Cloud Hosting

In this article we try to understand how machine learning can be beneficial for predicting disease outbreaks

By 2050, the 7.8 billion people who currently live on the planet are projected to number 9.7 billion. Unfortunately, the rate of infectious diseases is rising as a result of population rise. The development of diseases is influenced by several variables. These include urbanization, globalization, and climate change, and most of these phenomena are to some degree a result of human activity. Pathogens themselves may be prone to emerging, and the emerging pathogens tend to have more quickly evolving viruses. When a virus from one person infects a different person or an animal, an infectious illness ensues.

It can impact society on a large scale, similar to the coronavirus COVID-19, and is consequently a significant societal issue. It is viewed as a societal issue since it not only negatively affects people but also negatively affects society as a whole. Therefore, identifying high-risk locations for fatal infectious and non-infectious disease outbreaks is crucial in order to undertake prediction and detection of deadly disease outbreaks and improve the effectiveness of the response to these deadly disease outbreaks. To stop the spread of catastrophic infectious disease outbreaks (like COVID-19), health officials can use a variety of machine learning (ML) technologies.

This may be achieved by utilizing machine learning algorithms for both forecasting and identifying lethal infectious diseases as well as for responding to them. In order to predict disease outbreaks, machine learning algorithms can be used to learn datasets that include details about known viruses, animal populations, human demographics, biology and biodiversity data, readily accessible physical infrastructures, cultural/social practices around the world, and the geographic locations of the diseases. For instance, Support Vector Machine (SVM) and Artificial Neural Network (ANN) models can be used to predict malaria outbreaks.

Average monthly precipitation, temperature, humidity, the total number of positive cases, the total number of Plasmodium Falciparum (pF) cases, and the binary values of the number of outbreaks each month Yes or No, as the models performance is evaluated using their predictors, Root Mean Square Error (RMSE), and Receiver Operating Characteristic (ROC). The machine learning techniques may be integrated into an intelligent system to assess or mine social media data for any indicators of any outliers associated to uncommon flu symptoms in order to build efficient diagnostic methods.

As an illustration, Chae et al. suggested using deep learning to forecast infectious illnesses. In their study, deep learning algorithm parameters are improved while simultaneously using social media data to improve detection performance. The metrics include things like the frequency of verified infectious disease diagnoses, the volume of daily Google searches, the volume of Twitter mentions of the illness, and the average temperature and humidity in South Korea as a whole. After a disease event is recognized, responding to infectious disease outbreaks requires making a swift, well-informed choice in order to minimize the harm caused by the effect of the disease outbreaks.

In order to forecast the location and pace of spread of a disease, machine learning techniques may also learn integrated multi-source data from travel itinerary, population, logistics, and epidemiological data. Machine learning techniques can be applied by medical professionals to enhance the delivery of current therapies and hasten the development of novel ones. For instance, to learn any medical data gathered by the hospitals, they may utilize deep learning algorithms to model massive data sets. For instance, data from coronavirus patient clinical testing may be used as input for machine learning models to help clinicians identify the disease more quickly.

View original post here:
Role of Machine Learning in Predicting Disease Outbreaks - Analytics Insight

Read More..

Deep learning is a branch of the machine learning model based on neural networks. In the other machine model, the data processing to find the meaningful features is often done manually or relying on domain expertise; however, deep learning can mimic the human brain to discover the essential features, increasing the model performance.

There are many applications for deep learning models, including facial recognition, fraud detection, speech-to-text, text generation, and many more. Deep learning has become a standard approach in many advanced machine learning applications, and we have nothing to lose by learning about them.

To develop this deep learning model, there are various library frameworks we can rely upon rather than working from scratch. In this article, we will discuss two different libraries we can use to develop deep learning models: PyTorch and Lighting AI. Lets get into it.

PyTorch is an open-source library framework to train deep-learning neural networks. PyTorch was developed by the Meta group in 2016 and has grown in popularity. The rise of popularity was thanks to the PyTorch feature that combines the GPU backend library from Torch with Python language. This combination makes the package easy to follow by the user but still powerful in developing the deep learning model.

