A Naive Bayes machine learning approach to risk prediction using censored, time-to-event data

摘要

Predicting an individual's risk of experiencing a future clinical outcome isa statistical task with important consequences for both practicing cliniciansand public health experts. Modern observational databases such as electronichealth records (EHRs) provide an alternative to the longitudinal cohort studiestraditionally used to construct risk models, bringing with them bothopportunities and challenges. Large sample sizes and detailed covariatehistories enable the use of sophisticated machine learning techniques touncover complex associations and interactions, but observational databases areoften ``messy,'' with high levels of missing data and incomplete patientfollow-up. In this paper, we propose an adaptation of the well-known NaiveBayes (NB) machine learning approach for classification to time-to-eventoutcomes subject to censoring. We compare the predictive performance of ourmethod to the Cox proportional hazards model which is commonly used for riskprediction in healthcare populations, and illustrate its application toprediction of cardiovascular risk using an EHR dataset from a large Midwestintegrated healthcare system.