Research map / Probabilistic and scientific modeling / Molecules and proteins

Gradient-based Explanations for Deep Learning Survival Models

ICML 2025 · Probabilistic and scientific modeling › Molecules and proteins · approach: Sparse concepts and hypothesis testing · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Extends gradient-based attribution methods (Saliency, SmoothGrad, Gradient×Input, Integrated Gradients, GradSHAP) to deep survival networks with time-dependent explanations and proposes GradSHAP(t) as a faster gradient-based counterpart to SurvSHAP(t).

Problem

Deep learning survival models are black boxes, and existing survival-specific XAI methods (SurvLIME, SurvSHAP(t)) are model-agnostic and slow; gradient-based explanations have not been generalized to time-dependent survival outputs.

Approach

Applies gradient-based attribution at each discretized time point of survival-network outputs (survival, hazard, CHF), yielding time-dependent relevance curves, and introduces GradSHAP(t), which averages Integrated Gradients(t) over reference samples drawn from the data to approximate SHAP values over time.

Key results

• GradSHAP(t) achieves local accuracy comparable to SurvSHAP(t) while significantly reducing runtime, and outperforms SurvSHAP(t) in runtime across all feature set sizes and model classes.
• On the multi-modal glioma model (ResNet34 + DeepHit), C-index 0.713 and integrated Brier score 0.092 were achieved on 266 test samples.
• Global importance rankings of GradSHAP(t) and SurvSHAP(t) are nearly identical and match the data-generating process, while SurvLIME rankings are more uniform.

Measured effects reported in the paper

Surprising findings

• Gradient×Input yields only positive attribution curves for a feature with negative global effect when the feature value is negative, showing misleading interpretations due to local linearity assumption at zero.
• DeepSurv cannot capture time-dependent effects due to PH constraint, which gradient-based explanations reveal.

Limitations

• Explanations reflect model behavior, not causal relationships.
• Restricted to standard right-censored survival; competing risks, multi-state and recurrent events not covered.
• Gradient×Input relies on a local linearity assumption at zero that is poor for nonlinear survival curves.

Experimental setup

Datasets: simulated Weibull/Cox survival data TCGA LGG and GBM whole-slide histology and clinical data (diffuse glioma)

Benchmarks: simulated survival data with known ground-truth effects TCGA diffuse glioma multi-modal dataset

Metrics: C-index integrated Brier score normalized standard deviation local accuracy runtime global feature ranking

Baselines: SurvSHAP(t) SurvLIME

Builds on: SurvSHAP(t) SurvLIME Integrated Gradients SmoothGrad DeepHit

Closest papers by meaning

• SurvivalPFN: Amortizing Survival Prediction via In-Context Bayesian Inference (NeurIPS 2026 · similarity 0.86)
• Deep Copula-Based Survival Analysis for Dependent Censoring with Identifiability Guarantees (AAAI 2024 · similarity 0.85)
• Orthogonal Survival Learners for Estimating Heterogeneous Treatment Effects from Time-to-Event Data (NeurIPS 2025 · similarity 0.85)
• NeuralSurv: Deep Survival Analysis with Bayesian Uncertainty Quantification (NeurIPS 2025 · similarity 0.84)
• Toward a Well-Calibrated Discrimination via Survival Outcome-Aware Contrastive Learning (NeurIPS 2024 · similarity 0.84)
• A Scalable Nonparametric Continuous-Time Survival Model through Numerical Quadrature (NeurIPS 2026 · similarity 0.84)