There are a few standout PyTorch features that are enabled by the libraries, including a nice front-end, distributed training, and a fast and flexible experimentation process. Because there are many PyTorch users, the community development and investment were also massive. That is why learning PyTorch would be beneficial in the long run.

PyTorch building block is a tensor, a multi-dimensional array used to encode all the input, output, and model parameters. You can imagine a tensor like the NumPy array but with the capability to run on GPU.

Lets try out the PyTorch library. Its recommended to perform the tutorial in the cloud, such as Google Colab if you dont have access to a GPU system (although it could still work with a CPU). But, If you want to start in the local, we need to install the library via this page. Select the appropriate system and specification you have.

For example, the code below is for pip installation if you have a CUDA-Capable system.

After the installation finishes, lets try some PyTorch capabilities to develop the deep learning model. We will do a simple image classification model with PyTorch in this tutorial based on their web tutorial. We would walk on the code and have an explanation of what happened within the code.

First, we would download the dataset with PyTorch. For this example, we would use the MNIST dataset, which is the number handwritten classification dataset.

We download both the MNIST train and test datasets to our root folder. Lets see what our dataset looks like.

Every image is a single-digit number between zero and nine, meaning we have ten labels. Next, lets develop an image classifier based on this dataset.

We need to transform the image dataset into a tensor to develop a deep learning model with PyTorch. As our image is a PIL object, we can use the PyTorch ToTensor function to perform the transformation. Additionally, we can automatically transform the image with the datasets function.

By passing the transformation function to the transform parameter, we can control what the data would be like. Next, we would wrap the data into the DataLoader object so the PyTorch model could access our image data.

In the code above, we create a DataLoader object for the train and test data. Each data batch iteration would return 64 features and labels in the object above. Additionally, the shape of our image is 28 * 28 (height * width).

Next, we would develop the Neural Network model object.

In the object above, we create a Neural Model with few layer structure. To develop the Neural Model object, we use the subclassing method with the nn.module function and create the neural network layers within the__init__.

We initially convert the 2D image data into pixel values inside the layer with the flatten function. Then, we use the sequential function to wrap our layer into a sequence of layers. Inside the sequential function, we have our model layer:

By sequence, what happens above is:

Lastly, the forward function is present for the actual input process for the model. Next, the model would need a loss function and optimization function.

For the next code, we just prepare the training and test preparation before we run the modeling activity.

Now we are ready to run our model training. We would decide how many epochs (iterations) we want to perform with our model. For this example, lets say we want it to run for five times.

The model now has finished their training and able to be used for any image prediction activity. The result could vary, so expect different results from the above image.

Its just a few things that PyTorch can do, but you can see that building a model with PyTorch is easy. If you are interested in the pre-trained model, PyTorch has a hub you can access.

Lighting AI is a company that provides various products to minimize the time to train the PyTorch deep learning model and simplify it. One of their open-source product is PyTorch Lighting, which is a library that offers a framework to train and deploy the PyTorch model.

Lighting offers a few features, including code flexibility, no boilerplate, minimal API, and improved team collaboration. Lighting also offers features such as multi-GPU utilization and swift, low-precision training. This made Lighting a good alternative to develop our PyTorch model.

Lets try out the model development with Lighting. To start, we need to install the package.

With the Lighting installed, we would also install another Lighting AI product called TorchMetrics to simplify the metric selection.

With all the libraries installed, we would try to develop the same model from our previous example using a Lighting wrapper. Below is the whole code for developing the model.

Lets break down what happen in the code above. The difference with the PyTorch model we developed previously is that the NNModel class now uses subclassing from the LightingModule. Additionally, we assign the accuracy metrics to assess using the TorchMetrics. Then, we added the training and testing step within the class and set up the optimization function.

With all the models set, we would run the model training using the transformed DataLoader object to train our model.

With the Lighting library, we can easily tweak the structure you need. For further reading, you could read their documentation.

