TrajSurv-mPC: Learning From Prostate-Specific Antigen Trajectories for Interpretable and Generalizable Survival Prediction in Metastatic Prostate Cancer.

PURPOSE: Conventional prognostic models for metastatic prostate cancer (mPC) typically rely on summary features like prostate-specific antigen (PSA) doubling time, often obscuring informative longitudinal dynamics. We developed TrajSurv-mPC, a flexible deep learning framework that analyzes full premetastasis patient trajectories-including serial PSA and treatments-to improve postmetastasis overall survival (OS) prediction.

METHODS

The TrajSurv-mPC framework was developed and evaluated using two independent cohorts: Memorial Sloan Kettering (MSK; n = 403) and Veterans Affairs (VA; n = 9,984). To demonstrate its architectural robustness, the framework's sequential modeling was implemented with both Long Short-Term Memory (LSTM) and Transformer backbones. The models used baseline characteristics alongside longitudinal PSA and treatment data to predict OS from the time of metastasis detection. Performance was compared against models using summary features. Generalizability was assessed via external validation (training on a harmonized VA subcohort; testing on MSK). Interpretability was explored through post hoc hidden state clustering, a simulation study, and a case study.

RESULTS

Compared with summary-feature models (C-indices: 0.650-0.677 internally; 0.635-0.651 externally), the TrajSurv-mPC framework achieved C-indices of 0.723-0.747 (MSK) and 0.711-0.754 (VA) in internal validations, and 0.693 (LSTM, P = .02) and 0.708 (Transformer, P = .008) in cross-cohort external validations across both architectures. Clustering stratified distinct prognostic groups (P < .001). Furthermore, simulation and case studies showed that predicted risk varied in clinically plausible directions across different treatment response trajectories.

CONCLUSION

By modeling the complete premetastasis PSA trajectory, the TrajSurv-mPC framework improves prognostic risk prediction in mPC. Its robust performance across different sequential backbones demonstrates the generalizable added value of dynamic patterns in distinct populations. Combining this framework with complementary measures based on imaging or genomics may enhance personalization of mPC management.