Detecting Tardive Dyskinesia Using Video-Based Artificial Intelligence.

Abstract

Objective: Tardive dyskinesia (TD) is a late-onset adverse effect of dopamine receptor-blocking medications, characterized by involuntary movements primarily affecting the mouth, though other body parts may be involved. Severity of TD varies from mild to debilitating and is usually irreversible. Despite the existence of treatments such as VMAT2 inhibitors, TD remains underdiagnosed, with 40,000 patients treated of an estimated 2.6 million affected US individuals. This study demonstrates a novel, efficient, and reliable method to detect and bring TD to psychiatrists' attention using video-based artificial intelligence.

Methods: Individuals taking antipsychotic medications were enrolled in Study 1 (n = 46) between March and November 2018, in Study 2 (n = 136) between May 2023 and May 2024, and in Study 3 (n = 174) between July 2023 and May 2024. Participants completed video assessments. A vision transformer machine-learning architecture was evaluated by calculating the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity compared with a reference standard of the trained raters' evaluation of TD on the Abnormal Involuntary Movement Scale.

Results: The algorithm reached an AUC of 0.89 in the combined validation cohort across Studies 1, 2, and 3. The model demonstrated strong and reliable levels of agreement, outperforming human raters.

Conclusion: Our algorithm reliably detected suspected TD, reaching higher sensitivity and specificity than trained raters using the standard assessment. The algorithm can be used to monitor patients taking antipsychotic medications, allowing scarce resources to assess identified patients for a conclusive diagnosis by psychiatrists.