Predictive Model Estimates 180-Day Mortality Risk in Advanced Prostate Cancer

Global: Predictive Model Estimates 180-Day Mortality Risk in Advanced Prostate Cancer

Researchers have introduced a visit-level model that predicts the probability of death within 180 days for patients with metastatic castration-resistant prostate cancer (mCRPC). The model was built and externally validated using longitudinal data from two Phase III clinical cohorts comprising 526 and 640 patient visits, respectively. By labeling only visits with observable outcomes and excluding right-censored cases, the investigators aimed to provide clinicians with an early warning system for high-risk patients. The work was submitted to arXiv in January 2026.

Data and Cohorts

The development dataset consisted of 526 visits drawn from a Phase III trial, while a separate cohort of 640 visits served as the external validation set. Both datasets captured routine clinical measurements recorded at each patient encounter, enabling a visit-level analysis of mortality risk over a 180‑day horizon.

Model Architectures Compared

Five candidate approaches were evaluated: Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Cox proportional hazards, Random Survival Forest (RSF), and logistic regression. Each architecture leveraged the sequential nature of the longitudinal records to estimate short‑term survival probabilities.

Selection Criteria and Internal Performance

For each model, the smallest risk threshold achieving at least 85% sensitivity was identified. Internally, the GRU and RSF demonstrated the highest discrimination, each attaining a concordance index of 0.87. These results guided the selection of the GRU for further external testing.

External Validation Results

When applied to the independent cohort, the GRU model maintained strong calibration, with a slope of 0.93 and an intercept of 0.07. The precision‑recall area under the curve reached 0.87, indicating robust performance in identifying patients at imminent risk of death.

Clinical Impact Assessment

Impact analysis revealed a median time‑in‑warning of 151.0 days for true‑positive alerts, compared with 59.0 days for false‑positive alerts. The system generated approximately 18.3 alerts per 100 patient visits, suggesting a manageable alert frequency for clinical workflows.

Feature Importance

Permutation‑based importance ranking highlighted body‑mass index and systolic blood pressure as the most influential predictors, reflecting the relevance of frailty and hemodynamic stability in short‑term mortality among mCRPC patients.

Implications for Care Planning

The findings support the feasibility of using routinely collected clinical markers to forecast near‑term mortality, potentially enabling proactive care planning and timely interventions for individuals with advanced prostate cancer. Further research is warranted to assess integration into electronic health record systems and to evaluate performance across diverse patient populations.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.