Adjustment of survival curves with machine learning and cox regression: Application to a cardiac dataset.

Machine Learning (ML) is an effective approach for modeling large, complex data, including survival time-to-event data (SD), which are characterized by censoring. Survival curves are routinely generated and their adjustment for specific subgroups requires special care. Estimates of adjusted survival curves can be interpreted as the survival probabilities of groups with similar prognostic covariates. One popular method for analysis of SD is the Cox model, with Proportional Hazards (PH) or the more general stratified model. A standard approach for adjustment of survival curves using results from the Cox model is the means of covariates method, which is often the default in computer survival packages. Here we compared, on left-truncated and right-censored data, several SD-adapted methods (Cox adjusted corrected group prognosis method, Random Survival Forest, Survival Neural Network) with the Cox mean of covariates approach. In a graphic, survival estimates were calculated using a Monte Carlo simulation for pair-wise comparisons over given time periods. The model comparison was carried out on a mid-size real dataset not previously analyzed by these methods as well as on simulated survival datasets where the simulated ones mimicked the real datasets. [ABSTRACT FROM AUTHOR]