and 64.5mm/s were the optimum wear conditions of COF as determined by GA. Hybrid DOEGA results of minimum VL and COF at a single pass were 2.2266 E-7 m3 and 0.220278, respectively, obtained at 147.313mm/s and 5 N for VL as shown in Fig.18a and 5 N and 64.5mm/s for COF as shown in Fig.18b.

Optimum VL (a) and COF (b) by GA at 1P condition.

Optimum VL (a) and COF (b) by hybrid DOE-GA at 1P condition.

Subject to the wear boundary condition, the fitness function was the minimization of VL and COF at four passes, as defined by Eqs.13 and 14. The optimum values of VL and COF via GA shown in Fig.19a,b were 2.12638E8 m3 and 0.231302, respectively. The lowest reported VL was 5 N and 77.762mm/s. However, GA found that the optimal wear conditions for COF were 5 N and 64.5mm/s. In Fig.20a,b, the hybrid DOEGA findings for the minimum VL and COF at four passes were 2.12638E8 m3 and 0.231302, respectively. These results were achieved at 77.762mm/s and 5 N for VL and 5 N and 64.5mm/s for COF.

Optimum VL (a) and COF (b) by GA at 4Bc condition.

Optimum VL (a) and COF (b) by hybrid DOE-GA at 4Bc condition.

A mathematical model whose input process parameters influence the quality of the output replies was solved using the multi-objective genetic algorithm (MOGA) technique54. In the current study, the multi-objective optimization using genetic algorithm (MOGA) as the objective function, regression models, was implemented using the GA Toolbox in MATLAB 2020 and the P and V are input wear parameter values served as the top and lower bounds, and the number of parameters was set to three. After that, the following MOGA parameters were selected: There were fifty individuals in the initial population, 300 generations in the generation, 20 migration intervals, 0.2 migration fractions, and 0.35 Pareto fractions. Constraint-dependent mutation and intermediary crossover with a coefficient of chance of 0.8 were used for optimization. The Pareto optimum, also known as a non-dominated solution, is the outcome of MOGA. It is a group of solutions that consider all of the objectives without sacrificing any of them55.

By addressing both as multi-objective functions was utilized to identify the lowest possible values of the volume loss and coefficient of friction at zero pass. Equations(9) and (10) were the fitness functions for volume loss and coefficient of friction at zero pass for ZK30. The Pareto front values for the volume loss and coefficient of friction at zero pass, as determined by MOGA, are listed in Table 2. The volume loss (Objective 1) and coefficient of friction (Objective 2) Pareto chart points at zero pass are shown in Fig.21. A friction coefficient reduction due to excessive volume loss was observed. As a result, giving up a decrease in the coefficient of friction can increase volume loss. For zero pass, the best volume loss was 1.50096E06 m3 with a sacrifice coefficient of friction of 0.402941. However, the worst volume loss was 1.50541E06 m3, with the best coefficient of friction being 0.341073.

The genetic algorithm was used for the multi-objective functions of minimal volume loss and coefficient of friction. The fitness functions for volume loss and coefficient of friction at one pass were represented by Eqs.(11) and (12), respectively. Table 3 displays the Pareto front points of volume loss and coefficient of friction at one pass. Figure22 presents the volume loss (Objective 1) and coefficient of friction (Objective 2) Pareto chart points for a single pass. It was discovered that the coefficient of friction decreases as the volume loss increases. As a result, the volume loss can be reduced at the expense of a higher coefficient of friction. The best volume loss for a single pass was 2.22699E07 m3, with the worst maximum coefficient of friction being 0.242371 and the best minimum coefficient of friction being 0.224776 at a volume loss of 2.23405E07 m3.

The multi-objective functions of minimal volume loss and coefficient of friction were handled by Eqs.(13) and (14), respectively, served as the fitness functions for volume loss and coefficient of friction at four passes. The Pareto front points of volume loss and coefficient of friction at four passes are shown in Table 4. The Pareto chart points for the volume loss (Objective 1) and coefficient of friction (Objective 2) for four passes are shown in Fig.23. It was shown that when the volume loss increases, the coefficient of friction lowers. The volume loss can be decreased as a result, however, at the expense of an increased coefficient of friction. The best minimum coefficient of friction was 0.2313046 at a volume loss of 2.12663E08 m3, and the best minimum volume loss was 2.126397E08 m3 at a coefficient of friction of 0.245145 for four passes. In addition, Table 5 compares wear response values at DOE, RSM, GA, hybrid RSM-GA, and MOGA.