PyTorch is a library for developing deep learning models, and it provides an easy framework for us to access many advanced APIs. Lighting AI also supports the library, which provides a framework to simplify the model development and enhance the development flexibility. This article introduced us to both the library's features and simple code implementation.Cornellius Yudha Wijaya is a data science assistant manager and data writer. While working full-time at Allianz Indonesia, he loves to share Python and Data tips via social media and writing media.

Go here to see the original:
Introduction to Deep Learning Libraries: PyTorch and Lightning AI - KDnuggets

Read More..

Biomedical research aimed at improving human health is particularly reliant on publicly funded basic science, according to a new analysis boosted by artificial intelligence.

B. Ian Hutchins

What we found is that even though research funded by the National Institutes of Health makes up 10% of published scientific literature, those published papers account for about 30% of the substantive research the important contributions supporting even more new scientific findings cited by further clinical research in the same field, says B. Ian Hutchins, a professor in the University of WisconsinMadisons Information School, part of the School of Computer, Data & Information Sciences. Thats a pretty big over-representation.

Hutchins and co-authors Travis Hoppe, now a data scientist at the Centers for Disease Control and Prevention, and UWMadison graduate student Salsabil Arabi, published their findings recently in the Proceedings of the National Academy of Sciences.

Published research papers typically include lengthy sections citing all the previous work supporting or referenced within the study. Predicting substantive biomedical citations without full text, the paper by Hutchins and Hoppe that you are reading about right now, cited no fewer than 64 other studies and sources in its References section.

Citations represent the transfer of knowledge from one scientist (or group of scientists) to another. Citations are extensively catalogued and tracked to measure the significance of individual studies and of the individuals conducting them, but not all citations included in any given paper make equally important contributions to the research they describe.

Were taught that as scientists, when we make a factual claim, were supposed to back it up with some kind of empirical evidence, Hutchins says. Like in Wikipedia entries, you cant have the little citation needed here flag. You have to add that citation. But if that fact youre citing isnt actually describing key prior work that you built upon, then it doesnt really support the interpretation that the citation represents a necessary earlier step toward your results.

Hutchins and his collaborators figured citations added later in the publication process, like those that appear at the behest of peer reviewers the subject-matter experts that evaluate scientific papers submitted to journals are less likely to have been truly important to the authors research.

If youre building on other peoples work, you probably identify that work earlier on in the research process, Hutchins says. That doesnt mean all the references that are in an early version of the manuscript are important ones, but the important ones are probably more concentrated in that earlier version.

To make the early-late distinction, the researchers trained a machine learning algorithm to judge citations on their importance by feeding it citation information from a pool of more than 38,000 scholarly papers. Each papers citation data came in two versions: a preprint version, posted publicly before peer review, and the eventual published version that had undergone peer review.

The algorithm found patterns to help identify the citations that were more likely to be important to each piece of published science. Those results revealed NIH-funded basic biological science appearing in the weightier citations at a rate three times the size of its share of all published research.

Federal funding for basic research is under constant scrutiny from members of the public and congressional leadership, Hutchins says. This gives us some evidence, not just anecdotes, that this kind of basic research funding is really important for stimulating the kind of clinical research treatments and cures for people that Congress tends to be more receptive to funding.

Read more:
Machine learning analysis of research citations highlights ... - University of Wisconsin-Madison

Read More..

Study design

We conducted a retrospective study utilizing prospectively recorded Japanese Utstein-style EMS activity records. The Ethics Committee of Nara Medical University approved the study (No. 3353), and the requirement for informed consent was waived owing to the use of anonymized records. This study was conducted in accordance with the tenets of the Declaration of Helsinki.

Japan has an aging population as 28.9% of its 130 million people are aged>65years10. The country consists of 47 prefectures with varying population densities of 65.46,399.5 individuals/km2. EMSs respond to all emergency calls and transport approximately 125,000 patients with OHCA to hospitals annually11. Emergency protocols, based on the Japanese Resuscitation Councils Resuscitation Guidelines12 and revised every 5years, are developed and implemented by 250 regional health managers. Each medical control region is supervised by a council established in each prefecture, tailoring protocols to local conditions13,14,15. EMS activities are recorded in the Utstein style and verified by the medical control council, and all records are collected annually by the Fire and Disaster Management Agency11. Our analysis included prehospital records of patients with OHCA resuscitated by EMS and transported to hospitals in 47 prefectures between 2015 and 2020, excluding patients aged<18years and those with non-cardiogenic cardiopulmonary arrest to reduce pathology variability.

