Performance of Bayesian Additive Regression Trees (BART)-Survival implementation in Python: A Comparison with Traditional and R-Based BART Survival Analysis Methods

Abstract Background: Time-to-event analysis, widely used in public health to model outcomes such as disease progression and survival, has traditionally relied on maximum likelihood-based approaches like Cox proportional hazards models. Recent advances in non-parametric machine learning, notably Bayesian Additive Regression Trees (BART), have enhanced predictive performance in complex, high-dimensional data common to public health studies. BART has primarily been implemented for survival analysis in R, offering a flexible alternative to conventional models. With Python's growing adoption for large healthcare data analytics, this study evaluates a novel Python-based BART-Survival algorithm (p-BART), benchmarking its performance against an R-based BART implementation (r-BART) and traditional survival models using both simulated and real-world healthcare data. Methods The p-BART algorithm was evaluated using simulated datasets with varying complexity, including non-linear effects, interactions, and non-proportional hazards. Its performance was also assessed using real-world administrative healthcare data (Premier Healthcare Database). Key evaluation metrics included root mean squared error, bias, and coverage probability. Comparisons were made against Kaplan-Meier (KM), Cox proportional hazards model (CPHM), and r-BART implementation. Results In simpler scenarios, such as one or two sampling distributions and multivariate regression with proportional hazards, p-BART performed similarly to traditional statistical models and to r-BART. In more complex scenarios, including non-proportional hazards and models with nonlinear interactions, p-BART outperformed traditional methods while matching r-BART performance. In real-world data analysis, p-BART produced hazard ratio and survival probability estimates comparable to the CPHM and r-BART. Like other approaches, p-BART showed the greatest coverage loss and reduced reliability under extreme censoring scenarios. Conclusions This study demonstrated the accuracy and robustness of p-BART, even when the proportional hazards assumption was violated. It performed well in complex scenarios and was comparable to r-BART, making it valuable for public health researchers using Python. Like other methods, the performance of p-BART declines under high censoring. As Python and machine learning applications in public health expand, p-BART provides an advanced statistical method for analyzing large healthcare datasets, improving decision-making in disease in clinical practice and public health. Clinical Trial Number: Not applicable.