DynSurvX: Time-Resolved Explanations for Dynamic Survival Predictions

Van Tuan Nguyen, Agathe GuillouxarXiv 2026 · arXiv

TL;DR

DynSurvX is a new model-agnostic framework that brings "time-resolved" explainability to dynamic survival models. By using a hierarchical two-phase strategy based on Shapley values, it identifies not only which biomarkers drive a patient's risk but also exactly when those measurements became influential, all while reducing computational costs by up to 80%.

What Problem Does This Paper Address?

In modern healthcare, survival analysis—predicting the time until an event like death or disease recurrence—is increasingly "dynamic." Instead of a one-time prediction, models are updated as new longitudinal data (like repeated blood tests or scans) arrive. While deep learning models excel at this, they are often "black boxes," making it hard for clinicians to trust their outputs.

Existing explainability tools like SHAP are either designed for static data or are too computationally expensive to handle the "feature-time" space of longitudinal data. Clinicians need to know if a high-risk score is due to a patient's age (static) or a specific trend in a biomarker over the last few months. DynSurvX fills this gap by providing efficient, fine-grained temporal explanations.

Methods

DynSurvX is built on the foundation of Shapley values, a game-theoretic approach to feature attribution. Its core innovation is a hierarchical two-phase strategy:

  1. Phase 1: Feature Screening: The framework first treats each entire longitudinal trajectory as a single feature. It calculates overall importance scores and employs an "elbow method" to filter out non-influential features.
  2. Phase 2: Adaptive Temporal Refinement: For the influential features identified in Phase 1, DynSurvX zooms in. It adaptively partitions the time series into intervals. It recursively subdivides these intervals only where the "importance signal" is strong, ensuring high resolution where it matters most without wasting compute on less informative time periods.

This approach is model-agnostic, meaning it can be wrapped around any dynamic survival model, from traditional joint models to deep neural networks.

Datasets

The authors evaluated DynSurvX across three diverse settings:

  • Simulated Dataset: Generated using an Ornstein–Uhlenbeck process, providing a "ground truth" for theoretical feature and temporal importance.
  • PBCseq: A clinical dataset of 312 patients with Primary Biliary Cirrhosis, featuring five longitudinal biomarkers (e.g., serum bilirubin, albumin) and two static features.
  • MSK-Chord: A breast cancer cohort (3,253 patients) from Memorial Sloan Kettering, focusing on two longitudinal tumor markers (CEA and CA15-3) and five static covariates.

Key Insights and Ideas

  • Explanations are Model-Dependent: A critical finding is that the "quality" of an explanation depends on the underlying model's performance. As shown in Figure 1, models with better discrimination and calibration (higher C-index and lower Brier scores) produced explanations that aligned much more closely with the theoretical ground truth. Predictive performance of different models on simulated data
  • Efficiency is the Enabler: By decoupling feature screening from temporal refinement, DynSurvX makes real-time clinical use feasible. It avoids the "curse of dimensionality" inherent in joint feature-time attribution.
  • Capturing "Recency" Effects: In clinical datasets like PBCseq, DynSurvX successfully identified that recent measurements often exert more influence on survival predictions than older ones, a pattern consistent with clinical intuition but often hidden in black-box models.
Predictive performance of different models on simulated data

Key Results

  • Significant Speedup: DynSurvX reduced computation time by 41% to 80% compared to the WindowSHAP baseline. This makes repeated updates to explanations practical as new patient data flows in. Runtime comparison showing DynSurvX's efficiency gains
  • Superior Accuracy: On simulated data, DynSurvX more accurately recovered both global feature importance and the specific timing of those contributions compared to JointLIME and WindowSHAP. Comparison of feature importance recovery against baselines
  • Clinical Interpretability: In the MSK-Chord breast cancer cohort, the framework correctly identified tumor markers CEA and CA15-3 as the most influential longitudinal features, aligning with established oncological knowledge.
Runtime comparison showing DynSurvX's efficiency gains
Comparison of feature importance recovery against baselines

Limitations and Future Work

The authors acknowledge that DynSurvX, like many Shapley-based methods, uses "interventional" approximations which might generate unrealistic data perturbations if features are highly correlated. Furthermore, as a post-hoc method, it can only reflect what the model has learned; if the model is biased or inaccurate, the explanation will faithfully reflect those flaws.

Future directions include:

  • Integrating conditional Shapley estimators to better handle feature dependencies.
  • Developing uncertainty-aware explanations to signal when the explainer itself is unsure.
  • Extending the framework to competing-risk and multi-state survival models.

Summary updated

Abstract

Recent survival models integrating longitudinal data achieve strong predictive performance, but their predictions are often difficult to interpret, limiting their clinical use. We propose DynSurvX, a model-agnostic explainability framework for dynamic survival prediction with mixed static and longitudinal covariates. Based on Shapley values, DynSurvX introduces a hierarchical two-phase strategy that first identifies influential longitudinal features and then adaptively refines their temporal importance, enabling fine-grained explanations while reducing computational cost. The framework provides time-resolved explanations that evolve as new patient data become available. Experiments on simulated and two clinical datasets show varying agreement with theoretical importance across prediction models, with closer agreement for better-performing models, alongside clinically interpretable insights in real-world settings. It remains robust across prediction settings and reduces computation time by 41--80\% relative to the temporal explanation baseline, making dynamic explanations practical.
BibTeX
@misc{nguyen2026dynsurvx,
  author = {Van Tuan Nguyen and Agathe Guilloux},
  title = {{DynSurvX: Time-Resolved Explanations for Dynamic Survival Predictions}},
  year = {2026},
  eprint = {2611.21571},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.21571},
}
Figure