Identification of Early Signs of Mental Health Disorders in Older Survivors of Cancer Using Patient-Generated Health Data: Observational Study

Clinical Snapshot

Key Findings

• P Value: Not reported for classification performance

• Effect Size: Best performing model (smart plug + smart scale): F1-score 0.77, AUC 0.85

• Primary Outcome: Classification of PHQ-4-defined mental health risk using patient-generated health data

• Nnt Or Sensitivity: Smart plug alone achieved AUC 0.71 (SD 0.06), demonstrating reasonable standalone performance

• Confidence Interval: Standard deviations reported: F1-score SD 0.15, AUC SD 0.10

Clinical Application

Requires patient acceptance of multiple monitoring devices and sustained adherence over time Could complement existing cancer survivorship care plans under Medicare; would need TGA approval for clinical use and integration with My Health Record Older cancer survivors in community settings with access to smart home technologies

Abstract

BACKGROUND: Older survivors of cancer face heightened risk of depression and anxiety related to cancer experiences, fear of recurrence, and aging-related difficulties. Conventional mental health monitoring approaches, such as clinical assessments and even electronic patient-reported outcomes, are limited by recall bias, patient burden, and infrequent data collection. Emerging patient-generated health data from wearables and smart home devices offer passive, low-burden, continuous monitoring, but their ability to capture mental health risks in older survivors of cancer remains unclear. OBJECTIVE: This study aims to explore whether patient-generated health data collected in the wild, either passively or actively, can classify older survivors of cancer as having or not having signs of anxiety and depression based on Patient Health Questionnaire-4 (PHQ-4) scores and to assess the potential added value of passive monitoring modalities, such as smart plugs. METHODS: This study recruited 41 older survivors of cancer (mean age 72.3, SD 6.81 years) from the LifeChamps project. Over a 12-week period, participants were monitored using an activity tracker to measure physical activity, sleep, and physiological metrics; a smart scale to capture weight and body composition; and a smart plug to track television use as a proxy for sedentary television viewing. Mental health status was self-reported via the PHQ-4 questionnaire in a mobile app. Machine learning models were trained to classify mental health risk based on features derived from each sensor modality, both independently and in combination. RESULTS: Tree-based gradient boosting models showed good performance in classifying PHQ-4-defined mental health risk. The best-performing configuration, combining smart plug and smart scale features, achieved a mean F1-score of 0.77 (SD 0.15) and a mean area under the receiver operating characteristic curve (AUC) of 0.85 (SD 0.10) across 3 repeated train-test splits. Standalone smart plug models, based solely on passive television use patterns, achieved a mean F1-score of 0.66 (SD 0.04) and a mean AUC of 0.71 (SD 0.06), outperforming models that relied only on activity tracker data (mean F1 0.59, SD 0.2). Multimodal combinations tended to improve average performance but did not consistently yield large gains over the strongest single-modality configurations, likely reflecting adherence-related data loss for wearables and scales. Crucially, passive monitoring of television use patterns emerged as a promising behavioral proxy measure of mental health states. CONCLUSIONS: This study pioneers the use of passively collected data (eg, smart plugs) for mental health monitoring in older survivors of cancer, demonstrating their potential. Smart plugs capture behavioral patterns without user burden, with reasonable standalone performance (mean AUC 0.71, SD 0.06), positioning them as a promising low-burden modality. Future work should validate findings in larger independent cohorts and in prospective clinical workflows. Such technologies could transform monitoring for vulnerable populations, enabling scalable, inclusive care while reducing health care burdens.

References

1. 1.Petridis, G., Billis, A., Meditskos, G., Tsanousa, A., Lagakis, P., Naranjo, J. C., Zenóglio de Oliveira, J., Tavares, V., & Bamidis, P. D. (2026). Identification of Early Signs of Mental Health Disorders in Older Survivors of Cancer Using Patient-Generated Health Data: Observational Study. JMIR Cancer. https://doi.org/10.2196/75050