This page was translated automatically from Turkish. Read the original

This page is an AI-assisted Turkish summary of the study presented at the 26th World Congress of Psychiatry. The study belongs to the presenting authors; please refer to the original source when citing it.

EV0320

Beyond subjective assessment: using polysomnography and deep learning in the differential diagnosis of schizophrenia, bipolar disorder, and major depressive disorder

BEYOND SUBJECTIVE ASSESSMENT: LEVERAGING POLYSOMNOGRAPHY AND DEEP LEARNING FOR THE DIFFERENTIAL DIAGNOSIS OF SCHIZOPHRENIA, BIPOLAR DISORDER, AND MAJOR DEPRESSIVE DISORDER

A multimodal foundation model combining multicenter inpatient polysomnography recordings with electronic clinical records has been developed to transform overnight sleep signals into quantitative mental health assessments, psychiatric diagnostic support, and automated sleep reports.

The study aims to develop a framework based on polysomnography (PSG) and a large language model to provide a signal-based alternative to subjective assessment in psychiatric differential diagnosis. The poster reports 42,003 PSG recordings from a sleep cohort of psychiatric inpatients compiled across multiple centers in China and 119,192 PSG recordings from five publicly accessible source sites (Site 1: 43,572; Site 2: 26,364; Site 3: 18,180; Site 4: 16,046; Site 5: 15,030), with a total sleep duration exceeding 3 million hours. The multimodal PSG input includes EEG, EOG, EMG, ECG, airflow, and SpO2 channels; the model integrates local physiological windows with whole-night context through bidirectional cross-scale attention and feeds them into a PSG-LLM architecture that processes signal tokens and clinical language tokens in the same space. The clinical tasks presented include a brain health score (78/100 in the poster example), psychiatric disorder risk stratification (with branches for mood disorders, anxiety disorders, and sleep disorders), and overnight sleep metrics (TST, SE, SL, WASO, AHI, ODI). The poster also includes a sample clinical report generated by the model, which describes an impression of mild neurocognitive disorder, possible depression, and a possible contribution from obstructive sleep apnea, based on a presentation of progressive forgetfulness, low mood, and fragmented sleep over 9 months. The main output presented in the poster is the creation of a dataset pairing full-night PSG recordings with comprehensive clinical electronic health record (EHR) data and the development of an integrated PSG-EHR foundation model and language agent framework. The poster does not provide performance metrics such as sensitivity, specificity, or accuracy for diagnostic discrimination; therefore, the model's differential diagnostic performance cannot be evaluated from these data.

All WCP 2026 summaries