Dynamic survival predictions get time-resolved explanations: DynSurvX cuts computation by 41–80% and can refresh them during follow-up
Synopsis
Longitudinal follow-up data make survival predictions more accurate but harder to interpret, and existing explanation methods are either tied to specific model classes or limited to static covariates. DynSurvX splits temporal attribution into two steps: it first screens longitudinal features by overall importance, then adaptively subdivides time windows only for those features, yielding time-resolved explanations that update as new measurements arrive. On simulated and two clinical datasets, agreement with theoretical importance rises with prediction-model performance, and computation time falls 41–80% relative to the temporal explanation baseline WindowSHAP.
Interpretation
DynSurvX provides time-resolved explanations for dynamic survival prediction with both static covariates and longitudinal measurements, showing which features matter and during which part of follow-up the longitudinal measurements affect the prediction, updating as new data arrive. Earlier survival explanation methods were either tailored to specific model classes or restricted to static covariates and could not show how longitudinal importance evolves during follow-up; temporal attribution methods such as TimeSHAP, WindowSHAP and OrdSHAP were built mainly for regression and classification and attribute jointly over feature–time space at high cost. On simulated data, DynSurvX recovers overall and temporal importance more accurately than JointLIME and WindowSHAP; JointLIME misses several influential features, while Dynamic WindowSHAP detects the main longitudinal features but gives less accurate importance estimates and temporal localization.
The approach treats temporal explanation as a hierarchical attribution problem: Phase 1 uses KernelSHAP in the mixed static–longitudinal setting to score each longitudinal trajectory as a single feature, and only features above a threshold pass to Phase 2. Phase 2 extends Dynamic WindowSHAP: each selected trajectory starts as one interval and is recursively subdivided, with only intervals whose normalized importance exceeds a threshold split further, and the threshold updates adaptively during refinement so each prediction gets an appropriate temporal resolution. The evaluation spans one simulated dataset and two clinical datasets (312 primary biliary cirrhosis patients in PBCseq and 3,253 patients in the MSK-Chord breast cancer cohort), three dynamic survival models (Dynamic DeepHit, CoxSig and DySurv), and two prediction targets.
Perspective
The demonstrated scope is three dynamic survival models on one simulated and two clinical datasets, with expected remaining restricted survival time and median residual survival time as prediction targets; explanation quality depends on the underlying prediction model, and on MSK-Chord all three models show moderate predictive performance with markedly different attribution patterns.
The two-phase design may discard longitudinal features with low overall importance but strong localized temporal effects; varying the threshold around the elbow-selected value leaves PBCseq patterns largely unchanged, but this robustness does not remove that limitation, and the authors recommend treating the threshold as tunable and running sensitivity analysis. Interventional Shapley approximations may generate unrealistic perturbations when features are strongly dependent, and as a post-hoc framework it inherits the underlying model's limitations under distribution shift.
