Machine-Learning Tool Sorts Tics From Non-Tics on Video

COPENHAGEN — A novel machine-learning tool that can distinguish between tics in patients with tic disorders and non-tic movements in healthy controls could potentially save clinicians time and improve the accuracy of tic identification, German researchers suggest.

Videos of more than 60 people with tic disorders were assessed manually to provide a set of clinical features related to facial tics. These were then fed into a machine-learning tool that was trained on nearly 290 videos of patients and controls, and then tested on a further 100 videos.

The resulting tool is “useful to detect tics and distinguish between tics and other movements in healthy controls,” said lead author Leonie F. Becker, MD, Institute of Systems Motor Science, University of Lübeck, Lübeck, Germany, and colleagues.

The findings were presented here at the International Congress of Parkinsons Disease and Movement Disorders (MDS) 2023.

The applications of the machine-learning algorithm could eventually extend well beyond analyzing videos of patients recorded in the doctor’s office, said Becker.

“Having patients in our clinic is really artificial because they may suppress their tics,” she told Medscape Medical News. It is “a really different situation at home or at school.”

She hopes that in the future, patients could record themselves on video sitting at home and have that video analyzed by the machine-learning tool. The tool could even be used longitudinally to assess the impact of medication, she said.

For the moment, however, Becker stressed that they have a tool that can simply differentiate between tics and normal movements.

Before it can be released as a clinical application, the tool needs to be able to distinguish between “tics and functional tics, and between tics and myoclonus and every other hyperkinetic movement,” and it needs to be validated, she said.

“I think it’s years before we have this as an app for your patient.”

Tic or Extra Movement?

Becker explained that their group recently conducted a study of healthy individuals, demonstrating that “even people without a tic disorder sometimes move a little bit,” although all participants had been asked to sit still.

The team, therefore, wanted to develop a means of reliably distinguishing between these “extra movements” in healthy control participants and tics in people with tic disorders.

One challenge of this task is that rating tics on video recordings is time-consuming and cumbersome; the team reasoned that an automated, machine-learning system could be a more efficient means of assessment, as well as potentially improving classification accuracy.

The researchers used a dataset of 63 videos of people with tic disorders to train a Random Forest classifier to detect tics per second of video.

They first identified 170 facial landmarks and manually tracked the features of tics to indicate whether a tic greater than or equal to a predefined threshold for severity had occurred within 1 second. The severity threshold was chosen as a score of 3 on the Yale Global Tic Severity Scale, which Becker said is a tic which “everybody who looks at it would recognize.”

This information was fed into the machine-learning tool to train it to predict the presence of tics in each second. These per-second predictions were aggregated over the whole video to calculate a series of clinical “meta-features,” including the number of tics per minute, the maximum duration of a continuous tic, the average duration of tic-free segments, the average size of a tic cluster, and the number of clusters per minute.

The features were then combined into a logistic regression model, which was trained on a dataset of 124 videos of individuals with tic disorders, and 162 videos of health controls.

To determine the accuracy of the model, it was then tested on a dataset of 50 videos of patients with tic disorders and 50 videos of healthy controls.

The machine-learning tool was able to identify severe tics with a test accuracy of 84%, and an F-1 score, which combines the positive predictive value with the sensitivity, of 83%.

The area under the receiver operating characteristics curve was 0.896, and the authors report that the tool revealed significant differences in all meta-features between patients and healthy controls.

Data Quality and Privacy

Approached for comment, Christos Ganos, MD, Department of Neurology, Charité University Medicine Berlin, Germany, said that the current study is one of several looking at ways of “automatically classifying patterns of behavior.”

He told Medscape Medical News that it has the potential to not only “reinforce our clinical decision-making” by demonstrating that “the way we classify phenomenon has been correct all along,” but also by showing ways of improving it.

He noted that a new classification of facial tics is being developed, and the phenomenological aspect is “so broad” that machine-learning models could help with some aspects of this, although it will take some time to have useful information from current efforts.

He emphasized, however, that there are “several caveats” to the use of artificial intelligence in this manner, the first being the quality of the data that is fed into the machine-learning tools in the first place.

The information needs to be “correctly labelled,” said Ganos, and he is convinced that there will, initially, be a “lot of white noise” from studies that have trained tools using poorly classified data.

Another fundamental aspect, and one that is “going to be talked about a lot” in the future, is that of data protection, he added.

“I worry increasingly” over stories in the media of “videos being re-circulated and re-posted,” he said. “Many of these data…labeled and fed into certain algorithms will exist forever.”

“Forever means a long time,” he stressed, “and it has many implications for generations to come, so we should be aware of that.”

“Of course, [machine learning] has great possibilities to be used in therapeutic trials, to monitor symptoms over the large scale, and all of this is very positive,” Ganos told Medscape Medical News. “But our role, in many ways, is to make sense of the data, and of what data we feed into these type of approaches, and of how best to leverage it.”

“Strong Translational Value”

Davide Martino, MD, associate professor of neurology in the Department of Clinical Neurosciences at the University of Calgary in Canada, commented in a press release that “an algorithm that measures frequency and clustering of tics from video recordings has strong translational value in routine clinical practice and clinical research.”

This is because “it would likely optimize reliability and efficiency of these measurements,” he explained.

“Although limited to facial/head tics, the same approach can be extended to other body regions and phonic tics,” he added.

“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,” such as interference with daily routines and functional impact, “in order to achieve a truly comprehensive assessment of tics,” Martino underlined.

The study had no specific funding. The investigators report no relevant financial relationships.

International Congress of Parkinson’s Disease and Movement Disorders (MDS) 2023: Abstract 951. Presented August 29, 2023.

For more Medscape Neurology news, join us on Facebook and X (formerly Twitter)

Source: Read Full Article