Patient-reported outcomes as early indicators of recurrence, disease progression, and survival via machine learning in breast cancer.

Abstract

Patient-reported outcomes (PROs) offer a non-invasive, low-cost way to capture patients' symptoms, functioning, and quality of life. Yet, their potential as early indicators of tumor size, recurrence/disease progression, and survival remains unclear. We retrospectively analyzed 445,239 longitudinal PRO entries from 2738 patients with breast cancer, pooled from four clinical trials, including both early- and late-stage disease, covering 15 PRO measures. Among patients with radiographically confirmed recurrence or disease progression, 89.2% experienced at least one PRO deterioration prior to relapse detection (85 vs. 706 days), indicating that PROs often worsen before imaging-confirmed relapse. Cox proportional hazards models showed that PRO deterioration was significantly associated with metastatic sites, tumor size, and survival. Appetite loss was most strongly correlated with tumor size, while pain and diarrhea were the most prognostic symptoms for overall survival (OS) and progression-free survival (PFS). Gradient boosting models further showed that combining PRO deterioration times across all subscales with PFS best predicted OS, correctly classifying survival outcomes in over 93% of cases (AUC - ROC = 0.932), outperforming models using PROs alone (AUC - ROC = 0.806) or PFS alone (AUC - ROC = 0.88). This indicates that integrating PROs with PFS enhances the prediction of OS, providing a more powerful approach than using either measure alone. These findings suggest that PROs can serve as early, complementary predictors of recurrence, disease progression, and survival, supporting their use as patient-centered biomarkers in breast cancer management. Our findings align with FDA and EMA efforts to integrate PROs into oncology endpoints, supporting more patient-centered regulatory evaluation.