Variable Selection for Chronic Disease Outcome Prediction Using a Causal Inference Technique: A Preliminary Study

The ability to predict health outcomes of patients with chronic conditions has the potential for early risk factor identification, better treatment planning, and shared decision making. Compared to prediction tasks for acute conditions, modeling chronic diseases require careful adjustment for time-dependencies among treatments and responses, as well as variable selection to identify significant predictors. In this paper, targeting outcome prediction for chronic conditions which often require multiple medications, we applied causal inference techniques, specifically, the g-computation formula and marginal structural model, for the purpose of input variable selection prior to prediction using Bayesian networks. We propose that this approach allows for interpretable variable selection that also leads to better outcome prediction. An evaluation was performed using electronic health record data of a cohort of chronic kidney disease (CKD) patients to predict CKD progression. We identified effects of individual and concurrently used drugs on patients' kidney functions that are different across CKD stages. Lastly, using proposed variation selection technique, we predicted CKD progression with accuracy as high as 0.74, slightly outperforming logistic regression.