Explainable Deep Learning Models for Individualized Mental Health Risk Assessment Using Wearable and Smartphone Sensing

Abstract

Background: Mental health disorders represent a major global public health concern, creating a critical need for early detection and personalized intervention strategies. Advances in wearable devices and smartphone sensing enable continuous and unobtrusive monitoring of behavioral and physiological patterns, while deep learning offers powerful tools for analyzing such high-dimensional data. However, the limited interpretability of many deep learning models restricts their clinical adoption. Methods: The study presents a systematic literature review of explainable deep learning models for individualized mental health risk assessment using wearable and smartphone sensing data. A comprehensive search was conducted across IEEE Xplore, PubMed, Google Scholar, and the Directory of Open Access Journals (DOAJ). Studies were screened using predefined inclusion and exclusion criteria, resulting in the selection of 25 relevant articles for qualitative synthesis. Results: The reviewed studies demonstrate a clear shift toward multimodal sensing and advanced deep learning architectures, including attention-based and temporal models, to capture complex behavioral dynamics. Despite strong predictive performance, explainability techniques were inconsistently applied across studies. Challenges related to data quality, validation practices, generalizability, and real-world deployment were frequently identified. Conclusion: The findings highlight the importance of integrating explainability into deep learning-based mental health assessment systems to enhance trust, clinical relevance, and regulatory compliance. This review provides a structured synthesis of current approaches and outlines key research directions for developing transparent, reliable, and clinically meaningful mental health assessment frameworks.