Artificial intelligence: A real tool for advancing research on intellectual and developmental disabilities and beyond

Artificial Intelligence banner

By Charlene N. Rivera-Bonet, Waisman Science Writer

You’ve probably used artificial intelligence (AI) in the last few days – even unconsciously. Very likely in the last few hours. AI collects big data and uses computer algorithms to search patterns that are present in your daily life. Getting somewhere using Google Maps, your spelling autocorrected in a text, scrolling through social media, or seeing tons of ads about that item you googled just once. AI has taken many forms in the last decade, and its uses have expanded rapidly. Including its use in research.

AI model stepsFinding useful patterns in millions of medical records, analyzing multimodal data of tens of thousands of single cells, or creating full magnetic resonance images from very little input are just a handful of examples of what AI can do in research.

AI enables computers to mimic human intelligence. Computers are programmed to have the abilities of perception, cognition, and decision-making. The machines consume large quantities of labeled data (perception), analyze it to find relationships and patterns (cognition), and use what they learn to make predictions about future states (decision-making). Much like humans, they learn through training and repetition. If we take the AI chatbot ChatGPT as an example, it learns to generate humanlike conversations with users by reviewing millions of example conversations.

AI opened a world of possibilities for research on intellectual and developmental disabilities (IDD). Researchers from multiple disciplines at the Waisman Center are employing AI methods to gain a deeper, more holistic understanding of IDD with the hopes of developing better screenings, diagnostics, and patient classification; advancing the understanding of co-occurring conditions; and accelerating the identification of biomarkers and drug development.

Computers can now do work that would be tedious, too time consuming, or impossible for humans to do. AI allows for a more complete look into IDDs, including areas scientists haven’t thought to look into before.

AI use in IDD research

One of the most common types of AI used in research is machine learning. It is a mathematical model that takes large amounts of data, detects generalizations in the data – like patterns and relationships – and maps the outcome with high accuracy. But how is this useful for IDD research?

IDDs can be biologically complex. They often have overlapping factors, co-occurring conditions, heterogeneous phenotypes, multiple causes, environmental effects, among other complexities. This makes understanding IDD in depth difficult, because it is humanly impossible to look at every single aspect of an IDD, let alone measure how it all interacts to culminate in the symptoms an individual can present.

With the advances in scientific technology and the rise of data sharing there is an increasing amount of data on IDD available for study. The challenge becomes how to effectively integrate and analyze that volume of data, and the increasing solution, for many, is AI.

Many aspects of IDDs are being studied at the Waisman Center, starting from genes and single cells, to brain, behavior, and family units. Several Waisman investigators have adopted the machine learning approach into their research over the last decade in order to better integrate their data, find new patterns they hadn’t seen before, or have a less biased analyses.

Single-cell data integrated by AI

Daifeng Wang, PhD
Daifeng Wang, PhD

Researching IDDs at the single-cell level allows researchers to study differences in brain cells by profiling tens of thousands of individual cells at a time. The question becomes, how do you integrate this huge amount of data? Luckily for researchers, a large data set is where machine learning thrives. Daifeng Wang, PhD, associate professor of biostatistics and medical informatics, and computer sciences, is known as an AI guru at the Waisman Center. His research is entirely based on developing new machine learning and bioinformatics tools to better analyze multimodal data in neurological disorders in order to better understand their cellular and molecular mechanisms. Wang has come up with multiple machine learning algorithms that are capable of integrating various modalities of single-cell data to make it easier to analyze and understand. “How mutations affect gene expression and eventually affect neuronal functions, which may lead to the disease,” Wang says. “Machine learning is pretty useful to integrate and interpret that data.” The algorithm takes tens of thousands of data points from different biological processes within cells and finds how they relate to each other. Wang and John Svaren, PhD, professor of comparative biosciences, are venturing into a collaboration to better understand the genetic regulation of two types of cells of the nervous system that enable high speed transmission of nerve impulses, oligodendrocytes and Schwann cells. These cells are often impacted in disorders where neuronal communication is affected such as Charcot-Marie-Tooth disease.

