Improved diagnosis of bipolar versus unipolar depression by AI-based multimodal behavioural analysis

Background and study aim:Accurately distinguishing bipolar from unipolar depression—given their differing pathophysiology and treatment approaches—remains a major clinical challenge.The present study applied artificial intelligence (AI) to quantify multiple behavioural and physiological parameters —facial expressions, speech prosody, eye movements, and sleep patterns— in a large sample of individuals with bipolar disorder and unipolar depression. By integrating these multimodal data with clinical information through machine learning, we sought to identify optimal distinguishing features and to develop an objective diagnostic support system, aiming to advance diagnostic accuracy, personalize treatment, and improve clinical outcomes.

Methods

This cross-sectional observational study included adults with bipolar disorder or unipolar depression, who were recruited nationwide in Japan and assessed through video interviews (Pre-registered: UMIN000048854). All participants received a full explanation of the study procedures, and provided written informed consent. All procedures were approved by the Ethics Committee of Hamamatsu University School of Medicine (21-335). During the interview, diagnoses were established using the Structured Clinical Interview for DSM-5-Research Version (SCID-5 RV). Depressive and manic symptom severity was assessed using the Hamilton Depression Rating Scale (HDRS) and Young Mania Rating Scale (YMRS), respectively. Artificial intelligence (AI)-based integrative, quantitative analyses of behavioural characteristics, including facial expressions, eye gaze, and voice prosody were conducted on the recorded videos.

Results

The data collection was conducted from August 2022 to March 2024. Among the 505 subjects who provided written consent and participated in the study, 459 subjects were confirmed as eligible, and 379 participants completed the Zoom interview. Based on the lifetime diagnosis of SCID-5, 120 was diagnosed as bipolar disorder type 1, 93 as bipolar disorder type 2, 122 as major depressive disorder, and 74 has diagnosis of continuous depressive disorder, while 25 participants did not meet criteria for any of these diagnoses. The overlap with major depressive disorder and continuous depressive disorder was recognized in 55 participants. 366 of the 379 participants were analysed, since the 13 participants were excluded because of the incomplete recordings. AI-based diagnostic differentiation between bipolar and unipolar depression using four-fold, three-times repeated cross-validation achieved an overall accuracy of 78%, with an area under the curve of 0.76. Key features contributing to bipolar diagnosis included greater temporal instabilities in neutral facial expression, eye gaze, sad facial expression, and root mean square amplitude of speech. HDRS could be predicted in unipolar and bipolar depression with influential predictors of depression included sleep inefficiency, autistic traits, variability in sad facial expression, and comorbid anxiety disorder. YMRS could be predicted with major predictors including comorbid anxiety disorder, sleep inefficiency, and variabilities in sad facial expression and eye movement.

Conclusions

Integrating multimodal behavioural indicators with AI-based modelling enables clinically meaningful differentiation between bipolar and unipolar depression and supports the prediction of depressive —manic symptoms. Using large-scale, real-world data, these findings provide strong support for the clinical applicability of AI-assisted diagnostic assessment.

References

No references