Cloud Hosting

Introduction

About 4.5 million electroencephalogram (EEG) tests are performed in the US each year. Thats more than if every person in Oregon, Connecticut, or Iowa got an EEG. Compared to magnetic resonance imaging (MRI) scans, which use magnetic fields and radio waves to generate images of the structure of the brain, EEGs use wires placed on the scalp to record the brain function as seen through the electrical activity that the brain generates in the process of neurons in the brain sending signals to each other. The Brain Data Science Platform (BDSP), hosted on Amazon Web Services (AWS), is increasing EEG accessibility through cooperative data sharing and research enabled by the cloud.

Because they provide insights into brain activity and not just structure, EEGs are one of the most common tests ordered by doctors to help make a diagnosis for people with brain problems. This includes seizures and epilepsy, coma, stroke, developmental delays in children, and sleep disorders. However, in current practice, EEGs are not always part of a diagnostic plan, even when they could provide important information. Currently, experts trained in EEG are in short supply, and methods to automatically interpret EEGs are not yet advanced enough to fill this gap. For these reasons, diagnostic plans that include EEGs are limited. The cloudincreases EEG accessibility by facilitatingdata sharing and research innovation, makingEEGs more accessible formore patients medical care plans.

Brain cells talk to each other using tiny electrical impulses and are constantly active, even during sleep.Activity is measured by placing small, metal discs (electrodes), held in place by tape or glue, at different locations on your scalp.These electrodes detect the tiny voltage fluctuations from the activity of millions of neurons in the brain.The electrodes are connected to an amplifier, which magnifies the weak electrical signals picked up by the electrodes. The acquired signals are often weak and susceptible to various types of interference, including noise from the environment or other non-brain biological sources. The amplifier performs signal conditioning operations to filter out unwanted noise and artifacts while preserving the relevant brain activity. This can involve processes such as amplification, filtering, and isolation.

Once conditioned, signals are converted from analog to digital format using an analog-to-digital converter (ADC). The ADC samples the analog signals at a specific rate and converts them into a digital representation that can be processed and analyzed by a computer. The digitized EEG signals can undergo further processing, like additional filtering, artifact removal, and feature extraction. Various algorithms and techniques can be applied to extract meaningful information from the EEG signals, depending on the specific analysis or diagnostic purpose. For example, algorithms can detect EEG activity that is normal for a given age, as shown in Figure 1.

Figure 1. A 15-second excerpt from a normal EEG recording from a 42-year-old woman who is awake. The vertical bars are 1 second apart. The channel names indicate locations on the persons head.

Algorithms can also detect harmful brain activity like a seizure, as shown in Figure 2.

Figure 2. A 15-second excerpt from an EEG recording from a 22-year-old man. The high-amplitude rhythmic brain activity is a seizure. Detecting this kind of abnormal electrical activity in the brain helps establish a diagnosis of epilepsy and helps doctors choose a treatment that can prevent future seizures.

The processed EEG signals are stored for later review or analyzed in real time. The data is visualized as waveforms or through spectral analysis, event-related potential analysis, or analysis by machine learning (ML) algorithms to detect abnormalities or patterns of interest.

In the US, about 75 percent of EEGs are interpreted by neurologists without expertise in EEG interpretation. This can lead to mistakes in tricky cases, such as when an EEG looks abnormal but really is not, so some patients who really have a heart problem get misdiagnosed instead with epilepsy.In many parts of the world, patients are unable to get an EEG because the doctors available are not trained to interpret EEGs.People who have sleep problems face similar challenges in getting a diagnosis. Doctors can order a sleep test (which includes recording EEG and other signals overnight while sleeping) to help diagnose sleep problems. However, getting the sleep test done can take a long time or may not be possible because there is a shortage of sleep specialists.

A research team at Harvard Medical School, headed by Drs. Brandon Westover and Robert Thomas at Beth Israel Deaconess Medical Center, and Drs. Valdery Moura Junior and Sahar Zafar at Massachusetts General Hospital aims to make EEGs more easily accessible by using artificial intelligence (AI) to automate medical diagnosis based on EEG.They are joined in this effort by other scientists from several institutions.[i]

The team is working to automate EEG and sleep testing interpretation by developing the Brain Data Science Platform (BDSP), the worlds largest and most diverse set of EEG and sleep testing data. Using this data, the team is constructing algorithms that diagnose sleep disorders, detect seizures and other forms of harmful brain activity in hospitalized patients who are critically ill, predict the risk of future seizures, and calculate the probability that a patient with coma due to brain damage will be able to recover consciousness. To be useful in the real world, they need to cope with EEG patternsfrom all people, regardless of age, gender, race, and ethnicity, and across a vast number of different health conditions. Thus, the algorithms that underlie automated EEG and sleep test interpretation must be well-trained and well-tested so that the resulting diagnoses are just as reliable or more so than can currently be obtained when EEGs are interpreted by human experts with specialty medical training in EEG and sleep test interpretation.