John Svaren, PhD
John Svaren, PhD

Taking advantage of their single-cell data combined with other publicly available data, they want to identify mutations associated with disease or potential drug targets. “Assembly and integration of all our data sets and integrating with other data sets emerging from human epigenome analysis is a massive undertaking,” Svaren says. He believes they are currently only able to interpret a very low percentage of what the data is telling them. “I think having an unbiased procedure like AI can be transformative,” he adds. Recognizing the complexity of AI and machine learning, Wang also works with Robert Olson, PhD, research engineer at Waisman, to make machine learning algorithms more accessible and user-friendly through a web application for other biological researchers without computer science expertise. “Instead of having to do a lot of programming, [researchers] can click on a button and say, ‘I want to use this machine learning method to perform that transformation,’” Olson says. “Biologists who don’t necessarily have a lot of computational background can access these methods.”

Also, Waisman investigators Xinyu Zhao, PhD, Anita Bhattacharyya, PhD, André Sousa, PhD, and Qiang Chang, PhD, work with Wang to apply AI and machine learning to understand single cell (dys)functions in various IDDs including Rett syndrome, fragile X syndrome and Down syndrome.

Faster and quieter brain imaging

Neuroimages of the brain from magnetic resonance imaging (MRI) give insight into brain structure and function in IDD. Andrew Alexander, PhD, professor of medical physics and psychiatry, uses different AI approaches to improve the way MRI data are acquired and analyzed.

Andrew Alexander, PhD
Andrew Alexander, PhD

Alexander’s lab, as well as Wang’s, use an advanced and sophisticated version of machine learning called deep learning, which is able to acquire more pieces of information from the data it is given, compared to traditional machine learning.

Alexander and his lab used deep learning to create an algorithm that correctly identifies and labels specific groups of cells in a largely studied brain region called the amygdala. Before developing the algorithm, this took hours to do manually. “AI is very fast at identifying those sub nuclei. So, I think it can really reduce a lot of the manual labor and need for expert labeling,” Alexander says.

One of the biggest challenges with MRI is that it can be loud and slow, taking up to 10 minutes per scan. Using deep learning, Alexander’s lab is trying to improve the process of image reconstruction – or creating an interpretable image from the raw data – using shorter scan acquisitions. This, Alexander hopes, would reduce the scanning time from 10 to just a couple of minutes, which would be useful for studies of children or individuals with IDD. Along the same lines, they plan to use AI based methods to make the MRI scanner quieter.

Arezoo Movaghar, PhD
Arezoo Movaghar, PhD

Another area they anticipate using AI for is identifying a disease or pathology based on brain scans. “And I think the opportunities there are really exciting, but we also have to tread carefully. We need to make sure that we’re not biasing those results by training it on certain populations that may not be representative of a broader population,” Alexander says.

Identifying patterns and subgrouping in autism

One thing machine learning is really good at is identifying patterns. In fact, given large quantities of datasets, it can identify patterns that are not obvious to the human eye. Lauren Bishop, PhD, associate professor in the Sandra Rosenbaum School of Social Work uses machine learning in large administrative datasets including claims data to better understand patterns and subgroupings in autism. Using diagnostic codes for physical and mental health conditions, she classifies individuals into different groups, for example, correctly identifying someone who has autism versus someone who doesn’t based on health information.

Lauren Bishop-Fitzpatrick, PhD
Lauren Bishop, PhD

In forthcoming work, Bishop and her lab are using machine learning algorithms to classify people based on the way they received Applied Behavior Analysis (ABA) therapy, one of the most common interventions for autism. They then used these subgroupings to understand how ABA therapy during childhood impacts mental health in adulthood.

Up next, Bishop and collaborators will work on developing an algorithm that can predict mortality of autistic individuals by building a mortality-risk index based on Medicare records. “It will allow providers to flag folks who are at high, medium, or low risk for early mortality, and then apply treatments to help them live longer,” Bishop says.

There are many factors about autism that remain unknown. “[Machine learning] allows an algorithm to essentially build a model that takes into account these things that we don’t know about and therefore can’t effectively measure to then give us something that’s usable in practice,” Bishop says.

