Transforming Patient Voices into Early Predictors of Survival Using Nonlinear Mixed-Effect Models and AI/ML for Patient-Centered Decision-Making

Patient-reported outcomes (PROs) capture the patient voice and have been associated with improved clinical outcomes in oncology, but their prognostic and predictive value remains underutilized due to challenges in interpreting these highly variable and noisy PRO data. Here, we developed a quantitative modeling framework integrating nonlinear mixed-effects (NLME) and item response theory (IRT) to characterize symptom-level PRO trajectories and transform them into clinically actionable predictors. Using longitudinal PRO data from 589 patients with metastatic cancers in the PRO-TECT trial, we modeled 332,920 symptom responses to estimate patient-specific PRO trajectory parameters while accounting for variability and noise. IRT-NLME modeling captured heterogeneous symptom-level PRO dynamics and is more informative than modeling with composite PRO scores. PRO trajectory parameters were strongly associated with overall survival, acute care utilization, and treatment modifications. Machine learning models leveraging these parameters achieved robust prediction of survival (AUC-ROC 0.80) and retained prognostic performance using the first 30 - 180 days of PRO observations, with AUCs of 0.69-0.78. Similar predictive performance was observed for hospitalization (AUC 0.75), emergency department visit (AUC 0.65), treatment discontinuation (AUC 0.71), and dose reduction (AUC 0.67). These findings demonstrate that longitudinal PRO trajectories can serve as early, patient-centered biomarkers of clinical risk. By converting complex symptom data into interpretable and predictive metrics, this quantitative framework provides a practical pathway to integrate the patient voice into clinical decision-making and advance precision oncology. ClinicalTrials.gov registration: NCT03249090