In Japan, EMS is activated via a Communications Command Center upon receiving emergency calls. Bystanders may be instructed to administer cardiopulmonary resuscitation (CPR) over the telephone if cardiac arrest is suspected. Each ambulance includes a team of three, often featuring emergency life-saving technicians capable of advanced airway management and adrenaline administration for OHCA, under online medical control supervision. Additionally, hospital destinations are determined during field operations, and all patients, barring those with evident signs of death, are transported to a hospital.

We employed 23 factors and prefecture numbers from the Utstein-style EMS activity records as predictors, including county number, age, year and month of onset, bystander type, initial rhythm, number of defibrillations, number of adrenaline boluses administered, and elapsed time of each activity. Notably, the prefecture number was treated as a continuous variable due to its sequential allocation from north to south. This approach aimed to capture potential spatial correlations between adjacent prefectures. We also conducted a similar analysis using one-hot encoding for the prefecture numbers, and the outcomes did not contradict the results obtained when treating the prefecture number as a continuous variable. Categorical data were one-hot encoded. Remarkably, in the case of missing data, we refrained from substituting them with any particular value. Instead, the data missingness was coded as a separate category, which was incorporated into our analysis as a separate data element. Selected continuous variables were standardized using z-score normalization, a method that confers advantages in machine learning algorithms such as neural networks by aiding gradient descent convergence and mitigating issues related to weight initialization and gradient problems. Time factors, which were initially considered continuous variables, were one-hot encoded as categorical data16 because of their non-linear relationship with prognosis in cardiopulmonary resuscitation. The time factors were measured in minutes and thus represented as 1, 2, 3, 4, minutes.

Cases in which a specific intervention, such as defibrillation or drug administration, was not performed were also considered. These were coded as no intervention and incorporated into the contact-to-intervention column, allowing the model to reflect a comprehensive range of patient experiences. These steps resulted in 249 features (see Supplementary Table S1). Subsequently, we constructed a machine learning model to predict good neurological outcomes 1month after cardiac arrest, based on the cerebral performance category (CPC) score17a binary classification (Yes/No), with CPC1/2 signifying good neurological outcome and CPC3-5 indicating poor neurological outcomesourced from the Utstein records.

We stratified and randomly split the training and test datasets using an 8:2 ratio based on CPC1/2 to ensure a consistent ratio for predictive model construction. The prediction model was built using the neural network with the best average class sensitivity after several machine learning model trials. The compared methods included logistic regression, support vector machine, decision tree, random forest, and LightGBM9. To balance model bias (underfitting) and variance (overfitting), we applied a stratified cross-validation method (five-fold) using CPC1/2, along with batch normalization and dropouts in each neural network layer. The models accuracy plateaued after increasing the number of layers to five because of which we used a five-layer network to optimize learning costs. The sigmoid function served as the activation function and binary cross-entropy served as the loss function18. We measured model performance using area under the receiver operating characteristic curve (AUROC) and accuracy during training.

Imbalanced datasets significantly affect minority class performance. To address misclassification, we simulated based on predicted CPC1/2 numbers and employed class weighting during training to balance sensitivities, considering trade-offs. Our model aimed to maximize the majority class (CPC35) sensitivity without excessively reducing minority class (CPC1/2) sensitivity. We set CPC1/2 sensitivity at 80% and tested weights from 1 to 100 in 0.1 increments to optimize CPC3-5 sensitivity.

Additional training parameters included a batch size of 1,024,100 epochs, a learning rate of 0.001, and Adam optimizer. We conducted training using Python version 3.8.5 (Python Software Foundation, Beaverton, OR, USA).