This section proposed the optimal wear parameters of different responses, namely VL and COF of ZK30. The presented optimal wear parameters, such as P and V, are based on previous studies of ZK30 that recommended the applied load from one to 30 N and speed from 64.5 to 1000mm/s. Table 6 presents the optimal condition of the wear process of different responses by genetic algorithm (GA).

Table 7 displays the validity of wears regression model for VL under several circumstances. The wear models' validation was achieved under various load and speed conditions. The volume loss response models had the lowest error % between the practical and regression models and were the most accurate, based on the validation data. Table 7 indicates that the data unambiguously shows that the predictive molding performance has been validated, as shown by the reasonably high accuracy obtained, ranging from 69.7 to 99.9%.

Equations(15 to 17) provide insights into the relationship that links the volume loss with applied load and speed, allowing us to understand how changes in these factors affect the volume loss in the given system. The validity of this modeling was further examined using a new unseen dataset by which the prediction error and accuracy were calculated, as shown in Table 8. Table 8 shows that the data clearly demonstrates that the predictive molding performance has been validated, as evidenced by the obtained accuracy ranging from 69.7 to 99.9%, which is reasonably high.

Read the original post:
Optimization of wear parameters for ECAP-processed ZK30 alloy using response surface and machine learning ... - Nature.com

Cancer-associated genes and essentiality scores

We first determined whether cancer-related genes are likely to have high essentiality scores. We aggregated several essentiality scores calculated by multiple metrics5 for the list of genes identified in the COSMIC Census database (Oct 2018) and for all other human protein coding genes. Two different approaches to scoring genes essentiality are available. The first group of methods calculates the essentiality scores by measuring the degree of loss of function caused by a change (represented by variation detection) in the gene. It uses the following methods: residual variation intolerance score (RVIS), LoFtool, Missense-Z, the probability of loss-of-function intolerance (pLI) and the probability of haplo-insufficiency (Phi). The second group (Wang, Blomen and Hart- EvoTol) studies the impact of variation on cell viability. For all methods above measuring essentiality, a higher score indicates a higher degree of essentiality. Each method is described in detail in5.

We find that on average the cancer genes exhibit a higher degree of essentiality compared to the average scores calculated for all protein coding human genes and all metrics (Fig.1). We find that genes associated with cancer have higher essentiality scores on average in both categories (intolerance to variants and cell line viability), compared to the average scores across all human genes. P values are consistently<0.00001 (Table 1).

We also investigated whether Tumor Suppressor Genes (TSGs) or Oncogenes as distinct groups of genes would show different degrees of essentiality. (If the gene is known to be both an oncogene and a TSG, then the essentiality score of that gene would be present in both the oncogene and the TSG groups). We found no significant differences in the degrees of essentiality on average for either group compared to the set of all cancer genes (Table 1; Fig.1).

The results are particularly of interest in the context of cancer, as essential genes have been shown to evolve more slowly than nonessential genes20,21,22, although some discrepancies have been reported22. A slower evolutionary rate indicates less probability to evolve resistance to a cancer drug. This is particularly important in the case of anticancer drugs as it was reported that these drugs cause a change in the selection pressure when administered, leading to increased drug resistance23.

This association between cancer-related genes and essentiality scores prompted us to develop methods to identify cancer-related genes using this information. We used a machine-learning approach. A range of open-source algorithms were applied and tested to produce the most accurate classifier. We focus on properties related to proteinprotein interaction networks, as essential genes are likely to encode hub proteins, i.e., those with highest degree values in the network21,24.

