Deep Learning for Automated Tic Detection and Prediction in Tourette Syndrome Using Electromyography and Video

Tourette syndrome is a neurological condition marked by sudden, involuntary movements or sounds called tics. For people living with the disorder, these tics can range from a quick eye blink to a complex shoulder shrug or a vocal outburst. While the condition is well-known, measuring it accurately in a clinical setting remains difficult. Doctors currently rely on patients or caregivers to remember how often tics happen over the last week, or they watch short video clips recorded in an office. Both methods are imperfect; human memory is fallible, and a brief office visit might miss the full picture of a patient's daily experience. Because tics fluctuate in intensity and can be temporarily suppressed, capturing them objectively requires a tool that can watch and listen continuously without fatigue.

Researchers are increasingly turning to technology to solve this problem, using sensors to record muscle activity and cameras to record movement. One such effort comes from a team at the University of Florida, who set out to build a computer system capable of spotting tics in real-time. They wanted to know if artificial intelligence could learn to distinguish between a person resting, moving their muscles on purpose, and having a tic, using only the electrical signals from their muscles and video footage of their body. Their goal was to create a reliable way to monitor the condition, which could eventually help doctors adjust treatments or even trigger therapies automatically when a tic is about to happen.

The team gathered data from sixteen adults diagnosed with Tourette syndrome. Each person visited a clinic multiple times over three months, during which researchers recorded their muscle activity using eight small sensors placed on their arms and neck. At the same time, a high-definition camera filmed their movements. To ensure the computer learned the right lessons, four experts reviewed the video footage and marked exactly when a tic occurred, when the person was resting, and when they were moving voluntarily. The researchers then fed this paired data—muscle signals and video labels—into three different types of computer models designed to recognize patterns. They tested these models in two ways: first, by asking them to identify what was happening in the current moment, and second, by asking them to predict what would happen in the next moment.

A key part of the study involved testing how much time the computer needed to see before it could make a decision. The researchers initially guessed that because tics are very brief, the computer would work best if it looked at very short slices of time, perhaps just one second. They tested windows of one, five, and ten seconds. The results surprised them. When the task was to detect a tic that was already happening, the computer performed best when it looked at a five-second window. The model could identify tics with an AUC of 0.79 in this longer timeframe, suggesting that the electrical signature of a tic unfolds over a few seconds and benefits from a bit of context to be recognized clearly. In contrast, when the computer tried to predict a tic before it happened, it worked best with the shortest, one-second window. This indicates that the warning signs of an upcoming tic are very recent and fleeting, requiring the system to focus on the immediate past rather than a longer history.

The study found that the computer models were generally better at spotting a tic as it occurred than at predicting one before it happened. The most successful model for detection, which used a specific type of neural network designed for time-based data, reached an AUC of 0.79. While this is a strong result, the researchers noted that the system is not yet perfect. The models struggled more with predicting future tics, and their performance varied depending on the length of the time window they analyzed. The team also observed that the computer models learned best when they were trained on data where the muscle signals were matched with the expert video labels, rather than relying on the muscle signals alone.

The researchers were careful to point out the limits of their work. The data came from a specific group of patients who had already undergone deep brain stimulation surgery, which might make their muscle signals different from those of the general population. The sensors were placed on the arms and neck, meaning the system could not detect vocal tics or facial tics that did not involve those specific muscles. Furthermore, the experts who labeled the video data did not always agree perfectly on exactly when a tic started or stopped, which introduces a small amount of uncertainty into the training process. Despite these constraints, the study provides a clear benchmark for what is possible. It suggests that for detecting tics, a system needs a few seconds of data to be sure, but for predicting them, it must react instantly to the most recent signals.

This work lays the groundwork for future tools that could monitor Tourette syndrome continuously and objectively. Instead of relying on memory or brief clinic visits, a system built on these findings could potentially track a person's condition over days or weeks, providing a complete picture of their symptoms. While the current models are not yet ready for daily use, the study demonstrates that combining muscle sensors with video analysis and artificial intelligence offers a promising path forward. By understanding exactly how much time the computer needs to see a tic and how it differs from a voluntary movement, scientists can begin to design systems that help manage the condition more effectively, moving closer to a future where treatment can be adjusted based on real-time data rather than estimates.