Document Type : Systematic Review
Author
Shiraz University of Medical Sciences, Shiraz, Iran.
10.22034/hp.2026.608530.1111
Abstract
Background: Missing covariates and right-censored event times pose distinct challenges in medical survival analysis. Machine-learning and deep-learning approaches offer flexibility but differ in assumptions and performance. We aimed to review machine-learning and deep-learning methods for handling missing covariates and, separately, censored event times in medical time-to-event analyses, evaluating their roles, assumptions, and reported performance.
Methods: In this systematic review, PubMed, Embase, Scopus, and Web of Science Core Collection were searched from inception through 18 September 2026. Eligible studies evaluated ML/deep-learning methods managing incomplete data in medical survival analyses. Two reviewers screened studies and extracted methodological details, missingness mechanisms, censoring frameworks, and performance data. Appraisal used PROBAST domains for applied prediction studies and a domain-based framework for methodological/simulation studies. Heterogeneity necessitated structured narrative synthesis.
Results: Twenty-five studies (2009–2025) primarily addressed missing covariates with right-censored outcomes; one imputed censored event times. Methods included tree-based models, neural networks, autoencoders, and multiple-imputation workflows. El Badisy et al. (2024) found that CART/MissForest favoured selected imputation-accuracy outcomes, whereas miceCART/miceRF reduced Cox-coefficient bias and improved confidence-interval coverage. Cai et al. (2025) used CART-based multiple imputation before survival-model comparisons; Ogutu et al. (2025) used MissForest before deep-learning survival modelling. Wang et al. (2022) evaluated CondiS/CondiS-X for censored event-time imputation. No universal survival-performance benefit emerged: findings depended on missingness, incomplete-data characteristics, methods, models, validation, comparators, and metrics. Discrimination was reported more frequently than calibration; external validation and explicit leakage safeguards were uncommon.
Conclusion: ML-based methods may address incomplete covariates without consistent superiority across different performance outcomes. Censored event-time imputation requires separate consideration. Studies should clarify missingness and censoring assumptions, prevent leakage, and evaluate calibration and external validity.
Keywords