A total of nine different modelling approaches (or configurations) were run on the data to ensure the selection of the best performing approach (the list of these can be found in the Supplementary Information Table 2, along with their performance metrics). The performance metric used to rank the models was Logarithmic Loss (LogLoss), LogLoss is an appropriate and known performance measure when the model is of a binary-classification type. The LogLoss measures confidence of the prediction and estimates how to penalise incorrect classification. The selection mechanism for the performance metric takes the type of model (binary classification in this case) and distribution of values into consideration when recommending the performance metric. However, other performance metrics were also calculated (Supplementary Information Table 2). The performance metrics are calculated for all validation and test (holdout) sets to ensure that the model is not over-fitting. The particular model with best performance result (LogLoss) in this case was: eXtreme Gradient Boosted Trees Classifier with Early Stopping. The model shows very close LogLoss values for training/validation and holdout data sets (Table 2), demonstrating no over-fitting.

The model development workflow (i.e., the model blueprint) is shown in Fig.2. This shows the pre-processing steps and the algorithm used in our final model, and illustrates the steps involved in transforming input into a model. In this diagram, Ordinal encoding of categorical variables converts categorical variables to an ordinal scale while the Missing Values Imputed node imputes missing values. Numeric variables with missed values were imputed with an arbitrary value (default9999). This is effective for tree-based models, as they can learn a split between the arbitrary value (9999) and the rest of the data (which is far away from this value).

Model development stages.

To demonstrate the effectiveness of our model, a chart was constructed (Fig.3) that shows across the entire validation dataset (divided into 10 segments or bins and ordered by the average outcome prediction value) the average actual outcome (whether a gene has been identified as cancer gene or not) and the average predicted outcome for each segment of the data (order from lowest average to highest per segment). The left side of the curve indicates where the model predicted a low score on one section of the population while the right side of the curve indicates where the model predicted a high score. The "Predicted" blue line displays the average prediction score for the rows in that bin. The "Actual" red line displays the actual percentage for the rows in that bin. By showing the actual outcomes alongside the predictive values for the dataset, we can see how close these predictions are to the actual known outcome for each segment of the dataset. Also, we can determine if the accuracy diverges in cases where the outcome is confirmed as cancer or not, as the segments are ordered by their average of outcome scores.

The Lift Chart illustrating the accuracy of the model.

In general, the steeper the actual line is, and the more closely the predicted line matches the actual line, the better the model. A close relationship between these two lines is indicative of the predictive accuracy of the model; a consistently increasing line is another good indicator of satisfactory model performance. The graph we have for our model (Fig.3) thus indicates high accuracy of our prediction model.

In addition, the confusion matrix (Table 3) and the summary statistics (Table 4) show the actual versus predicted values for both true/false categories for our training dataset (80% of the total dataset). The model statistics show the model reached just over 89% specificity and 60% sensitivity in predicting cancer genes. This means that we are able to detect over half of cancer genes successfully while only misclassifying around 10% of non-cancer genes within the training/validation datasets. The summary statistics (Table 4) also shows the F1 score (harmonic mean of the precision and recall) and Matthews Correlation Coefficient (MCC is the geometric mean of the regression coefficient) for the model. The low F1 score reflects our choice to maximise the true negative rate (preventing significant misclassification of non-cancer genes).

To further confirm the models ability to predict cancer genes, we used the model on 190 new cancer genes that had been added to the COSMIC Cancer Census Genes between October 2018 and April 2020. Applying the model, we were able to predict 56 genes out of the newly added 190 genes as cancer genes, all of which were among the false positives detected by the model. This indicates that the model is indeed suitable to use to predict novel candidate cancer genes that could be experimentally confirmed later. A full ranked list of candidate genes predicted to be cancer associated by our model is available in Supplementary Information Table 3.

Another way to visualise the model performance, and determine the optimal score to use as a threshold between cancer and non-cancer genes, is the prediction distribution graph (Fig.4) which illustrates the distribution of outcomes. The distribution in purple shows the outcome where gene is not classified as a cancer gene while the second distribution in green shows the outcomes where gene is classified as a cancer gene. The dividing line represents the selected threshold at which the binary decision creates a desirable balance between true negatives and true positives. Figure4 shows how well our model discriminates between prediction classes (cancer gene or non-cancer gene) and shows the selected score (threshold) that could be used to make a binary (true/false) prediction for a gene to be classified as a candidate cancer gene. Every prediction to the left of the dividing line is classified as non-cancer associated and every prediction to the right of the dividing line is classified as cancer associated.

The prediction distribution graph showing how well the model discriminates between cancer and non-cancer genes.

