Discrete Survival Knowledge Distillation for Competing Risks Analysis

Accurate prediction in survival analysis with competing risks is challenged by rare event rates and limited effective sample sizes. Knowledge distillation offers a promising way to transfer information from an external teacher to improve a local student, but existing methods are overwhelmingly developed for uncensored outcomes and do not directly extend to survival analysis, where censored observations provide only partial information. Moreover, prior work often assumes that teacher and student share identical outcome definitions, whereas in competing risks settings, they may differ in outcome granularity and event definitions, further complicating knowledge transfer. To address these gaps, we propose DiSKD (Discrete Survival Knowledge Distillation), a deep learning framework for discrete-time competing risks that integrates teacher predictions via a cause-specific, time-dependent Kullback-Leibler divergence. DiSKD enables flexible and privacy-conscious transfer without requiring raw data sharing, remains robust to model misspecification or outcome-definition heterogeneity, and adaptively weights teacher guidance by emphasizing compatible teachers while down-weighting less relevant ones. Simulation studies and real-world applications demonstrate improved discrimination and calibration.

Many medical decisions depend on predicting not only whether a patient may experience a bad outcome, but also which type of outcome may happen and when. This is especially important in kidney transplantation, where doctors and patients must weigh different risks, such as graft failure and patient death, when deciding whether to accept a donated kidney. However, accurate prediction is difficult because these serious outcomes are rare, many patients have limited follow-up, and recent clinical practice can differ from older data. This paper proposes DiSKD, a machine learning method that helps a new prediction model learn from existing models without requiring access to the original patient records used to train them. Instead, DiSKD uses the predictions from an external model as guidance while still adapting to the local patient population. The method is designed for settings where different outcomes compete with each other, such as death versus graft failure, and where the external model may use a different outcome definition or a different set of patient information. Across simulations and real-world kidney transplant data, DiSKD improves prediction accuracy and reliability compared with training only on the local data. More broadly, this work provides a practical way to reuse medical prediction knowledge across institutions and time periods while respecting data-sharing constraints.