441. Optimizing machine learning models for screening psychiatric disorders in older adults: the role of depression, stress, and anxiety measures

Abstract

Abstract Background Psychiatric disorders are prevalent among community-dwelling older adults; however, many face limited access to clinical specialists due to various barriers, such as physical distance and social stigma. Self-report questionnaires offer a practical solution. Integrating self-report questionnaires including PHQ-9, PSS, GAD-7, and WHOQOL-BREF with machine learning (ML) techniques may facilitate more effective screening in this population. Aims & Objectives The aim of this study is to exploit ML to identify questionnaire items that effectively screen psychiatric disorders in older population. We employed linear discriminant analysis (LDA) to evaluate the predictive performance and determine the most effective screening model. Method A total of 232 community-dwelling older adults were included after excluding individuals with cognitive impairment (K-MMSE ≤ 24). Participants completed self-report questionnaires assessing depressive symptoms (PHQ-9), perceived stress (PSS-10), anxiety (GAD-7), and quality of life (WHOQOL-BREF). Psychiatric diagnoses were determined by trained clinicians using the Mini-International Neuropsychiatric Interview (MINI) and used as reference labels. ML models were developed using different combinations of questionnaires. Model performance was evaluated using repeated stratified 5-fold cross-validation with 100 iterations. AUC was used as the primary performance metric. Results Using LDA, PHQ-9 alone demonstrated moderate classification performance (AUC = 0.720). Adding perceived stress information (PHQ-9 + PSS) substantially improved discrimination (AUC = 0.835). Further inclusion of anxiety symptoms (PHQ-9 + PSS + GAD-7) yielded the highest performance (AUC = 0.842). In contrast, adding quality-of-life measures (all questionnaires) did not further improve performance and resulted in a lower AUC (0.803). These findings indicate a stepwise improvement in screening performance with the addition of stress and anxiety measures, followed by a plateau or decline when broader quality-of-life domains were included. Discussion & Conclusions The PHQ-9, PSS, and GAD-7 combination yielded the most robust classification performance. This suggests that a multidimensional approach—incorporating depression, stress, and anxiety—is essential for the comprehensive detection of psychiatric conditions in the elderly. Interestingly, adding quality-of-life measures (all questionnaires) resulted in decreased accuracy, indicating that overly broad data may introduce statistical noise to the ML model. This study demonstrates that an optimized ML model can provide a screening accuracy comparable to clinical interviews performed by trained clinician.