Comparison of oblique random survival forest, random survival forest, and statistical models for time-to-event data...

Area of Science:

• Biostatistics
• Machine Learning in Healthcare
• Prognostic Modeling

Background:

• Time-to-event (TTE) machine learning (ML) algorithms are increasingly used in prognostic models.
• A systematic evaluation of their strengths and limitations compared to traditional statistical models (SMs) is lacking.

Purpose of the Study:

• To compare the predictive performance and computational time of TTE ML algorithms (Oblique Random Survival Forest - ORSF, Random Survival Forest - RSF) against SMs (Cox Proportional Hazards - Cox PH, Penalized Cox PH).
• To evaluate algorithm performance across various scenarios with differing censoring rates, sample sizes, and predictor effects.

Main Methods:

• Generated 18 scenarios assuming the proportional hazards (PH) assumption.
• Evaluated performance using Harrell's C-index and Integrated Brier Score (IBS).
• Assessed performance differences using One-Way Repeated Measures ANOVA.

Main Results:

• In linear scenarios, SMs outperformed RSF; ORSF variants showed comparable performance to SMs.
• Under non-linear scenarios, SMs consistently achieved higher C-indices than RSF, with minimal differences from ORSF.
• RSF demonstrated inferior discrimination compared to SMs and ORSF; ORSF variants showed no significant differences in discrimination or calibration.

Conclusions:

• Traditional SMs outperform ML models in TTE prediction at higher censoring rates.
• SMs matched ORSF performance at lower censoring rates, while RSF was generally inferior.
• ORSF-net exhibited the longest training time among ML models evaluated.