Artificial Intelligence enters the HD space as a diagnostic tool

Artificial Intelligence, or AI, has become an everyday feature of the world we live in. Internet browsers have an ‘AI mode’ and even our refrigerators and vacuum cleaners now include AI features! While there are many ongoing discussions regarding the uses and drawbacks of AI, there is no denying that in some fields it is proving invaluable. One of these fields is medical diagnostics, and diseases like Huntington’s disease (HD) are an excellent candidate for AI-based tools due to both their complex nature and diverse diagnostic features that cover both physical and mental symptoms. 

What is AI?

Before diving into some of the tools being developed, it is helpful to understand what exactly AI is. In the broadest sense, AI is designed to be able to do things that are conventionally thought to require human intelligence, such as tasks that involve understanding language or recognizing faces. 

At the most basal level, AI operates by learning patterns and using those patterns to make very smart guesses very quickly. Older AI systems learnt patterns using rules provided for them, while newer AI such as the Machine Learning models (ML), will look through defined datasets and create their own rules based on the data. 

For example, old spam filters in our email inboxes were told to look for certain keywords and could then learn our personal preferences based on our manual input (‘mark X as spam’ or ‘this is not spam’). Now, an ML model will be given a large set of emails marked ‘spam’ or ‘not spam’ and would figure out the patterns it needs to recognize to categorize your emails on its own, without explicit keywords being set for it. 

Deep Learning (DL) models are a more complex version of ML models that have multiple learning ‘layers’ – these need large amounts of data but can find patterns within ‘unstructured data’ such as images and text.

How can AI help in healthcare?

There are many advantages to using AI in healthcare, particularly in cases involving HD and other neurodegenerative disorders. These tools are more accessible than medical care involving multiple healthcare professionals. 

For example, if wearable data could be processed by AI and used for motor assessment, it would reduce the time and frequency of hospital visits for people with HD. This would make things more convenient for people taking the assessments and caregivers. This is especially true during later disease stages or for people in more remote locations. It would also make medical care more financially sustainable. 

What can AI do for the HD community at this time?

Using AI to identify “genetic modifiers”

Current research mostly focuses on using AI to model disease onset and progression, and on using AI as a diagnostic tool to monitor disease states. For example, a recent study used genetic data from 9,000 people with HD to try to answer the question: why do people with the same number of CAG repeats have different ages of disease onset? 

The same genetic data used in this study has been analyzed by others before to identify genes which act as ‘modifiers’, genes other than the disease-causing gene that influence the age of onset. You may have heard about some of these modifier genes before, like MSH3 or PMS1, since they’re being pursued as potential treatments by other groups. 

However, with the use of AI models, this study was able to identify genes that were not identified in the original analyses. Interestingly, this study also suggested that age of symptom onset may be modified by different genes depending on the number of CAG repeats present. Analyses like these could be used to develop more personalized treatment plans for HD based on the genetic profile of the individual. 

Using AI for clinical trial recruitment

Another study was aimed at improving recruitment for HD clinical trials. They used an AI model to predict how soon someone would start developing symptoms. Accurate prediction of disease onset will be critical as trials move toward testing people before they start to develop symptoms This type of approach could reduce bias between treatment groups and increase the statistical power of the trial results. 

The scientists conducting this study used data from natural history studies, like PREDICT-HD, TRACK-HD, TrackON-HD and IMAGE-HD. Their AI model was trained using brain scans from these studies and metrics such as cognitive and motor assessment scores. 

This model was then able to predict when someone would start to develop symptoms of HD 24% better than previous studies, allowing also for more accurate classification for clinical trials. The tipping point for the computer models over human analysis was the addition of the brain scan data and the scoring metrics. That’s because a major advantage of AI is its ability to recognize complex patterns in images. 

Using AI to track movement changes

There are also multiple studies that use data from ‘wearables’ such as smartwatches or cellphones. One such study uses data from wrist wearables to monitor variations in walking patterns for people with HD. 

To do this, they trained an AI model to accurately differentiate between the involuntary movements caused by HD and the voluntary movement of the individual. This would allow clinicians to get a more accurate estimate of changes in movement abilities as the disease progresses. 

Another study was done that used publicly available walking pattern data to diagnose HD. This data used three parameters – stride interval or the time between steps, swing interval or the time any foot is in the air, and stance interval or the time the foot is on the ground. 

This study compared different AI learning models to see which model could diagnose HD most accurately. It also looked at which of these parameters was most effective at correctly predicting the presence of HD. The scientists found that three of their models were accurate over 80% of the time, and that for each model, a different parameter was most accurate (between 90%-100%). 

Where is AI in healthcare?

So why haven’t we started using AI much more extensively in healthcare? The problem lies in the nature of our current learning models. 

The most advanced models are also the most opaque – they cannot tell you why they came to a particular conclusion. Since the stakes in medical care are so high, we cannot have a system with decision making capacities that cannot give explanations. 

To solve this issue, the AI community is working on interpretable and explanatory models, which will be immensely helpful in the medical fields. 

The role of the HD community in developing AI based tools

The HD community is also crucial in the development of relevant AI-based tools. All AI models are only as good as their training data. The more data the model has, and the better organized it is, the better the model is likely to perform. But in many cases, generating medical data is very time consuming and expensive, as you’d need people with the appropriate medical knowledge to parse it. 

But one thing the HD community does very well is participate! This is one of the reasons why drug companies have gravitated toward studying HD. Because the community is so keen to participate, we have resources like PEDICT-HD, TRACK-HD, and TrackON-HD studies. If you’re interested in contributing to natural history studies like these that have helped advance AI research for HD, you can go to https://enroll-hd.org/ to learn more about the ongoing Enroll-HD study that tracks people with HD as they naturally live and age.

Due to the diligent efforts of the HD community to collect and classify this data, and make it available freely on many platforms, the AI models trained on data from people with HD perform well. 

Scientists who access this data for research purposes are asked to briefly describe their research project and the role of this data in it. The current entries show multiple projects using AI to improve disease prediction, develop more highly personalized predictions and even attempt to find new HD biomarkers! 

While the AI field is rapidly growing and evolving, we hope the development of more interpretable models and the existing presence of HD related datasets will lead to AI being more widely used in diagnostics and disease prognosis to help improve the lives of the HD community.

Summary

• Artificial intelligence (AI) is being used in HD research as a diagnostic and monitoring tool, taking advantage of the rich datasets the HD community has helped build over decades
• A study using genetic data from 9,000 people with HD used AI to identify genetic “modifiers,” genes that influence age of symptom onset, including some that previous analyses missed
• An AI model trained on brain scans and clinical scores from natural history studies (PREDICT-HD, TRACK-HD, and others) predicted symptom onset 24% better than previous methods, which could improve clinical trial recruitment
• Wearables like smartwatches are being paired with AI to track HD-related movement changes
• A current limitation is that the most powerful AI models can’t explain their reasoning, which is a major barrier to clinical use, but the field is actively working on more interpretable models
• The HD community’s strong participation in natural history studies is a competitive advantage that has generated high-quality, well-organized, and freely available data, which is why HD-trained AI models tend to perform well