The prediction distribution graph can be interpreted as follows: purple to the left of the threshold line is for instances where genes were correctly classified as non-cancer (true negatives). Green to the left of the threshold line is for instances were incorrectly classified as non-cancer (false negatives). Purple to the right of the threshold line, is for instances that were incorrectly classified as cancer gene (false positives). Green to the right of the threshold line, is for instances were correctly classified as cancer genes (true positives). The graph again confirms that the model was able to accurately distinguish cancer and non-cancer genes.

Using the receiver operating characteristic curve (ROC) curve produced for our model (Fig.5), we were able to evaluate the accuracy of prediction. The AUC (area under the curve) is a metric for binary classification that considers all possible thresholds and summarizes performance in a single value, with the larger the area under the curve, the more accurate the model. An AUC of 0.5 shows that predictions based on this model are no better than a random guess. An AUC of 1.0 shows that predictions based on this model are perfect. (This is highly uncommon and likely flawed, indicating some features that should not be known in advance are being used in model training and thus revealing the outcome.) As the area under the curve is of 0.86, we conclude that the model is accurate. The circle intersecting the ROC curve represents the threshold chosen for classification of genes. This is used to transform probability scores assigned to each gene into binary classification decisions, where each gene would be classified as a potential cancer gene or not.

The receiver operator characteristic (ROC) curve indicating model performance.

Feature impact measures how much worse a models error score would be if the model made predictions after randomly shuffling the values of one field input (while leaving other values unchanged) and thus shows how useful each feature is for the prediction. The scores were normalised so that the value of the most important feature column is 100% and the other subsequent features are normalised to it. This helps identify those properties that are particularly important in relation to predicting cancer genes and would aid in further our understanding of the biological aspects that might underline the propensity of a gene to be a cancer gene.

Closeness and degree are ranked as the properties with the highest feature impact (Fig.6). Both are proteinprotein interaction network properties, indicating a central role of the protein product within the network. We find that both correlate with likelihood of cancer association. Other important properties such the phi essentiality score (probability of haploinsufficiency compared to baseline neutral expectation) and Tajimas D regulatory (measures for genetic variation at intra-species level and for proportion of rare variants) show that increased essentiality accompanied with occurrence of rare variants increase the likelihood of pathological impact and for the gene to be linked to cancer initiation or progression. We also note that greater length of a gene or transcript increases the likelihood of a somatic mutation, so increasing the chance of a mutation within that gene, thus increasing the likelihood of it being a cancer gene.

The top properties ranked by their relative importance used to make the predictions by the model.

To confirm that the selected model performance is optimal based on the input data used, we created a new blended model combining the best 2nd and 3rd modelling approaches from all modelling approaches tested within our project and compared the performance metric (AUC) of our selected model with the new blended model. We found that improvement is small (0.008), despite the added complexity, where the blended model achieved an AUC of 0.866 and our single selected model achieved an AUC of 0.858.

We have also retrained our model using a dataset that excludes general gene properties and found that a reduction in models performance was evident but very small. The model trained on this dataset achieved an AUC of 0.835 and a sensitivity of 55% at a specificity of 89%. This small reduction in the predictability of the models indicates that essentiality and proteinprotein interaction network properties are the most important features predicting cancer genes and that information carried by gene general properties can be in most part be represented by information carried by these properties. This can be rationalised, as longer genes (median transcript length=3737) tend to have the highest number of proteinprotein interactions25.

According to a recent comprehensive review of cancer driver genes prediction models, currently the best performing machine learning model is driverMAPS with an AUC of 0.94, followed by HotNet2 with an AUC of 0.814. When comparing our model performance using AUC to the other 12 reviewed cancer driver genes prediction models, our model would come second with an AUC of 0.86. Our predictive model achieved better AUC measured performance when compared to the best model that used a similar network based approach (HotNet2 with AUC=0.81) and better than the best function-based prediction model (MutPanning with AUC=0.62). The strong performance of our model indicates the importance of combining different and distinctive gene properties, when building prediction models, while avoiding reliance on the frequency approach that could mask important driver genes that were detected in fewer samples. Despite the apparent success and high AUC score reported by our model, this should be treated with some caution. The AUC value is based on the ROC curve which is constructed by varying the threshold and then plotting the resulting sensitivities against the corresponding false positive rates. Several statistical methods are available to use to compare two AUC results and determine if the difference is significant26,27,28. These methods require the ranking of the variables in its calculations (e.g., to calculate the variance or covariance of the AUC). The ranking of predicated cancer associated genes was not available from all the other 12 cancer driver genes prediction methods. Thus, we were not able to determine whether the difference between the AUC score of our method and the AUC scores of these methods is significant.

