A random forest machine learning model with time-varying covariates improved sudden cardiac death prediction, achieving an AUC of 0.89 at 6 years compared to 0.77-0.83 for Poisson regression models.
Cohort (n=15,661)
Sí
Does a machine learning model incorporating time-varying predictors improve the prediction of sudden cardiac death compared to standard Poisson regression models in a general population cohort?
Applying a novel machine learning approach with time-varying predictors significantly improves the prediction of sudden cardiac death in the general population compared to traditional regression models.
Tasa de eventos absoluta: 0.89% vs 0.83%
Abstract Introduction Sudden cardiac death (SCD) is a devastating consequence often without antecedent expectation. Current risk stratification methods derived from baseline independently modeled risk factors are insufficient. Novel random forest machine learning (ML) approach incorporating time-dependent variables and complex interactions may improve SCD risk prediction. Methods Atherosclerosis Risk in Communities (ARIC) study participants were followed for adjudicated SCD. ML models were compared to standard Poisson regression models for interval data, an approximation to Cox regression, with stepwise variable selection. Eighty-two time-varying variables (demographics, lifestyle factors, clinical characteristics, biomarkers, etc.) collected at four visits over 12 years (1987-98) were used as candidate predictors. Predictive accuracy was assessed by area under the receiver operating characteristic curve (AUC) through out-of-bag prediction for ML models and 5-fold cross validation for the Poisson regression models. Results Over a median follow-up time of 23.5 years, 583 SCD events occurred among 15,661 ARIC participants (mean age 54 years and 55% women). Compared to different Poisson regression models (AUC at 6-year ranges from 0.77-0.83), the ML model improved prediction (AUC at 6-year 0.89). Top predictors identified by ML model included prior coronary heart disease, which explained 47.9% of the total phenotypic variance, diabetes mellitus, hypertension, and T wave abnormality in any of leads I, aVL, or V6. Using the top ML predictors to select variables, the Poisson regression model AUC at 6-year was 0.77 suggesting that the non-linear dependencies and interactions captured by ML, are the main reasons for its improved prediction performance. Conclusions Applying novel ML approach with time-varying predictors improves the prediction of SCD. Interactions of dynamic clinical characteristics are important for risk-stratifying SCD in the general population.
Yu et al. (Sun,) conducted a cohort in Sudden Cardiac Death (n=15,661). Random Forest for Survival, Longitudinal, and Multivariate data analysis (RF-SLAM) vs. Poisson regression models was evaluated on Area under the receiver operating characteristic curve (AUC) for sudden cardiac death prediction at 6 years. A random forest machine learning model with time-varying covariates improved sudden cardiac death prediction, achieving an AUC of 0.89 at 6 years compared to 0.77-0.83 for Poisson regression models.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: