基于人工智能对可穿戴设备衍生生物特征的分析以刻画实体器官癌症化疗的生理反应

Objective: (1) To use wearable-derived biometrics to examine the relationship between physiologic patterns and adverse events during chemotherapy for solid organ cancers. (2) To use machine-learning analysis of these physiologic patterns to predict adverse events. Methods: Patients with solid organ cancers receiving systemic therapy were enrolled in an ongoing trial evaluating the effects of chemotherapy/ immunotherapy on physiologic patterns. Patients continuously wore a Fitbit for >5 days before and during treatment. We analyzed intra- and inter-personal variation in normalized daily average resting heart rate (RHR) across cycles. We developed a novel, personalized, baseline-referenced statistical framework based on the Change-of-Heart machine learning algorithm, “HeartSense,” which standardizes each participant's RHR against their own baseline during sleep and non-sleep periods, generating a score that quantifies daily physiologic stability. Higher scores reflect RHR patterns consistent with pre-treatment baseline and lower scores indicate greater deviation from baseline. We compared lowest HeartSense scores by patients who experienced adverse events vs those who did not in cycle 1. We also examined whether scores from cycle 1 were associated with adverse events in cycle 2. Results: A total of 54 cancer patients who underwent > 1 cycle of systemic therapy were included in analysis. Distinct physiologic patterns emerged during therapy. One patient showed cyclic decreases in heart rate after each infusion with recovery to baseline (Figure A), while another exhibited progressive physiologic disruption preceding treatment termination (Figure B). Intrapersonal variation in sleep and non-sleep RHR was significantly smaller than interpersonal variation (Figure C). Lower median HeartSense scores within a cycle were significantly more likely to experience adverse events (Figure D). Lower scores were also associated with adverse events in cycle 2 (Figure E). Conclusions: Patients exhibit distinct physiologic changes in resting heart rate during systemic therapy and deviation from normal patterns may indicate treatment-related adverse events. Our novel machine learning algorithm identified that lower scores-reflecting abnormal physiologic regulation-were linked to a higher risk of adverse events both during and after the current treatment cycle, suggesting that chemotherapy-related toxicity can be preceded by early physiologic warning signals.