SurvivalBench: Empirical Benchmark for Survival Models

• SurvivalBench is a standardized framework that offers reproducible pipelines, curated datasets, and explicit preprocessing protocols for survival analysis.
• It addresses inconsistencies in EHR preprocessing, incomplete datasets in finance and omics, and fragmented evaluation protocols to ensure model comparability.
• The framework features configurable, transparent pipelines with rigorous missingness control and multi-modal support, advancing reproducible research in survival modeling.

SurvivalBench refers to a series of large-scale, empirically rigorous benchmark frameworks, datasets, and open-source pipelines developed for standardized evaluation of survival analysis models. The term encompasses both end-to-end data pipelines (as in the clinical multi-modal SurvBench pipeline (Mesinovic et al., 14 Nov 2025)) and curated evaluation suites for survival modeling tasks in biomedicine, finance, and computational statistics. SurvivalBench addresses the reproducibility, comparability, and methodological challenges historically associated with survival analysis across diverse domains.

1. Motivations and Historical Context

Survival analysis, encompassing time-to-event prediction under censoring, is foundational in fields such as clinical prognosis, risk scoring, and reliability engineering. Progress in deep learning and statistical modeling for survival data has been impeded by the lack of standardized, transparent, and reproducible preprocessing and benchmarking protocols. SurvivalBench-type frameworks were developed to address several critical constraints:

• Inconsistent preprocessing in electronic health record (EHR) applications, leading to data leakage, irreproducibility, and modality fragmentation.
• Absence of large-scale, realistic open datasets outside of medicine, particularly for domains like decentralized finance where public, high-frequency transaction data allows for new classes of survival tasks.
• Fragmentation in evaluation protocols for classical and modern survival models, including inadequate calibration assessment and improper handling of censoring, especially for high-dimensional or multi-task settings (Herrmann et al., 2020).

2. Framework Architecture and Principles

The core design objectives of SurvivalBench frameworks are:

• Raw-to-tensor transparency: All transformations are performed directly from raw tabular files (e.g., flat CSVs from PhysioNet or DeFi ledgers), with explicit aggregation, filtering, scaling, masking, and imputation steps—no hidden intermediates or non-reproducible preprocessing.
• Multi-domain and multi-modal support: Support for diverse data types, including time-series (physiological signals, transaction histories), static demographics, textual embeddings (e.g., radiology BERT), and structured event codes (e.g., ICD diagnoses).
• Survival-specific label engineering: Standardized handling of right-censoring, horizon truncation, time discretization for discrete-time models, and encoding of competing risks.
• Explicit missingness and leakage control: Each tensor output is accompanied by missingness masks, and splitting is always at the patient/entity level to prevent label leakage, with stratified train/validation/test splits.
• Configuration-driven reproducibility: All preprocessing parameters, including windows, missing thresholds, feature selection, and included modalities, are specified in version-controlled YAML or equivalent config files, enabling exact reruns and systematic ablations.

3. Reference Implementations: Clinical and Financial Pipelines

3.1 SurvBench for EHR Data (Mesinovic et al., 14 Nov 2025)

SurvBench provides a fully transparent, multi-modal preprocessing pipeline for survival analysis on critical care datasets (MIMIC-IV, eICU, MC-MED):

Dataset	Static Features	Dynamic Features	Codes/Embeddings
MIMIC-IV	Demographics, admission details	High-freq vitals, labs (binned)	ICD-10 (multi-hot), rad-BERT
eICU	Demographics	Periodic/aperoidic vitals, labs	None by default
MC-MED	Demographics, chief complaint	Minute-level numerics, labs	ICD history, rad-BERT

Each data instance produces tensors $(N, W, F)$ (patients, time-windows, features), with associated binary missingness arrays and split-specific files. Survival labels are constructed as (duration, event indicator) pairs for both single-risk and competing-risk settings.

3.2 FinSurvival for DeFi (Green et al., 7 Jul 2025)

FinSurvival algorithmically generates 16 right-censored survival prediction tasks from over 7.6 million DeFi transactions, using index-outcome pairs (e.g., borrow→liquidation) with precise event/censoring construction. Feature sets are broad (base fields, user/market history, time features), and two task types are provided: survival regression (with censoring) and classification (via RMST thresholding).

