Wearable-derived digital biomarkers for predicting systemic therapy toxicity and survival in oncology.

Severe (Grade ≥3) systemic therapy toxicities compromise treatment delivery and survival in oncology. Traditional risk stratification using ECOG performance status and laboratory values lacks real-time granularity. In this study, we evaluated whether wearable-derived digital biomarkers enhance the prediction of toxicity and survival beyond clinical standards and conducted a retrospective multicentre cohort study (N = 720) of adults with breast, lung, or colorectal cancer initiating systemic therapy (2018-2025). Primary exposures included baseline and early-treatment (days 1-30) heart rate variability (HRV-RMSSD), daily steps, and sleep efficiency. The primary outcome was Grade ≥ 3 toxicity (CTCAE v5.0) within six months. Gradient boosting machines compared wearable versus clinical model performance (AUC). Cox models assessed survival associations with stratification by cancer type and therapy class. Mediation analysis quantified baseline activity's role in the HRV-toxicity pathway. The clinical model (ECOG, albumin, demographics) significantly outperformed the wearable-only model for toxicity prediction (AUC 0.797 vs. 0.601; Δ=-0.20, p < 0.001). However, the prognostic value of HRV decline was highly heterogeneous: in lung cancer patients receiving immunotherapy, early HRV decline was associated with improved survival (HR=0.58, 95% CI: 0.48-0.70). Baseline steps mediated 44% (95% CI: 36-52%) of the HRV decline-toxicity relationship (p < 0.001), indicating a modifiable resilience pathway. While clinical factors remain superior for acute toxicity prediction, wearable biomarkers provide therapy-specific prognostic insights, particularly for immunotherapy response, and identify baseline fitness as a quantifiable mediator of treatment tolerance. Integration of digital metrics should be context-specific rather than universal.