Machine learning tool can distinguish tics from spontaneous movements
Presenter: Leonie Felicia Becker, MD, Institute of Systems Motor Science, University of Lübeck, Germany.
New machine learning approaches in tic detection: Seeking to learn more about the characteristic of tics. Abstract 951. Presented August 27, 2023
Using a dataset of videos of people with tic disorders to train a Random Forest classifier to identify tic movements, a single detection tic score developed from various predictable features had an accuracy of 83% in classifying patients and healthy controls.
“It can be difficult to distinguish tics in people with tic disorders from extra movements in healthy controls,” according to the researchers, led by Leonie Felicia Becker, MD, Institute of Systems Motor Science, University of Lübeck, Germany. “Furthermore, rating tics based on video recordings is time consuming and cumbersome. Machine learning has the potential to aid and complement clinical assessment and also to yield additional information.”1-3
Their study used a dataset of 63 videos of people with tic disorders to train a Random Forest classifier for second-wise tic detection. The trained classifier was used to predict the presence of tics in patients and extra movements in healthy controls. The second-wise predictions were aggregated to calculate various tic meta-features: number of tics per minute, maximum duration of a continuous non-tic segment, maximum duration of a continuous tic, average duration of tic-free segments, number of changes from tic to non-tic segments and vice versa per minute, average size of a tic cluster, and number of clusters per minute.
A logistic regression model with parameters obtained from a dataset of 124 videos of individuals with tic disorders and 162 videos of healthy controls was then created from the tic features. A test dataset of 50 videos of patients and 50 videos of healthy controls was used to assess the accuracy of the logistic regression model in classifying patients and healthy controls.
A test accuracy of 84% and an F1 score (classification accuracy) of 83% were achieved using a predefined threshold of “3” for tic severity. Paired sample T-tests revealed significant differences between patients and healthy controls in all meta-features.
The machine learning algorithm could thus be further developed into a clinically applicable tool, the investigators note. “To improve classification accuracy, our next step is to fine-tune the tic detection score,” they wrote in their poster presentation. “Also, we aim to analyze the significance of each feature to determine which characteristics are most helpful in differentiating between the two groups. The proposed algorithm might also be helpful to distinguish between tics and functional tic-like movements.”
Davide Martino, MD, PhD, University of Calgary, Canada, who was not involved with the study, commented on the findings. “The frequency and characteristic cluster aggregation of tics are key determinants of tic severity,” he said. “Wearable sensors recording tics in patients’ natural environment are currently under exploration, but the anatomical distribution and diverse phenomenology of tics hinder the routine clinical applicability of these sensors. Tic frequency and phenomenology are also routinely assessed using video recordings usually obtained in a clinical setting, a methodology often used in clinical trials. The study reports a very good classification accuracy of the algorithm (83%), although the composition and accuracy of the tic detection score is still in progress.”
Martino noted that an algorithm that measures frequency and clustering of tics from video recordings has strong translational value in routine clinical practice and clinical research, as it would likely optimize reliability and efficiency of these measurements. Although limited to facial and head tics, the same approach can be extended to other body regions and phonic tics. It is also important to point out that video recording-based measures will inevitably still need to be integrated with other domains of tic severity (eg, interference with daily routines and functional impact) in order to achieve a truly comprehensive assessment of tics.
References
- Bartha S, Bluschke A, Rawish T, et al. Extra movements in healthy people: challenging the definition and diagnostic practice of tic disorders. Ann Neurol 2023; 93(9):472-478. doi: 10.1002/ana.26586.
- Brügge NS, Sallandt GM, Schappert R, et al. Automated motor tic detection: a machine learning approach. Mov Disord 2023; 38(7):1327-1335. doi: 10.1002/mds.29439.
- Wu J, Zhou T, Guo Y, et al. Tic detection in Tourette syndrome patients based on unsupervised visual feature learning. J Healthc Eng 2021:5531186. doi: 10.1155/2021/5531186.