Better diagnosis and understanding of fragile X syndrome

Fragile X syndrome (FXS) is the most common form of inherited intellectual disability, and arises as a result of a mutation in the gene FMR1. FXS remains underdiagnosed, and the current approaches to identify patients are not very effective. Arezoo Movaghar, PhD, postdoctoral fellow in the lab of Marsha Mailick, PhD, emeritus vice chancellor of research and graduate education, uses AI to create screening approaches that can find potential cases of FXS based on the medical records of screened individuals, and identify the factors that contribute to health differences in individuals with FXS.

Marsha Mailick, PhD
Marsha Mailick, PhD

In order to achieve these goals, they use electronic medical records from their partner clinic or hospital, including people with different conditions, not just FXS. “AI can find patterns that we can’t find. Electronic medical records include information on millions of medical encounters. So, it is impossible to manually examine the data to find specific patterns. It has to be a technique that can find all of these nonlinear, non-obvious patterns from these datasets,” Movaghar says.

Most of these records come from the Marshfield Clinic in Wisconsin, which include data from 1.7 million patients. The Waisman Center has access to these medical records because of a unique partnership developed between Murray Brilliant, PhD and Mailick. Brilliant was the director of the Marshfield Clinic Personalized Medicine Research Project and is now a senior scientist at the Waisman Center. The Marshfield dataset includes information about all the medical diagnoses that patients received in their lifetime. Even though these patients were not seeking care for FXS specifically for all of their visits, the researchers can look for patterns that frequently are reported in the medical records of patients with FXS. Then they can train the algorithm in a way that can identify patients who might likely have FXS based on the differences in health patterns compared to other patients. Recently, they were also able to partner with a medical network in Chicago that gives them access to a more diverse population in terms of race, ethnicity and socioeconomic status. This ensures that the algorithm can be based on diverse populations, and produce more generalizable results. Using machine learning, they have been able to identify conditions that co-occur with FXS but have never been identified or studied, some of which have a large impact on the life of people with FXS, such as heart issues.

“AI has been very, very good at picking up the signal of things that are less evident when using conventional approaches to analysis of data,” Mailick says.

Challenges and considerations

Artificial intelligence simplifies things, but it is complex. Although it has proven to be very useful, it does come with a few caveats. The first one is it’s “black box” characteristic. It doesn’t tell you how it makes the decisions that it makes nor underlying full biological mechanisms, so you need to know the data or rely on prior knowledge in order to appropriately interpret it. Because of this, researchers need to validate their output as much as they can, Alexander says.

Rob Olson
Rob Olson, PhD

AI may be really good at its job, but “the algorithm only works as well as the training set represents what results you’re expecting to get in future cases,” Alexander says. If your input is biased, your results will be biased as well. Good training data are needed in order to obtain useful results. AI also needs a lot of data. “The one drawback of AI or machine learning is that it needs a lot of data to train a good model,” Wang says. He says the reason machine learning has been so successful in the last 10 years is because

we have a lot of data coming in and can train a highly accurate model of machine learning. Moreover, input data processing and curation to train a successful AI model should also be considered.

“I think something that’s really important to bring up is that this is all very new,” Bishop says. “We don’t, at this point, have a sense of the great potential of AI as well as the potential detriments of AI. So, I think that proceeding with caution is important, and ensuring that we replicate findings in studies that use AI with more overtly ‘knowable’ methods.”

AI presents a useful tool for researchers in the field of IDD, and provides hope for better understanding of conditions, diagnostic tools, personalized treatments, and identification of drug targets. Not every study or researcher using AI at the Waisman Center was represented in this article, but there are many experts working to incorporate this powerful tool in their research to further improve the knowledge and care of intellectual and developmental disabilities.

Your support makes a difference. Donate now to advance knowledge about human development, developmental disabilities, and neurodegenerative diseases through research, services, training, and community outreach. DONATE NOW

Waisman Center Anniversary Logo | 1973-2023 Learn more about the Waisman Center's 50th Anniversary, including events, history, stories and images:
50 Years | 1973 - 2023