How to Screen Remotely for Suspected Tardive Dyskinesia
How can clinicians use smartphone video and AI to screen patients taking antipsychotics for suspected tardive dyskinesia remotely?
Patients taking antipsychotic medications need ongoing tardive dyskinesia monitoring, but repeated in-person AIMS assessments are difficult to deliver consistently in routine and telepsychiatry practice. This workflow describes the remote video capture approach studied here for initial TD risk stratification, with clinician follow-up reserved for diagnosis and management.
-
Identify patients eligible for remote TD screening
Use this workflow for patients who are taking antipsychotic medication and therefore are at risk for TD. In the studies, the principal inclusion criterion was antipsychotic exposure for at least 90 days. The article frames the tool as a way to augment routine TD monitoring in this population.
-
Set up standardized smartphone-guided video capture
Have the patient complete the assessment through a smartphone app that guides them through a standardized protocol. In the home-use protocol from Study 3, the assessment consisted of 15 seconds of tapping a hand on the shoulder, 30 seconds of opening the mouth and sticking out the tongue followed by 30 seconds of sitting still, and responses to 2 open-ended questions. The protocol is designed to capture facial, shoulder, trunk, arm, and hand movements as well as speech video and audio.
-
Optimize video conditions before analysis
Make sure the patient's face is visible and the recording is usable before relying on the result. Participants were excluded when poor lighting or improper positioning prevented facial assessment, when instructions were not followed, or when sunglasses or masks interfered with detection of facial movements. Patients with chewing gum or loose dentures were asked to remove them in the protocol.
-
Run AI-based risk stratification on the recorded videos
Submit the video responses for algorithmic analysis rather than treating the recording itself as the final assessment. The model was trained to predict a continuous total AIMS-based risk score and then group that output into binary TD versus no-TD classifications for reported performance. The article also notes that the continuous score can support multiple thresholds such as low, medium, and high risk if a program chooses to calibrate sensitivity or specificity.
-
Interpret the result as a screening output, not a diagnosis
Use the AI result to identify suspected TD rather than to confirm the disorder. At the reported operating point in the discussion, the model had sensitivity of 0.820 and specificity of 0.821 when the threshold was set at 5.1, and overall AUC ranged from 0.85 to 0.98 when trained on all available data. The article explicitly states that the tool augments initial remote screening and does not replace physician diagnosis.
Clinical Considerations
- The protocol does not directly assess legs, feet, or toes, so isolated lower-extremity TD may be missed.
- In Study 3, 17% of participants were excluded because video quality was below the threshold required for analysis.
- The model predicts suspected TD relative to AIMS-based reference ratings and cannot function independently as a diagnostic tool.
- Performance was reported as uniformly high across analyzed age, gender, and ethnicity subgroups, but the studies were conducted in specific clinical and community populations.
Bottom Line
Use smartphone-guided video AI as a standardized remote triage screen for suspected TD in patients taking antipsychotics, then rely on clinician assessment for confirmation.