We assessed the association of EMS activity duration with predicted CPC1/2 counts by simulating the constructed prediction model on a test dataset (n=92,108), containing all previously split prefectures from the training set. The simulation methodology involved three time factors: elapsed time from EMS arrival to hospital arrival (a), EMS arrival to first defibrillation (b), and EMS arrival to first drug administration (c).

Previous studies have shown that these temporal factors are important prognostic predictors of EMS activity time19,20,21,22,23,24,25,26. For example, shorter time from EMS arrival to defibrillation19,25 and from EMS arrival to drug administration20,21,22,23,24,25 are associated with better survival and improved neurological outcomes in OHCA patients. The prognostic impact of EMS providers staying on scene and performing their activities has also been reported26. Patients with non-shockable initial rhythm were excluded for (b), and those with EMS-witnessed cardiac arrest were excluded for (c). Time factors increased or decreased by5 to+5min for defibrillation and drug administration, and from5 to+10min for EMS arrival to hospital arrival time, in 1-min increments. We created a dataset adjusting each time factor in the test dataset and calculated the average predicted CPC1/2 score using the created prediction model. Then, we determined the percentage change in mean predicted CPC1/2 count to assess the association of time increase/decrease with the unadjusted data. We focused on percentage change relative to unadjusted data for a prefecture-specific analysis. A heat map visualized and evaluated the proportion of change between time adjustment and mean predicted CPC1/2 count.

We employed the same time adjustment method to estimate and visualize predicted CPC1/2 counts for the test dataset split by prefecture. We identified the time adjustments most associated with prognosis in each prefecture for the combinations (a) & (b) and (a) & (c), revealing treatment and EMS arrival to hospital arrival time adjustments with the greatest potential to improve predicted prognosis.

Patient characteristics are summarized as medians and interquartile ranges (IQRs) for continuous variables and counts and percentages for categorical variables. Additionally, the evaluation metric for the five models is expressed as meansstandard deviations. The standard deviations were calculated based on the variations in the evaluation metric across the five-fold cross-validation.

See the original post:
Machine learning-based analysis of regional differences in out-of ... - Nature.com

Read More..

Introduction

We are pleased to present Mehdi Rasul, a talented teenager excelling in his 11th-grade studies within the A-Level program at LANDAU School in Azerbaijan. Since a young age, he has been interested in Artificial Intelligence and Machine Learning, especially in Natural Language Processing and deep learning technologies. Over the years, Mehdi has become proficient in programming languages such as Python, Numpy, and Javascript. His exceptional academic performance is a testament to his unwavering quest for knowledge. Beyond the classroom, Mehdi's accomplishments extend to a diverse range of extracurricular activities, including community service, internships in various AI-related projects, active participation in MUN conferences, membership in the school newsletter, the completion of the IT Essentials program at Cisco Networking Academy, etc. In 2021 and 2022, Mr. Rasul won the competitions hosted by the "CyberMath Academy" and in 2022, he made a notable appearance in the American Mathematics Olympiad. Mehdi's research project has already been featured on different platforms and published in esteemed scientific journals in both Azerbaijan and Switzerland. His research delves into the comparison of diverse machine learning algorithms for sentiment analysis in both Azerbaijani and English texts. The study's findings underscore the pivotal role of national language corpora in achieving precise outcomes in natural language processing, particularly in sentiment analysis.

The article is provided below for your perusal.

The prediction of the sentiment of the text within different business spheres has been a challenging problem for various languages. Sentiment analysis, also known as opinion mining, is an active area of research in natural language processing (NLP) and computational linguistics that involves using text analysis and classification methods to identify and extract subjective information such as opinions, emotions, and attitudes from text data. Sentiment analysis has many applications in business, politics, and social media monitoring. Using sentimental analysis of textual data, companies can evaluate customer feedback, monitor reputation, forecast future and user behavior, etc., which can lead to better performance, efficiency, and increased profits.CREDIT: Mehdi Rasul

Studies have shown that machine learning algorithms, particularly those using supervised learning and deep learning, produce satisfactory results in the automation of textual sentiment analysis. Recent research illustrates that sentiment analysis of given texts can be automated, relying on the combination of computer science and mathematics, while also improving accuracy.