Beyond the diagnostics information currently available from EEGs, the team believes that there is hidden information, especially during sleep, that reveals insights into the health of the brain and which even experts cannot see. The understanding is that each of the different stages of sleep rapid-eye movement (REM), and light and deep stages of non-REM sleep has certain patterns that are normal for a given age and gender. Divergence from these norms can indicate positive or negative deviations from normal health.

The team is developing AI algorithms that use sleep signals to detect early signs of diseases like Parkinsons disease and Alzheimers disease. Earlier detection helps treatments to be given earlier when they can be more effective. The team has already found that information from sleep can predict life expectancy. Finally, one member of the research team, Dr. Haoqi Sun, has developed a way to measure brain age, as distinct from chronologic age. This validates the concept that someone who is 80 can have a brain that functions like someone 20 years younger. Accelerated brain aging (brain age older than chronologic age) is linked to a variety of brain health problems, including declining cognitive functioning and diseases like Alzheimers. The team believes that the ability to measure brain age, and similar types of hidden health information in sleep, may enable doctors to treat diseases more efficiently and effectively while providing more direct ways to measure the effects of those treatments on brain health.

The team has assembled a massive collection of EEG and sleep data currently more than 200,000 EEG recordings and 26,000-plus sleep tests. These span all medical settings where EEGs and sleep tests are performed, including outpatient neurology clinics, epilepsy centers, sleep centers, and home settings where data is collected using wearable consumer devices.For more valuable research, additional metadata is being collected as well. This includes diagnoses, medications, laboratory testing results, and brain imaging including head computed tomography (CT) scans and brain MRI images.

Dr.Westoverintends for any clinic or hospital in the world to benefit from the models that BDSP researchers develop. The benefits would not be limited to clinical groups with access to on-premises powerful servers and clusters.

We intend to offer the machine learning EEG interpretation models as an online service, where sites can upload their EEG data and get their results back within seconds, says Dr.Westover. If we want to engage clinical sites without such a level of internal IT infrastructure and bandwidth, we need to offer them a simple way to access our developments. This is where the cloud is going to be crucial.

Dr. Westover is collaborating with AWS to increase access to brain data and support researchers focusing on brain health. Twelve other hospitals have already committed to adding their data to BDSP on AWS, further increasing the possibilities for new discoveries by making it possible to study rare diseases, which are typically not seen often enough at any single hospital to allow rigorous research. Dr.Westoverbelieves BDSP will transform the field of brain health, paving the way for more personalized and precise ways to diagnose, treat, and prevent neurological disease.

These datasets are going to let us launch the field of precision brain health, said Dr. Westover.

With support from the AWS Open Data Sponsorship Program, the BDSP datasets are now openly available at no cost to researchers around the world. The BDSP dataset is one of the largest collections of brain data in existence.

[i] Collaborators include Junior Moura, PhD, Umakanth Katwa, MD, Wolfgang Ganglberger, PhD, Thijs Nassi, MSc, Erik-Jan Meulenbrugge, MSc, Yalda Amidi, PhD, Jin Jing, PhD, Haoqi Sun, PhD, Mouhsin Shafi, MD, PhD, Daniel Goldenholz, MD, PhD, Arjun Singh, MD, Sahar Zafar, MD, Shibani Mukerji, MD, PhD, Jurriaan Peters, MD, and Tobias Loddenkemper, MD at Harvard Medical School; Aaron Struck, MD at University of Wisconsin; Jennifer Kim, MD, PhD at Yale University; Emmanuel Mignot, MD, PhD and Chris Lee-Messer, MD, PhD at Stanford University; Gari Clifford, PhD, Samaneh Nasiri, PhD, and Lynne Marie Trotti, MD at Emory University; and Dennis Hwang, MD at Kaiser Permanente.

Read the original post:

Brain Data Science Platform increases EEG accessibility with open data and research enabled by AWS - AWS Blog

When Colin Twomey became interim executive director of the Data Driven Discovery Initiative (DDDI) last summer, he says, his background in behavioral ecology meant that he had a good idea of the data science needs for his own field and some idea for biology, genetics, and evolution. However, with DDDI serving as the hub for data science education and research across the School of Arts & Sciences, Twomey says he found his understanding of the needs for chemistry, sociology, and other fields to be lacking.

