One item per disorder: a cross-diagnostic optimization framework for simultaneous prediction of six psychiatric disorders

Abstract Digital mental health systems increasingly rely on multimodal, AI-driven approaches, yet remain constrained by lengthy, disorder-specific questionnaires. This burden is amplified in patients with overlapping symptoms, creating a measurement bottleneck that limits scalable multi-disorder screening. To address this, we developed a cross-diagnostic item selection framework that identifies one representative item from each of six psychiatric instruments—PHQ-9, GAD-7, PCL-5, AUDIT, PDSS, and DSI-SS—via multi-objective optimization balancing predictive utility and redundancy. This reduces assessment length from 57 to 6 items (~ 90% reduction). In a general population sample (N = 1,600), a CatBoost multi-output model using these six items plus age and sex achieved high performance on an independent lockbox dataset (AUROC: 0.909–0.977; AUPRC: 0.829–0.903). Cross-diagnostic integration improved performance for depression, anxiety, PTSD, and alcohol use, while panic and suicidality remained anchor-dominant, consistent with SHAP and GGM analyses. This compact framework enables scalable, interpretable, multi-condition screening.