3.3 Multi-omics SurvivalBench (Herrmann et al., 2020)

This benchmark aggregates 18 TCGA cancer datasets, with up to 100,000 features spanning CNVs, transcriptomics, miRNA, and somatic mutations, plus clinical variables. It enables comparison of penalized regression (lasso, group lasso, IPF-lasso, etc.), boosting (CoxBoost, glmboost), and random forest/block-wise methods.

4. Evaluation Metrics and Scoring Protocols

SurvivalBench frameworks adhere to rigorous, multi-metric evaluation, supporting:

• Discrimination: Harrell’s C-index, Uno’s C-index, concordance statistics for risk ranking.
• Calibration: 1-Calibration (Hosmer–Lemeshow at fixed times), D-Calibration (global uniformity of survival probability estimates), and van Houwelingen’s $\alpha$.
• Scoring rules: Integrated Brier Score (IBS), right-censored log-loss (RCLL), and survival log-likelihoods.
• Point estimation losses: Margin L1 (with mean-residual imputation for censored cases), log-L1 for relative timing error.
• Auxiliary metrics: Area under ROC for (RMST-based) classification, consistency, refusal rates, and chain-of-thought reflectiveness for agentic LLM settings (Lu et al., 5 Mar 2026).

All evaluation is conducted with nested cross-validation or resampling, with strict control over hyperparameter tuning and multiple random splits.

5. Supported Model Classes

Benchmarks support wide-ranging models for survival prediction, individual survival distribution estimation, and causal treatment effect inference:

• Classical: Cox proportional hazards, parametric AFT (Weibull, log-normal), flexible spline, penalized Cox variants (lasso, elastic net).
• Machine learning: Random survival forests (classical, oblique, conditional inference), survival boosting (model-based, CoxBoost), XGBoost (Cox/AFT objectives), support vector survival machines.
• Neural survival models: DeepSurv, DeepHit, MTLR (multi-task logistic regression), discrete-time PyCox.
• Causal effect estimators: Causal Survival Forests, SurvITE, meta-learners (S/T/X/DR), Double-ML, outcome imputation approaches (Noroozizadeh et al., 5 Mar 2026).
• Group-aware methods: BlockForest, IPF-Lasso, priority-lasso favoring clinical, GRridge.

6. Reproducibility, Extensibility, and Deployment

Every SurvivalBench release provides:

• Configurable pipelines: All settings are exposed for modification via YAML or code-level configuration, with modular loader designs for new datasets/modalities.
• Plugin support: Users can implement new data loaders, feature extractors, imputation strategies, or model classes and benchmark them under identical protocols.
• Integration with model libraries: Outputs are compatible with PyCox, scikit-survival, survival meta-learners, and deep-learning frameworks.
• Reproducible environments: Shipping with full dependency pinning (e.g., requirements.txt, Dockerfiles) and explicit scripts for end-to-end runs.

7. Recommended Practices and Limitations

Best-practice SurvivalBench use requires:

• Reporting both discrimination and calibration metrics for all survival models.
• Using ISD models (e.g., MTLR, RSF-KM, penalized Cox) when patient-specific full survival distributions are needed rather than only rank orderings.
• Careful preprocessing: Impute or drop missing features, control for leakage, and filter low-variance predictors.
• Critical awareness of model limitations: In low-dimensional settings, Cox PH often suffices and outperforms more complex methods (Burk et al., 2024). For multi-omics or high-dimensional regimes, group structure-aware methods protect against omission of low-dimensional prognostic blocks. In high-censoring or strongly nonlinear domains (e.g., financial DeFi), specialized architectures or further feature engineering may be required.
• Extending beyond classical settings: SurvivalBench frameworks are being actively extended to support competing risks, longitudinal covariates, heterogeneous treatment effect estimation, and robust evaluation under causal assumption violations (Noroozizadeh et al., 5 Mar 2026).

SurvivalBench-type resources have established a new standard for empirical rigor, comparability, and transparency in survival analysis, underpinning methodological innovation in EHR outcomes research, financial risk modeling, and omics-based predictive oncology.