Externally Validated Longitudinal GRU Model for Visit-Level 180-Day Mortality Risk in Metastatic Castration-Resistant Prostate Cancer

TL;DR

This study tackles the challenge of predicting short-horizon mortality in metastatic castration-resistant prostate cancer by leveraging longitudinal, visit-level data. It develops and externally validates a GRU-based model that updates risk at each clinical encounter, compared against five competing architectures, and uses an 85% sensitivity threshold to reflect clinical safety requirements. The GRU demonstrated strong external calibration (slope ≈ 0.93) and discrimination (AUC ≈ 0.89) with a median lead time of ~151 days and a modest alert burden (~18 alerts per 100 visits), while permutation analyses highlighted BMI and systolic BP as key drivers. The findings support integrating longitudinal risk signals into proactive care planning and indicate the approach generalizes across distribution shifts, with implications for automated EHR-based decision support in late-stage cancer care.

Abstract

Metastatic castration-resistant prostate cancer (mCRPC) is a highly aggressive disease with poor prognosis and heterogeneous treatment response. In this work, we developed and externally validated a visit-level 180-day mortality risk model using longitudinal data from two Phase III cohorts (n=526 and n=640). Only visits with observable 180-day outcomes were labeled; right-censored cases were excluded from analysis. We compared five candidate architectures: Long Short-Term Memory, Gated Recurrent Unit (GRU), Cox Proportional Hazards, Random Survival Forest (RSF), and Logistic Regression. For each dataset, we selected the smallest risk-threshold that achieved an 85% sensitivity floor. The GRU and RSF models showed high discrimination capabilities initially (C-index: 87% for both). In external validation, the GRU obtained a higher calibration (slope: 0.93; intercept: 0.07) and achieved an PR-AUC of 0.87. Clinical impact analysis showed a median time-in-warning of 151.0 days for true positives (59.0 days for false positives) and 18.3 alerts per 100 patient-visits. Given late-stage frailty or cachexia and hemodynamic instability, permutation importance ranked BMI and systolic blood pressure as the strongest associations. These results suggest that longitudinal routine clinical markers can estimate short-horizon mortality risk in mCRPC and support proactive care planning over a multi-month window.

Figures (4)

• Figure 1: Comparison of Model Calibration for 180-Day Mortality Risk on the Development Cohort. The reliability diagram illustrates the predictive accuracy of the candidate architectures at the visit level. This plots the mean predicted mortality risk against the observed proportion of death events. Perfect calibration is shown as the diagonal line.
• Figure 2: Mortality Risk Calibration and Distribution for the External Validation Cohort. Both models are evaluated on their ability to predict death within a 180-day look-ahead window after each index visit date. Thresholds are optimized to prioritize high sensitivity for clinical alerting, with the GRU model leveraging temporal markers to update risk scores at each subsequent clinical visit.
• Figure 3: Comprehensive Operating-Characteristic Evaluation of the GRU Risk Model on the External Validation Cohort at the Visit Level. (a) ROC curve with 95% confidence interval. To account for within-patient correlation in visit-level evaluation, we estimated confidence bands via patient-clustered nonparametric resampling, with details provided in the code. (b) Precision--Recall curve highlighting performance under class imbalance relative to the baseline event rate. (c) Sensitivity and specificity as a function of risk-score threshold, with the chosen operating threshold $\theta$ selected to satisfy the predefined sensitivity constraint. (d) PPV and NPV as a function of threshold, showing the resulting false-alarm versus missed-event trade-off at the selected 85% sensitivity target.
• Figure 4: Permutation-Based Feature Importance Ranking for the GRU survival model on the Internal Validation Cohort. To quantify importance, we measured the mean decrease in the Concordance Index (C-index) after permuting each feature across 3-fold stratified cross-validation. Horizontal bars show the mean decrease in C-index, and error bars show $\pm 1$ standard deviation (SD). The dots show fold-specific importance scores.