DynSHAP: Towards Explainable Dynamic Survival Analysis
Nastasya Anokhina, Jonas Jürß, Pietro Liò
cs.LG · cs.AI
Quick verdict
DynSHAP extends SHAP to dynamic survival analysis via time–feature players; promising and honest, but gains are model-specific and faithfulness mixed.
Summary
The paper introduces DynSHAP, a SHAP framework for explainability of dynamic survival analysis (DSA) models such as Dynamic-DeepHit and DySurv, which consume irregular longitudinal data and output survival functions. DynSHAP discretizes the time domain into bins, treats time–feature pairs as Shapley players, and extends marginal estimators (permutation sampling and kernel regression); Temporal DynSHAP additionally fits linear-Gaussian autoregressive conditionals across bins and uses conditional sampling for out-of-coalition values. On a synthetic Weibull survival dataset with ground-truth conditional SHAP values, Temporal DynSHAP roughly halves RMSE versus marginal estimators on DySurv (0.0032 vs 0.0064) but fails on Dynamic-DeepHit (0.0273, near the random-noise control 0.0265). Perturbation-based faithfulness on the PBC and MS clinical datasets is mixed: both estimators beat landmarking and random removal on PBC, but on MS/DySurv the random baseline (0.0097) slightly exceeds Kernel (0.0094) and Temporal (0.0079). Temporal DynSHAP is faster yet less stable (median log RIS about 3× higher on synthetic and PBC) and yields spurious attributions on independent features (RMSE 0.0498 in the linear-model check). The ground truth used for evaluation is itself conditional, favoring the proposed estimator. Code is public; the MS dataset is simulated from anonymized data. An honest workshop contribution whose central advantage is model- and metric-dependent.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Existing SHAP extensions cannot jointly handle irregular longitudinal inputs and functional survival outputs, and DynSHAP fills this gap by treating time–feature pairs as Shapley players.
- Temporal DynSHAP recovers ground-truth attributions of temporally dependent features about 2× more accurately than marginal estimators on DySurv, though not on Dynamic-DeepHit.
- On real clinical datasets, marginal estimators systematically collapse attribution onto the patient's most recent visit while Temporal DynSHAP recovers historical attribution, at the cost of reduced stability under input perturbation.
- DynSHAP attributions on PBC align with established clinical findings, matching the Mayo model's key biomarkers (albumin, bilirubin, edema, ascites, prothrombin).
Strengths
Clear formalization of SHAP properties for DSA
Appendix A redefines local accuracy, missingness, and consistency for time–feature players, and Figure 6 empirically verifies local accuracy residuals, exposing landmarking's violation rather than merely asserting it.
Well-targeted synthetic benchmark with shortcut detection
Appendix D's four covariates each isolate a failure mode (AR(1) biomarker β=2.5, treatment onset β=−2.5, static risk factor, zero-effect noise to catch shortcut learning), with Weibull inverse-cumulative-hazard event times and 30% administrative censoring — a rigorous ground-truth setup.
Above-average transparency about negative results
Clinical plausibility anchored to an external baseline
Usable artifact and reproducibility support
Weaknesses
Central advantage is architecture-dependent
Ground-truth circularity favors the proposed estimator
Real-data faithfulness evidence is inconsistent
Temporal estimator's pathologies are material
Weakened real-world evidence base
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
DynSHAP (arXiv 2609.13042, submitted to the RCMLR workshop at NeurIPS 2026) addresses a real gap: no existing XAI method jointly handles irregular longitudinal inputs and functional survival outputs (Section 2). The framework discretizes follow-up time into K bins, promotes each time–feature pair to a Shapley player (P = K×M), and estimates per-horizon attributions with three estimators: permutation Sampling DynSHAP, Kernel DynSHAP (Eqs. 3–5), and Temporal DynSHAP, which fits per-bin ℓ2-regularized linear-Gaussian autoregressive conditionals and imputes out-of-coalition pairs by conditional sampling (Eqs. 6–7). Appendix A usefully redefines local accuracy, missingness, and consistency for this setting. Central claims: (1) the gap is real; (2) Temporal DynSHAP recovers temporally dependent ground truth better than marginal estimators; (3) real-data attributions are faithful and clinically plausible.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is DynSHAP?
DynSHAP is a SHAP-based explainability framework for dynamic survival analysis models. It discretizes a patient's follow-up time into bins and treats each time–feature pair as a 'player' in the Shapley value game, producing attributions that say which feature mattered and when, while jointly handling irregular longitudinal inputs and survival-function outputs. A 'Temporal' variant learns linear dependencies between features across time and imputes out-of-coalition values by conditional sampling instead of independent background draws.
How does DynSHAP differ from SurvSHAP(t), TimeSHAP, or WindowSHAP?
SurvSHAP(t) handles survival models' functional outputs but assumes static instances; TimeSHAP and WindowSHAP handle sequences but target standard classifiers rather than survival functions, and all rely on marginal sampling that can create unrealistic counterfactuals with correlated time series. DynSHAP is designed to jointly cover irregular longitudinal inputs and survival outputs, and its Temporal variant adds conditional sampling to respect temporal dependence.
Does Temporal DynSHAP always beat standard (marginal) SHAP?
No. On the synthetic ground-truth benchmark it recovers conditional attributions about twice more accurately on the DySurv model (RMSE 0.0032 vs 0.0064), but on Dynamic-DeepHit it performs no better than a random-noise control (0.0273 vs 0.0265). It is also less stable under input perturbations and can introduce spurious attributions when features are actually independent.
Is the code and data available?
The authors provide a public Python package at github.com/tasyaa04/dynShap. The PBC dataset is a public Mayo Clinic liver-cirrhosis dataset; the MS dataset was simulated by sampling from anonymized patient records, and the paper does not state that this simulated dataset or its generator will be released.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are extracted as text (axis labels and captions only), so curve shapes in Figures 4 and 10–13 could not be inspected; qualitative curve judgments rely on the summary numbers reported in captions and tables.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.