Systematic review and meta-analysis of performance of wearable artificial intelligence in detecting and predicting depression

Depression, affecting approximately 260 million people globally, presents significant challenges for traditional diagnostic methods. Current approaches rely on subjective clinical observations, time-consuming questionnaires (like the PHQ-9), and suffer from inaccuracies and a shortage of mental health professionals, particularly in low-income countries. The stigma surrounding mental health disorders further hinders early detection. Wearable devices, offering objective, efficient, and real-time data collection (heart rate, physical activity, sleep patterns, etc.), present a potential solution. This review aims to systematically evaluate the performance of wearable AI in detecting and predicting depression, addressing the limitations of conventional methods and exploring the potential of this emerging technology.

The introduction mentions a scoping review by the authors themselves (Abd-Alrazaq et al., 2023) on wearable AI for depression recovery, and other reviews focusing on mobile and wearable technology for monitoring depression in children and adolescents (Sequinoa et al., 2020). Additional relevant literature, including studies on the use of AI in child and adolescent psychiatry (Welch et al., 2020), EEG-based depression detection (Yasin et al., 2021), and machine learning methods for depression screening using PHQ-9 (Kim & Lee, 2021), are cited. This highlights a growing body of research in the field, focusing on various AI approaches and data sources.

This systematic review and meta-analysis searched eight electronic databases (MEDLINE, PsychINFO, EMBASE, CINAHL, IEEE Xplore, ACM Digital Library, Scopus, and Google Scholar) for relevant studies published since 2015. A three-stage selection process (duplicate removal, title/abstract screening, full-text review) was independently conducted by two reviewers, achieving high inter-rater reliability (κ=0.85 and 0.92). Data extraction included study metadata, wearable devices used, AI algorithms employed, and performance metrics (accuracy, sensitivity, specificity, RMSE). A modified QUADAS-2 tool was used to assess risk of bias and applicability concerns. A three-level random-effects meta-analysis was performed to synthesize the results, accounting for the nested structure of the data (repeated analyses within studies). Subgroup analyses explored the influence of various factors (algorithms, devices, data sources, data types, reference standards) on the performance of wearable AI.

The review included 54 studies (published between 2015 and 2022, with a peak in 2022), involving a total of 249,203 participants. The most common wearable devices were Actiwatch AW4 and Fitbit series, predominantly wrist-worn. The meta-analysis revealed the following pooled mean performance metrics: highest accuracy (0.89, 95% CI 0.83–0.93), lowest accuracy (0.79, 95% CI 0.62–0.78), highest sensitivity (0.87, 95% CI 0.79–0.92), lowest sensitivity (0.61, 95% CI 0.49–0.72), highest specificity (0.93, 95% CI 0.87–0.97), and lowest specificity (0.73, 95% CI 0.62–0.83). Highest RMSE was 4.55 (95% CI 3.05–5.05) and lowest was 3.76. Subgroup analyses indicated statistically significant differences in performance across different AI algorithms (AdaBoost generally outperformed others, while logistic regression and decision trees performed poorly), and wearable devices. The most frequent algorithms were Random Forest (59.3%), Logistic Regression (24.1%), and Support Vector Machine (20.4%). Data sources were mixed (closed and open). The most common data input were physical activity, sleep, and heart rate data. Reference standards for depression diagnosis varied (MADRS, PHQ-9, DSM-IV/V, HDRS, etc.).

The findings suggest that wearable AI shows promise in detecting and predicting depression, achieving reasonably good performance in some instances. However, the overall performance is not yet optimal, particularly in sensitivity (correctly identifying those with depression). The substantial heterogeneity across studies highlights the need for standardization in data collection, algorithm development, and validation. The superior performance of AdaBoost compared to other algorithms warrants further investigation. The observed differences in performance based on wearable devices might be due to variations in data quality or sensor capabilities. Future research should focus on combining wearable data with other modalities (e.g., neuroimaging), improving algorithm robustness, and addressing the limitations identified in the risk of bias assessment.

This meta-analysis demonstrates that wearable AI offers potential for depression detection and prediction, but its performance is inconsistent and not yet sufficient for clinical use. Further research is needed to improve accuracy, particularly sensitivity, by standardizing methods and incorporating multimodal data. The identification of superior algorithms like AdaBoost suggests valuable avenues for future development. Ultimately, wearable AI may serve as a valuable supplementary tool, but not a replacement, for traditional diagnostic methods.

The review's limitations include potential publication bias, heterogeneity across studies due to variations in methods and populations, and the reliance on existing datasets, which might not be fully representative. The modified QUADAS-2 tool, while adapted for this context, may not capture all potential biases. The number of studies reporting RMSE was limited. Future research could use larger, more diverse datasets, and apply more sophisticated bias reduction techniques.