Photo Credit: Nathaphat

AI-powered screening shows promise for improving early detection of tardive dyskinesia, highlighting its potential to ease the clinical workforce burden.

“Early detection of [tardive dyskinesia (TD)] allows the deployment of effective interventions to mitigate morbidity,” researchers wrote in The Journal of Clinical Psychiatry. “Traditional assessment is accomplished with one of several validated scales. However, it is difficult even for skilled diagnosticians to devote in-person resources, as often as 2–4 times per year, as would be necessary to provide every patient with the recommended standard of care regarding monitoring for TD.”

Furthermore, there is no consensus regarding the conditions that can be managed by telepsychiatry alone, which would reduce the in-person burden for diagnosing and monitoring TD, noted Owen Scott Muir, MD, and colleagues.

Traditional screening for TD involves in-person visits and assessment tools, such as the Abnormal Involuntary Movement Scale (AIMS), which is most effective when administered by clinicians familiar with diagnosing and treating TD. Dr. Muir and colleagues sought to incorporate the use of video-based AI for initial screenings of TD. To test the technology, researchers recruited patients who were at risk for TD to participate in three studies.

Testing AI to Screen for TD

Study #1

 The researchers recruited 46 patients for the first study. Each participant was assessed using AIMS, but the screening was recorded by a device using a smartphone app instead of being administered by a human assessor. The device captured close-up video of each participant’s face, trunk, and hands.

After the AIMS assessment, participants were asked six open-ended questions, and their answers were recorded on video. The app used an algorithm to assess each participant’s data, after which three trained raters also viewed and scored each participant’s assessment.

Study #2

 The second study included 136 patients. Similar to the first study, participants in the second study were administered an AIMS assessment, which was recorded using a smartphone app. However, instead of asking six open-ended questions, Dr. Muir and colleagues chose the three questions from the first study that elicited the most detailed responses. Three trained raters also reviewed these responses.

Study #3

In the third study, 174 patients underwent an AIMS assessment administered by a single rater. The researchers also collected demographic data and reconfigured the smartphone app’s assessment protocol to record three videos in which participants:

1. Tapped a hand on a shoulder for 15 seconds
2. Opened their mouth and stuck out their tongue for 30 seconds, then sat still for 30 seconds
3. Answered two open-ended questions

Comparing AI to In-Person Assessment

To effectively compare AI screening to in-person assessment for TD, Dr. Muir and colleagues made sure to include a mix of participants who had already been diagnosed with TD and those who exhibited no signs of the condition. They excluded data from 72 participants whose video quality was too low for accurate assessment.

The researchers used a visual transformer algorithm to assess participants for TD using AIMS ratings from experienced clinicians. The algorithm was updated with data from each study to achieve high accuracy in screening participants for TD, resulting in an area under the curve (AUC) ranging from 0.85 to 0.98. They found that the algorithm could outperform human raters and that, while the app wasn’t equipped to assess movements in the legs and feet, it could predict a total AIMS score.

“The combined studies demonstrate that self-administered, smartphone-recorded video interviews can reliably yield data scored using algorithms [and] produced using highly discriminating machine learning approaches,” Dr. Muir and colleagues wrote. “The underdiagnosis of TD, exacerbated by insufficient professional training and the time constraints of assessments, underscores the value of our AI-based tool. Enabling efficient, accurate, and scalable detection of TD, followed by a psychiatrist’s diagnostic assessment completed following the algorithm’s detection, this technology has the potential to significantly improve early diagnosis and patient outcomes, especially in remote care settings where resources are the scarcest.”