Early Detection of Depression Using Wearable Sensor Data: A Comparative Study of Machine Learning Models and Explainable AI Techniques

Abstract

Depression screening relies largely on periodic self-report instruments, which can miss gradual behavioural change between assessments. Wearable devices continuously record sleep, physical activity and heart rate, yet published studies differ widely in data type, labelling and validation rigour, and evidence on participant-independent generalisation remains limited. This paper presents a complete, pre-specified study design for early depression-risk screening support from wearable-derived data, with a leakage-controlled comparative framework for classical machine learning, deep learning and statistical time-series baselines, and a multi-method explainable-AI analysis (SHAP, LIME, permutation importance, partial dependence). We audit four public or access-controlled datasets (Depresjon, GLOBEM, RADAR-MDD, StudentLife), document which required modalities each contains, and assign them to separate, non-pooled experiments. We define sleep, activity, physiological and circadian features mathematically, specify nine experiments (A–I) covering modality ablation, model comparison, cross-subject and temporal generalisation, and robustness to missing or noisy data, and describe calibration, cluster-bootstrap confidence intervals and multiplicity-corrected significance testing. No experiments have been executed for this manuscript: all result tables are marked templates, and every quantitative statement about prior work is attributed to its source. The framework targets research screening support, not clinical diagnosis.

Introduction

The text presents a research framework for using wearable devices, machine learning (ML) and explainable artificial intelligence (XAI) to monitor depression and identify potential depressive changes at an early stage. Major Depressive Disorder (MDD) is traditionally assessed through clinical interviews and standardised screening tools such as PHQ-9, MADRS and BDI-II. However, these assessments are conducted periodically and may not capture changes in an individual's mental health that occur between clinical evaluations.

The study highlights the potential of passive sensing, also known as digital phenotyping, which uses wearable devices to continuously collect information about physical activity, sleep patterns, heart rate and circadian rhythms. Changes in these behavioural and physiological parameters may provide useful indicators of depression-related conditions.

Literature review and existing research

Previous studies using datasets such as Depresjon, StudentLife, GLOBEM and RADAR-MDD have demonstrated the potential of wearable and smartphone-based sensor data for identifying depression-related behavioural patterns. Machine-learning algorithms, particularly Random Forest, Support Vector Machine (SVM) and deep learning models, have been widely explored.

Although previous studies have reported promising detection performance, systematic reviews have identified limitations such as small sample sizes, heterogeneous datasets, retrospective analysis and inadequate validation across independent participants. Therefore, wearable-based AI systems are not yet sufficiently validated for routine clinical application.

Research gaps

The study identifies four major research gaps:

1. Data leakage: Some studies use random data splitting, allowing information from the same participants to appear in both training and testing datasets, resulting in potentially overestimated performance.

2. Early prediction: Most existing studies focus on detecting current depression rather than predicting future depressive symptoms among individuals who are currently non-depressed.

3. Multimodal analysis: The separate and combined contributions of sleep, physical activity and physiological signals require further investigation.

4. Explainable AI: Existing studies often rely on a single explanation method without examining the consistency and reliability of explanations across different models and methods.

Research objectives and contributions

The proposed research addresses six major objectives:

• Identify wearable-derived features associated with depression-related outcomes.

• Investigate whether machine learning can identify potential depression risk before significant deterioration is detected through conventional assessments.

• Examine whether combining sleep, activity and physiological features improves predictive performance.

• Compare different machine-learning algorithms to identify their relative performance.

• Apply explainable AI techniques to improve model transparency and interpretability.

• Evaluate model robustness across individuals, demographic groups and different time periods.

Conclusion

We specified a reproducible, leakage-controlled framework for evaluating wearable-based depression-risk screening, from a dataset audit through nine experiments, calibration-aware evaluation, multiplicity-corrected statistics and a multi-method XAI protocol. The immediate next step is executing Experiments A–I on GLOBEM (and RADAR-MDD if access is granted) and replacing every template with measured values. Future directions include personalised baselines, multi-site prospective validation, uncertainty-aware abstention, federated or on-device learning, and fairness-constrained training.