The driverMAPS (Model-based Analysis of Positive Selection) method (the only method with higher AUC compared to our model) identifies cancer candidate genes using the assumption that these genes would exhibit elevated mutation rates in functionally important sites29. Thus, driverMAPS combines frequency- and function-based principles. Unlike our model that uses certain cohorts of genes properties, the parameters used in driverMAPS are mainly derived and estimated from factors influencing positive selection on somatic mutations. However, there are few features in common between the two models, such as dN/dS.

Despite driverMAPS had the overall best performance, network-based methods (like our method) showed much higher sensitivity than driverMAPS therefore potentially making more them more suited to distinguish cancer driver from non-driver genes. The driverMAPS paper29 provides a list of novel driver genes. We found that 35% of these novel candidate genes were also predicted by our model. Differenced in genes identified as cancer-related in the two approaches could be attributed to the different nature of features used by the two models. We believe that there is evidence30 pointing to genes with low mutation rates, but with important roles in driving the initiation and progression of tumours. Genes with high mutation rates were also shown to be less vital than expected in driving tumor initiation31. This variability in the mutation rate correlation with identified driver genes might explain some genes that our model does not identify as cancer-related genes where driverMaps does. Our model uses properties that are available for most protein coding genes, while driverMaps applies to genes already identified in tumour samples and predicts their likelihood to be driver cancer genes. Thus, the candidate list of genes provided by driverMaps is substantially smaller than our list. Using an ensemble method that evaluates both driverMAPS score and our models score for each gene, may produce more a reliable outcome. This would require further validation.

Enriching the models training dataset with added properties that show correlations with oncogenes could enhance the model prediction ability and elevate further the accuracy of the model. One potential feature is knowing whether a gene is an Ohnolog gene.

Paralogs retained from whole genome duplications (WGD) events that have occurred in all vertebrates, some 500 Myr ago are called ohnologs after Susumu Ohno32. Ohnologs have been shown to be prone to dominant deleterious mutations and frequently implicated in cancer and genetic diseases32. We investigated the enrichment of ohnologs within cancer-associated genes. Ohnolog genes can be divided into three sets: strict, intermediate and relaxed. These three sets are constructed using statistical confidence criteria32 . We found that 44% of the total number of cancer-associated genes (as reported in COSMIC census) belongs to an ohnolog family (using strict and intermediate thresholds). Considering that 20% of all known human genes are ohnologs (strict and intermediate) and that cancer-associated genes comprise less than 4% of all human genes, the enrichment of ohnolog genes with cancer-related genes is two times higher than expected. If only ohnologs that pass the strict threshold were considered, the fraction of cancer-related genes that are ohnologs is still high at 34%.

When performing pathway analysis (carried out using PANTHER gene ontology release 17.0), we found that cancer associated ohnologs show statistically significant enrichment (>tenfold) in many pathways and particularly within signalling pathways known to be cancer associated such as Jak/STAT, RAS and P53 (Supplementary Information Table 4). On the other hand, ohnologs that are not cancer associated are present in fewer signalling pathways and at enrichment (

The rest is here:
Essentiality, proteinprotein interactions and evolutionary properties are key predictors for identifying cancer ... - Nature.com

Enter the world of ML development services, a land where everything is in constant change due to technological advancements and data-driven innovation solutions.

In recent years, ML has become a groundbreaking technology that revolutionized various sectors such as health care, finances and transportation among others. The demand for ML development services has been growing at an extremely fast pace due to the rise of digitization that is taking place in various companies and doesnt seem like it will reduce any time soon. However, what is the future of machine learning in this fast-growing field? In this post, we will analyze the newest tendencies and make some forecasts on how ML development companies may change our world in a few years. Prepare for an adventurous journey into the world of existing technologies and their future possibilities!

