An AI video tool catches what busy clinicians miss
STUDY: Sterns AA et al, J Clin Psychiatry 2025;86(3):25m15792
STUDY TYPE: Diagnostic accuracy study
FUNDING: National Institute of Mental Health; partial industry funding (Teva provided registry data for Study 3; multiple authors hold equity in the sponsor company)
Background
An estimated 2.6 million Americans have tardive dyskinesia (TD), but only about 40,000 are being treated. TD is underdiagnosed partly because the standard assessment — the Abnormal Involuntary Movement Scale (AIMS) — requires trained, in-person raters performing evaluations 2–4 times per year, a standard that’s rarely met in practice.
The Study
356 patients on antipsychotic medications completed standardized smartphone video assessments across three studies. A vision transformer AI model analyzed the videos and was compared to trained human raters scoring the AIMS.
The algorithm achieved an AUC of 0.89 (which means it is 89% accurate at diagnosis) in the full combined cohort, with sensitivity of 82% and specificity of 82%. It outperformed human raters on interrater reliability — the algorithm’s Cohen’s κ was 0.61 versus 0.57 for raters who had already calibrated through consensus discussion. Performance held up consistently across sex and race subgroups.
Limitations: About 17% of participants in Study 3 (conducted at home) had to be excluded for poor video quality.
Practice Implications
- TD is difficult to detect, particularly in tele-health. Catch it early; use the app.
- The TDScreen app is free, so what’s the catch? The company is offering it in hopes that clinicians will use their other AI medical and EMR software.
- Studies find patients are more likely to stick with treatment if they are involved and informed about risks, including TD.
— Chris Aiken, MD
Director, Psych Partners
Editor in Chief, Carlat Psychiatry Report
What’s Your Take? Share in Comments
- Have your patients used the app? What’s their experience?