References

[1] K. Kroenke, R. L. Spitzer and J. B. W. Williams, “The PHQ-9: Validity of a brief depression severity measure,” J. Gen. Intern. Med., vol. 16, no. 9, pp. 606–613, 2001. [2] S. A. Montgomery and M. Åsberg, “A new depression scale designed to be sensitive to change,” Br. J. Psychiatry, vol. 134, pp. 382–389, 1979. [3] A. T. Beck, R. A. Steer and G. K. Brown, Manual for the beck depression inventory-II, Psychological Corp., 1996. [4] J.-P. Onnela and S. L. Rauch, “Harnessing smartphone-based digital phenotyping to enhance behavioral and mental health,” Neuropsychopharmacology, vol. 41, no. 7, pp. 1691–1696, 2016, doi: 10.1038/npp.2016.7. [5] D. C. Mohr, M. Zhang and S. M. Schueller, “Personal sensing: Understanding mental health using ubiquitous sensors and machine learning,” Annu. Rev. Clin. Psychol., vol. 13, pp. 23–47, 2017. [6] F. Matcham et al., “Remote assessment of disease and relapse in major depressive disorder (RADAR-MDD): A multi-centre prospective cohort study protocol,” BMC Psychiatry, vol. 19, pp. 72, 2019, doi: 10.1186/s12888-019-2049-z. [7] N. C. Jacobson et al., “Digital biomarkers of mood disorders and symptom change,” npj Digit. Med., vol. 2, pp. 3, 2019. [8] A. Abd-Alrazaq, R. AlSaad, F. Shuweihdi, A. Ahmed, S. Aziz and J. Sheikh, “Systematic review and meta-analysis of performance of wearable artificial intelligence in detecting and predicting depression,” npj Digit. Med., vol. 6, pp. 84, 2023, doi: 10.1038/s41746-023-00828-5. [9] J. Liu, J. Wang, Z. Wu and M. I. S. B. A. Assim, “The performance of wearable device-based artificial intelligence in detecting depression: Systematic review and meta-analysis,” JMIR Ment. Health, vol. 13, pp. e85319, 2026, doi: 10.2196/85319. [10] A. Abd-Alrazaq et al., “Wearable artificial intelligence for anxiety and depression: Scoping review,” J. Med. Internet Res., vol. 25, pp. e42672, 2023, doi: 10.2196/42672. [11] E. Garcia-Ceja et al., “Depresjon: A motor activity database of depression episodes in unipolar and bipolar patients,” in Proc. 9th ACM multimedia systems conf. (MMSys), 2018. [12] P. Jakobsen et al., “Applying machine learning in motor activity time series of depressed bipolar and unipolar patients compared to healthy controls,” PLoS ONE, vol. 15, no. 8, pp. e0231995, 2020, doi: 10.1371/journal.pone.0231995. [13] J. G. Rodr??guez-Ruiz et al., “Comparison of night, day and 24 h motor activity data for the classification of depressive episodes,” Diagnostics, vol. 10, no. 3, pp. 162, 2020, doi: 10.3390/diagnostics10030162. [14] R. Wang et al., “StudentLife: Assessing mental health, academic performance and behavioral trends of college students using smartphones,” in Proc. ACM int. Joint conf. Pervasive ubiquitous comput. (UbiComp), pp. 3–14, 2014, doi: 10.1145/2632048.2632054. [15] S. Saeb et al., “Mobile phone sensor correlates of depressive symptom severity in daily-life behavior: An exploratory study,” J. Med. Internet Res., vol. 17, no. 7, pp. e175, 2015. [16] P. Pedrelli et al., “Monitoring changes in depression severity using wearable and mobile sensors,” Front. Psychiatry, vol. 11, pp. 584711, 2020. [17] X. Xu et al., “Leveraging routine behavior and contextually-filtered features for depression detection among college students,” Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., vol. 3, no. 3, pp. 116, 2019. [18] X. Xu et al., “GLOBEM dataset: Multi-year datasets for longitudinal human behavior modeling generalization,” in Proc. NeurIPS datasets and benchmarks track, 2022. [19] F. Matcham et al., “Remote assessment of disease and relapse in major depressive disorder (RADAR-MDD): Recruitment, retention, and data availability in a longitudinal remote measurement study,” BMC Psychiatry, vol. 22, pp. 136, 2022, doi: 10.1186/s12888-022-03753-1. [20] J. Ø. Berle, E. R. Hauge, K. J. Oedegaard, F. Holsten and O. B. Fasmer, “Actigraphic registration of motor activity reveals a more structured behavioural pattern in schizophrenia than in major depression,” BMC Res. Notes, vol. 3, pp. 149, 2010, doi: 10.1186/1756-0500-3-149. [21] S. Woll, D. Birkenmaier, et al., “Applying AI in the context of the association between device-based assessment of physical activity and mental health: Systematic review,” JMIR mHealth uHealth, vol. 13, pp. e59660, 2025. [22] S. Kapoor and A. Narayanan, “Leakage and the reproducibility crisis in machine-learning-based science,” Patterns, vol. 4, no. 9, pp. 100804, 2023. [23] X. Xu et al., GLOBEM dataset: Multi-year datasets for longitudinal human behavior modeling generalization (version 1.1), doi: 10.13026/r9s1-s711. [24] R. J. A. Little and D. B. Rubin, Statistical analysis with missing data, Wiley, 2002. [25] Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology, “Heart rate