CleanSurvival uses reinforcement learning to auto-select preprocessing pipelines for survival analysis, improving predictive performance over simple baselines on real-world benchmarks

It introduces CleanSurvival, extending automated preprocessing to survival analysis, a task previously lacking tailored automated solutions. Existing AutoML pipelines have begun integrating data preprocessing for classification and regression, but this integration is lacking for time-to-event models with censored data; the work builds a dedicated framework for that gap. The abstract states the gap and the framework's positioning and provides an available Python package (GitHub link), but the loaded text contains no implementation details or code-level validation.

The framework uses Learn2Clean's Q-learning to search over combinations of imputation, outlier detection, and feature extraction techniques to improve the target time-to-event model. It transfers reinforcement-learning-driven preprocessing selection from general tasks to survival analysis, handles continuous and categorical variables, and supports Cox, random forest, neural network, or user-supplied models. The abstract describes the method's composition and supported model types; search-space size, reward design, and training details are not given in the loaded text.

On real-world dataset benchmarks, Q-learning-based preprocessing improved predictive performance relative to simple baselines. It provides empirical signal that automated preprocessing can outperform simple baselines in survival analysis, rather than only proposing a method. The abstract reports benchmark results on real-world datasets but gives no dataset names, sample sizes, or effect-size values.

A simulation study shows the approach is effective across different types and levels of missingness and noise. It extends evaluation from real data to controlled missingness and noise conditions, adding a robustness dimension to the evidence. The abstract states the simulation study's conclusion but provides no simulation setup, parameter ranges, or quantitative results.

The work targets researchers and practitioners doing time-to-event modeling with censored data, applicable where one needs to automatically select combinations of imputation, outlier detection, and feature extraction for Cox, random forest, neural network, or user-supplied models; the authors release it as an open-source Python package for direct trial in survival studies. Its conclusions rest on real-world dataset benchmarks and a simulation study, with runtime behavior being condition-dependent and most clearly interpretable in the best-covered benchmark cells.

The loaded text is incomplete, containing only the abstract, keywords, and references, with figures and tables absent, so specific datasets, benchmark cells, sample sizes, effect sizes, and concrete runtime behavior cannot be verified. The abstract says runtime behavior is condition-dependent but does not state how it varies under which conditions; the types and levels of missingness and noise covered by the simulation study are also not listed. In addition, implementation-level questions such as the Q-learning search-space size, reward design, and training cost need confirmation in the full text.