To tackle the problem, he followed his instinct as an ecologist: map out the system and get a big-picture view before digging into the details. What resulted is a work-in-progress map intended to capture all published research by current faculty in SAS, including their work before coming to Penn, encompassing research that spans several decades. It uses the same technology as ChatGPT and similar large language models (LLMs).

I really think of it as like a Google Maps for research. It gives you a very fast way to get oriented to a really big and complex research environment like Penn, Twomey says. He built what he calls the University Atlas Project, or uAtlas for short, during his personal time, and its just one of the ways Penn is leading in data-driven research, teaching, and applications.

At first glance, it might look like a single-cell atlas to a scientist or an abstract design to an artist. While the map is still being worked on, each of the more than 40,000 dots is a different publication by a professorcolor-coded by their departmentand zooming in shows labels for 240 topics. Departments are assigned a specific color. Red is economics. Highlighter-orange is chemistry. Pastel yellow is psychology. Robins-egg blue is Africana studies. Hot pink is cinema and media studies and so forth.

The spatial arrangement shows how thematically similar each paper is in relation to another and illustrates the interdisciplinary pursuits of Penn faculty. Theres all sorts of really unexpected overlaps, and it also doesnt put anyone into a box, Twomey says.

The Department of Physics and Astronomy shows up as very broad, Twomey says. It has its tendrils into everything, which is kind of amazing; it really does accommodate a very broad range of interests, from social sciences and psychology to chemistry.

The multicolored pattern of dots around labels such as inequality, bioethical dilemmas, and COVID-19 impact show how researchers in psychology, sociology, political science, philosophy, economics, Africana studies, and more are leading on the great challenges of our time.

The map is also searchable by name, which shows the varied interests and cross-disciplinary work of Penn faculty. For example, the spread-out clusters for physics professor Vijay Balasubramanian reflect his interests in string theory and neuroscience.

Users can also adjust the view to show only works published before or after a selected year. Twomey was struck by a bridge of green dots, for earth and environmental science, connecting hard science subjectsand specifically the topic of past climate variabilityto the social sciences. The bridge labeled climate communication, Twomey says, didnt start appearing until after about 2004, pointing to research led by Michael Mann.

Twomey says the tool has been useful to him in identifying what is going on in different departments. And he says it can also help faculty identify potential collaborators and prospective graduate students and postdocs determine with whom they want to work. My other hope for this is that, once you do this for long enough, you get these pictures of where the University is evolving over time, where research has moved, Twomey says.

Bhuvnesh Jain, the Walter H. and Leonore C. Annenberg Professor in the Natural Sciences and faculty co-director of DDDI, says he loves that Twomeys map is both sophisticatedin its use of an LLM to embed research papers onto an abstract spaceand visually intuitive.

The map transcends discipline and sub-discipline labels and shows how closely connected a lot of our work is, Jain says, adding that he had fun brainstorming with Twomey on the applications of this tool. I am confident that the users will range from incoming Penn undergraduates to the deans of our schools, who will be able to rapidly visualize the hubs of activity, the interconnections of different research efforts, and the growth areas in different fields.

To build this map, Twomey began by figuring out the affiliations of SAS faculty, which he says was a challenge because the data live in many places across the University. He then used Python to distill the data and a large language model to map the semantic content of each publication into a high-dimensional embedding space. But Twomey says visualizing hundreds of dimensions simultaneously is impractical, so the final map compresses data into a two-dimensional representation that best preserves the relationships between papers that address similar topics.

He next used the programming language Elixir to build a custom web server so the map would appear on a user-friendly website. Twomey then used an LLM again to add the research topics, choosing a labeling system that he felt was neither too dense nor too sparse, so its not overwhelming but still gives you enough waypoints.

To date, the map captures most but not all School of Arts & Sciences faculty as Twomey continues to work on the project. He also notes that some data from indexes like Google Scholar and OpenAlex may be incorrect, meaning a professor may show up as incorrectly attached to a paper or the year is wrong, so additional validation is needed. Twomeys goal is to eventually include research from graduate students and postdocs as well and to expand beyond SAS.

The School of Arts and Sciences has 28 departments and 34 centers, and seeing how all those intersect is super fascinating, but thats just one piece, one school, Twomey says. I want to have this Penn-wide and even scale it beyond Penn in the future.

More here:

Using AI to map research in the School of Arts & Sciences - Penn Today