Variable Selection by Domain Experts vs. Filter Algorithms for Clinical Predictive Modeling

Abstract

Studies on variable selection in the medical field have focused largely on algorithms with little attention paid to domain experts in this regard. This chapter compared the performance of domain experts with filter algorithms in variable selection for clinical predictive modeling. Five clinical datasets on bacterial survival, neonatal birthweight, breast cancer, diabetes, and myocardial infarction were employed. For each dataset, fifteen domain experts were requested to rank the importance of the variables on a five-point Likert scale. The same variables were ranked using four algorithms, namely, chi-squared, Fisher score, Pearson's correlation, and varImp function. Results of classification models showed that both methods performed competitively. This means human expertise and experience are important in clinical predictive modeling and must not be mortgaged to algorithms. Further studies should focus on developing automated platforms that codify domain knowledge and experience to facilitate real-time, speedy, and seamless variable selection.