Speaker
Description
Patient monitoring after radiotherapy for prostate cancer relies on population-based thresholds of rising prostate-specific antigen (PSA), ignoring patient-specific tumor dynamics and uncertainty. To avoid delays in recurrence detection and treatment, we propose a personalized Bayesian mechanistic framework to forecast post-radiotherapy PSA dynamics. These forecasts enable the definition of risk-aware biomarkers of biochemical relapse, such as surviving tumor cell proliferation rate and time to progression. The mechanistic model is calibrated using longitudinal measurements, yielding posterior distributions of PSA over time and the model-based biomarkers. To quantify biochemical relapse risk, we summarize posterior distributions using α-superquantiles to capture adverse tail behavior and evaluate them using leave-one-out cross-validated ROC analyses and comparison across post-treatment time horizons. The most informative model-based biomarkers achieve high discriminative performance, which is superior to standard PSA-based metrics (e.g., PSA nadir, time to PSA nadir). To further assess clinical relevance, we introduce a days gained metric, representing the time gained by model-based relapse detection compared with standard-of-care PSA-based criteria (e.g., nadir+2 ng/mL). Pareto front analyses reveal trade-offs between classification accuracy (sensitivity/specificity) and early detection, enabling risk-aware selection of biomarker thresholds for clinical decision-making.
