AI-Enabled IoT-Based Monitoring for Early Detection of Stress and Depression

Abstract

Stress and depression are among the most common mental health challenges worldwide, yet they often go unnoticed until symptoms reach a clinical severity. Traditional diagnostic methods primarily rely on self-reported assessments and infrequent clinical observations, which can be subjective, episodic, and ineffective at capturing subtle daily behavioural and physiological changes. While recent advancements in wearable and mobile technology allow for continuous data collection, existing monitoring systems face limitations related to low accuracy of indicators, individual variability, data noise, privacy concerns, and insufficient clinical interpretability. This study proposes an AI-driven, IoT-based smart mental health monitoring framework that integrates multimodal sensing with machine learning for continuous behavioural and physiological analysis. The system collects data on activity patterns, sleep behaviour, heart rate variability (HRV), and environmental context, comparing observations against personalized baseline profiles to detect individual deviations. Adaptive predictive models categorize mental states into three groups: Normal Condition, Moderate Stress Risk, and High Depression Risk, while SHAP-based explainable AI outputs enhance clinical transparency. The evaluation was conducted using simulated physiological and behavioural datasets generated from established clinical reference ranges. Among four machine learning classifiers, the Neural Network achieved the highest performance on 350 test samples, yielding an accuracy of 94.0%, precision of 93.7%, recall of 93.0%, and an F1 score of 93.3%, with an ROC-AUC of 0.981. HRV and the Sleep Disruption Index were identified as the most predictive features, together accounting for 43.9% of the total SHAP weight. Longitudinal monitoring over a simulated 90-day period confirmed the framework's capability to detect gradual psychological decline before clinical thresholds are surpassed. Compared to traditional assessment benchmarks reported in previous literature, the proposed system exhibited a reduction in false alert rates from 22.4% to 6.2% and a decrease in mean time to detection from 18.6 days to 4.3 days.