variability: Standards of measurement, physiological interpretation, and clinical use,” Circulation, vol. 93, no. 5, pp. 1043–1065, 1996. [26] A. J. K. Phillips et al., “Irregular sleep/wake patterns are associated with poorer academic performance and delayed circadian and sleep/wake timing,” Sci. Rep., vol. 7, pp. 3216, 2017. [27] R. J. Cole, D. F. Kripke, W. Gruen, D. J. Mullaney and J. C. Gillin, “Automatic sleep/wake identification from wrist activity,” Sleep, vol. 15, no. 5, pp. 461–469, 1992. [28] E. J. W. Van Someren et al., “Bright light therapy: Improved sensitivity to its effects on rest-activity rhythms in Alzheimer patients by application of nonparametric methods,” Chronobiol. Int., vol. 16, no. 4, pp. 505–518, 1999. [29] C. Cortes and V. Vapnik, “Support-vector networks,” Mach. Learn., vol. 20, no. 3, pp. 273–297, 1995. [30] L. Breiman, “Random forests,” Mach. Learn., vol. 45, no. 1, pp. 5–32, 2001. [31] T. Chen and C. Guestrin, “XGBoost: A scalable tree boosting system,” in Proc. 22nd ACM SIGKDD int. Conf. Knowl. Discovery data mining, pp. 785–794, 2016. [32] G. Ke et al., “LightGBM: A highly efficient gradient boosting decision tree,” in Advances in neural information processing systems (NIPS), pp. 3146–3154, 2017. [33] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural Comput., vol. 9, no. 8, pp. 1735–1780, 1997. [34] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2015. [35] G. E. P. Box and G. M. Jenkins, Time series analysis: Forecasting and control, Holden-Day, 1970. [36] D. R. Roberts et al., “Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure,” Ecography, vol. 40, no. 8, pp. 913–929, 2017. [37] R. D. Riley et al., “Calculating the sample size required for developing a clinical prediction model,” BMJ, vol. 368, pp. m441, 2020. [38] G. S. Collins, J. B. Reitsma, D. G. Altman and K. G. M. Moons, “Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD),” Ann. Intern. Med., vol. 162, no. 1, pp. 55–63, 2015. [39] G. S. Collins et al., “TRIPOD+AI statement: Updated guidance for reporting clinical prediction models that use regression or machine learning methods,” BMJ, vol. 385, pp. e078378, 2024. [40] T. Saito and M. Rehmsmeier, “The precision-recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets,” PLoS ONE, vol. 10, no. 3, pp. e0118432, 2015. [41] G. W. Brier, “Verification of forecasts expressed in terms of probability,” Mon. Weather Rev., vol. 78, no. 1, pp. 1–3, 1950. [42] A. Niculescu-Mizil and R. Caruana, “Predicting good probabilities with supervised learning,” in Proc. 22nd int. Conf. Mach. Learn. (ICML), pp. 625–632, 2005. [43] B. Efron and R. J. Tibshirani, An introduction to the bootstrap, Chapman & Hall, 1993. [44] E. R. DeLong, D. M. DeLong and D. L. Clarke-Pearson, “Comparing the areas under two or more correlated receiver operating characteristic curves: A nonparametric approach,” Biometrics, vol. 44, no. 3, pp. 837–845, 1988. [45] T. G. Dietterich, “Approximate statistical tests for comparing supervised classification learning algorithms,” Neural Comput., vol. 10, no. 7, pp. 1895–1923, 1998. [46] C. Nadeau and Y. Bengio, “Inference for the generalization error,” Mach. Learn., vol. 52, no. 3, pp. 239–281, 2003. [47] J. Demšar, “Statistical comparisons of classifiers over multiple data sets,” J. Mach. Learn. Res., vol. 7, pp. 1–30, 2006. [48] S. Holm, “A simple sequentially rejective multiple test procedure,” Scand. J. Statist., vol. 6, no. 2, pp. 65–70, 1979. [49] S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” in Advances in neural information processing systems (NIPS), pp. 4765–4774, 2017. [50] M. T. Ribeiro, S. Singh and C. Guestrin, ““Why should I trust you?”: Explaining the predictions of any classifier,” in Proc. 22nd ACM SIGKDD int. Conf. Knowl. Discovery data mining, pp. 1135–1144, 2016. [51] A. Fisher, C. Rudin and F. Dominici, “All models are wrong, but many are useful: Learning a variable’s importance by studying an entire class of prediction models simultaneously,” J. Mach. Learn. Res., vol. 20, no. 177, pp. 1–81, 2019. [52] J. H. Friedman, “Greedy function approximation: A gradient boosting machine,” Ann. Statist., vol. 29, no. 5, pp. 1189–1232, 2001. [53] C. Rudin, “Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead,” Nat. Mach. Intell., vol. 1, pp. 206–215, 2019. [54] M. Hardt, E. Price and N. Srebro, “Equality of opportunity in supervised learning,” in Advances in neural information processing systems (NIPS), pp. 3315–3323, 2016. [55] Z. Obermeyer, B. Powers, C. Vogeli and S. Mullainathan, “Dissecting racial bias in an algorithm used to manage the health of populations,” Science, vol. 366, no. 6464, pp. 447–453, 2019.

Copyright

Copyright © 2026 Prajwal R P. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.