A well-defined English language corpus for model training helps build powerful and highly accurate models for English texts. Given the extensive array of libraries offering diverse natural language processing techniques in English, an automated system for analyzing texts and extracting meaningful insights or summaries from input paragraphs is relatively easy to create. The abundance of useful libraries in computer linguistics for English, combined with numerous natural language processing tools, such as word correction, grammar checking, text generation, word tokenization, etc., enables the construction of highly accurate sentiment analysis models.

Despite its utilization with major languages such as English, research on sentiment analysis of under-resourced languages like Azerbaijani is still relatively limited. Azerbaijani is the official language of the Republic of Azerbaijan and has distinct linguistic characteristics. However, it still lacks labeled datasets and the sophisticated language corpus for training machine learning models in sentiment analysis of the texts. Therefore, building sentiment analysis for Azerbaijani appears to be uniquely challenging, due to the lack of built-in NLP techniques and language corpus designed specifically for Azerbaijani.

This study encompasses a comprehensive comparison of machine learning algorithms. These algorithms are evaluated to compare their effectiveness in achieving sentiment classification for Azerbaijani. The same algorithms were tested on the English version of the dataset to address the importance of the built-in language corpus and the advancement of currently applicable techniques for NLP tasks in Azerbaijani.

Recent studies have shown that using machine learning techniques, such as supervised and deep learning algorithms, has contributed to significant advancements in improving and automating sentiment analysis for textual data. In politics, machine learning is also used to determine the sentiment of texts. In 2016, Heredia et al. collected political tweets to predict the U.S. 2016 election. Researchers collected three million location-based tweets related to Donald Trump and Hillary Clinton and trained them on the deep convolutional neural network (CNN) to predict the election results and attained an accuracy score of 84%.

The sentiment analysis of the texts in Azerbaijani is limited due to the lack of the labeled dataset and the language corpus. In 2013, research by Neethu and Rajasree on sentiment analysis of Twitter using machine learning algorithms produced satisfactory results in terms of classification of tweets into positive and negative classes.

Generally, machine learning algorithms perform well in the classification of texts in both Azerbaijani and English. The major algorithms utilized in text classification problems are Support Vector Machines, Nave Bayes, and Logistic Regression, which help to identify the patterns in both type and sentiments. Decision trees are supervised machine-learning algorithms used for classification and regression tasks. The goal is to create a model that predicts the value of a target variable by using decision rules created during the training process from the input features.

In this research, various machine learning algorithms were used to predict the sentiment of movie reviews in both English and Azerbaijani. The study has indicated the achievements attained in building models with various techniques. TF-IDF and BOW (or Count Vectorizer) have been implemented for feature extraction methods from the texts and tested in different models, including Logistic Regression, Nave Bayes, SVM, Decision Tree, Random Forest, AdaBoost, and XGBoost.

For the Azerbaijani version of the dataset, using the TF-IDF feature extraction approach, Logistic Regression and SVM algorithms produced better results compared to other models. However, the decision tree had 64% accuracy. When BOW feature extraction was used, Logistic Regression produced 84% accuracy, while SVM reached an accuracy of 82%.

The same dataset in English has also been modeled to compare the results with different preprocessing techniques that were not applied in Azerbaijani, such as stemming, stopwords list, etc. Overfitting was much less of an issue for models based on English datasets. The highest score was attained with the BOW feature extraction method using the Logistic Regression method. Additionally, SVM and Naive Bayes algorithms performed well with the TF-IDF feature extraction method and achieved an 85% accuracy score. The research indicates that the results are similar for both language models.

Python has enriched libraries for building language models in English and includes a list of stopwords, word tokenization, stemming, and lemmatization techniques. The English language corpus is well developed, which helps achieve higher results, unlike Azerbaijani, which lacks the language corpus, making it harder to build generalized models. Additional resources should be dedicated to research and creating a language corpus for Azerbaijani, which would help attain greater accuracy.

This article was written in cooperation with Mehdi Rasul

Visit link:
Azerbaijani young talent: From computer science to AI - The Jerusalem Post

Read More..