First, we will address the described tendencies and forecasts without going into deeper details regarding why machine learning is gaining popularity in todays digital reality. This usefulness can be credited to the unmatched capacity to process vast tracts of data and make inferences or choices devoid of software. The advent of big data brought some enormous opportunities and challenges, high on the list of which is my favorite technology machine learning (ML). Importantly, it has already disrupted sectors such as healthcare services and finance industries especially when artificial intelligence is applied. Nevertheless, other applications of this technology are almost limitless to various areas and beyond; thus displaying the broad range of influence that transformative machine learning has.

Recently, there has been a significant increase in cloud-based machine learning capabilities. Most vendors, enterprises or individuals will find these platforms to be cost-effective means of deploying ML-based applications. Cloud-based solutions for the development of ML have three main benefits scalability, availability and automation. They provide an opportunity for developers to apply complex ML models and do not distract attention from important infrastructure details. In addition, the ML cloud platforms contain many tools and APIs for pre built models that result in development speed faster. The industry-wide adoption of ML-oriented products has determined the development of cloud-based platforms where solutions based on machine learning can be constructed. Because technology is developing every single day, we can assume that in future these platforms are going to be more complicated and provide developers with better choices of options as well as skills for AI.

With the above great leaps in machine learning for developers, there have been increasing conversations surrounding one field and it is interpretability. In other words, producing outputs is not enough for AI; the developers and users must come to grips with how those results were arrived at or what factors are involved. It is especially important for such areas as healthcare or finance since decisions made by AI models can influence significantly there. As a result, there is an elevated need for the generation of models that are easily transparent and interpretable to the needs shown. This is such a key achievement in ensuring that Artificial Intelligence becomes reliable and answerable to everything it offers.

The business need for integration with other growing technologies is because technology continues to evolve at the rate of exponential function. Scalable development is supported by artificial intelligence solutions for machines in different remote locations as we can see the increased popularity among manufacturers through Industrial Internet manufacturing and distribution. By integrating the above technologies it becomes possible to develop new competencies, improved decision making as well enhanced customer service. However, in the modern market, it is no longer possible to perceive these emerging technologies as standalone elements but more so as a constituent of the technology within which they operate. Integration strategy will result in development by a business or the adoption of some other software that is there and they would eventually benefit from this because it makes things much easier for them.

https://www.thewatchtower.com/blogs_on/supervised-machine-learning-its-advantages

Increased demand for personalized and customized ML solutions: With more companies embracing the use of machine learning to have an upper hand, the demand for specially tailored solutions shall grow. This will hence demand that machine learning development services like N-ix.com customize their solutions according to the specific needs and preferences of each client. Advancements in natural language processing (NLP): However, NPL has certainly come a long way and it continues to organize new language machines increasingly with effectiveness. With further advancements that lie ahead, NLP will evolve to even higher levels offering more advanced conversational AI and text analysis in the future.

Continued focus on ethics: However, as AI technologies continue their blend into different sectors of human life and activities in general, there will be an increased interest regarding the ethical development and deployment principles related to these emerging systems. The concern for these companies that provide the standards and guidelines will be for the government to model their operations by strict ethical practices to establish trust with clients as a well-behaved entity. In conclusion, machine learning development services have no limit to their possibilities in the future. Technological progress and wider adoption of AI solutions will surely keep the development in the field actively progressing, turning ML into a sphere with no boundaries for growth and innovation. Machine learning has a transforming effect on the world that is happening right under our noses, and it is quite thrilling for business owners as well as developers.

The trend of ML development services has tremendously changed. With the emergence of big data as a rapidly advancing trend and increasing demands for intelligent software, developers need to change their direction rather fast. Currently, ML algorithms are developed for application in various sectors such as medical care services or the financial sphere and other areas. Given that firms are increasingly embracing the creative development of approaches geared towards the promotion and support for complete production value, as well as other client relations enhancing such a trend is bound to be here with us. It is also clear that, as the demand for ML development services rises, there will be an increased number of innovative solutions to offer businesses a competitive edge. While much about ML remains unknown, there is no denying that such technologies have the potential to reform our lives and business operations.

Go here to see the original:
The Future of ML Development Services: Trends and Predictions - FinSMEs