A Personalised Clinical Dynamic Prediction Model to Characterise Prognosis for Patients with Localised Prostate Cancer: analysis of the CHHiP Phase III Trial.

Abstract

BACKGROUND: The CHHiP trial assessed moderately hypofractionated radiotherapy in localised prostate cancer. We utilised longitudinal prostate-specific antigen (PSA) measurements collected over time to evaluate and characterise patient prognosis. METHODS: We developed a clinical dynamic prediction joint model to predict the risk of biochemical or clinical recurrence. Modelling included repeated PSA values and adjusted for baseline prognostic risk factors of age, tumour characteristics and treatment received. We included 3,071 trial participants for model development using a mixed-effect submodel for the longitudinal PSAs, and a time-to-event hazard submodel for predicting recurrence of prostate cancer. We evaluated how baseline prognostic factor subgroups impacted on the nonlinear PSA levels over time and quantify the association of PSA on time-to-recurrence. We assessed bootstrapped optimism-adjusted predictive performance on calibration and discrimination. Additionally, we performed comparative dynamic predictions on patients with contrasting prognostic factors and investigated PSA thresholds over landmark times to correlate with prognosis. RESULTS: Patients that developed recurrence had generally higher baseline and overall PSA values during follow-up and had an exponentially rising PSA in the two-years before recurrence. Additionally, most baseline risk factors were significant in the mixed-effect- and relative risk submodels. PSA value- and rate-of-change was predictive of recurrence. Predictive performance of the model was good across different prediction times over an 8-year period, with an overall mean AUC of 0.70, mean Brier score of 0.10, and mean integrated calibration index of 0.048; these were further improved for predictions after 5 years of accrued longitudinal post-treatment PSA assessments. PSA thresholds less than 0.23ng/mL after 3 years were indicative of a minimal risk of recurrence by 8 years. CONCLUSIONS: We successfully developed a joint statistical model to predict prostate cancer recurrence, evaluating prognostic factors and longitudinal PSA. We showed dynamically updated PSA information can improve prognostication, which can be used to guide follow-up and treatment management options.