Wearable-derived physical activity monitoring with AI-based detection of treatment-related adverse events during neoadjuvant chemotherapy for early breast cancer: TRACK-BC.

1631 Background: Neoadjuvant chemotherapy (NAC) for early breast cancer (EBC) is commonly delivered in the outpatient setting, yet within-cycle changes in adverse events reduce physical activity (PA). Wearable devices may capture these short-term changes. However, prospective data describing activity patterns during modern regimens, including pembrolizumab (PEMB)-based therapy, are limited. This study evaluated changes in PA using Fitbit-derived digital biomarkers (dBM) and explored associations with patient-reported outcomes (PROs). Methods: We conducted a multicenter, prospective observational study in patients (pts) with stage I–IIIA EBC initiating NAC. Eligible pts were ≥18 years old, smartphone users, and able to wear a Fitbit daily. Chemotherapy regimens were grouped into weekly, q2w, or q3w cohorts. Fitbits were worn from ≥7 days before treatment initiation through four cycles. Total PA (TPA), moderate-to-vigorous PA (MVPA), heart rate, sleep, and other metrics were collected. PRO-CTCAE and HADS were completed weekly electronically (ePROs). The primary outcome was change in TPA and MVPA from baseline to the week following cycle 2. As a secondary endpoint, machine learning approaches extending conventional multivariable regression incorporated multiple wearable-derived parameters to individually predict the worst change in ePRO scores from baseline. Model performance was evaluated by cross-validation, and feature contributions were examined using SHapley Additive exPlanations (SHAP) to enhance clinical interpretability. Results: A total of 101 pts were enrolled across the three predefined chemotherapy cohorts. Fitbit use was feasible with high adherence. The primary outcome indicated a consistent decline in PA during NAC. From baseline to the week following cycle 2, mean TPA decreased by 141.4 METS-min/day and MVPA by 33.7 min/day. ePROs revealed worsening symptoms including fatigue, insomnia, appetite loss, and neuropathy. Symptoms varied according to the chemotherapy schedule and generally aligned with changes in PA levels and dBM. The secondary outcomes identified multiple vital signals showing absolute correlations >0.3 with all PRO and HADS metrics. For fatigue, the best model (ridge regression) yielded mean R² 0.33±0.20 and RMSE 1.86±0.26; SHAP analysis highlighted contributions from nocturnal heart-rate-variability minima and overall variability. Conclusions: Continuous wearable monitoring during NAC for EBC was feasible. Objective declines in PA occurred early and paralleled symptom worsening captured by ePROs. AI-based models explained a modest but clinically meaningful proportion of symptom deterioration with acceptable error. These findings support the potential utility of wearable-derived dBM for real-time surveillance of treatment-related toxicities during NAC. Clinical trial information: UMIN000053991.