Abstract
Machine learning (ML) has been increasingly applied to depression prediction among older adults, creating opportunities for AI-enabled mental health screening and decision support. However, existing research remains fragmented, with limited consensus on which ML approaches, data inputs, and validation strategies are more suitable for practical deployment. This scoping review synthesizes 35 studies on ML-based late-life depression prediction. The reviewed literature shows rapid recent growth, frequent use of aging cohort datasets, and strong reliance on tree-based ensemble models. Model inputs were highly heterogeneous, spanning physiological, social, psychological, behavioral, sociodemographic, and environmental domains. Some studies identified risk factors based on predefined model inputs, whereas others relied on model outputs or explainability methods. Although reported model performance was generally promising, direct comparison across studies was limited by differences in samples, outcome definitions, predictor spaces, and validation designs. External validation, calibration assessment, explainability, and cross-context transportability remain insufficiently addressed. Future research should move beyond performance-oriented model development toward explainable, externally validated, context-sensitive, and ethically responsible AI-enabled decision support for older adult mental health care.
Recommended Citation
Zhou, Xi; Lyu, Jue-Ni; and Cheung, Christy M. K., "Machine Learning-Based Depression Prediction in Older Adults: A Scoping Review of Modeling Approaches, Risk Factors, and Validation Strategies" (2026). ICEB 2026 Proceedings (Beijing, China). 56.
https://aisel.aisnet.org/iceb2026/56