Machine Learning and Causal Approaches to Predict Readmissions and Its Economic Consequences Among Canadian Patients With Heart Disease: Retrospective Study (Preprint)

Abstract

UNSTRUCTURED Patient readmission is preventable and predictable. However, it poses a major problem in healthcare worldwide. Successful prediction of readmission based on a patient’s past conditions, physical traits and current diagnosis would be a productive step forward towards reducing the stress readmission has on the healthcare system as well as the financial stress associated with it. The goal of this study is to design a predictive model that describes patient readmission in an optimally cost-effective manner that minimizes false negatives, and compare it to the cost of prolonging a patient’s length of stay. An ensemble prediction model was built using 5 classification submodels aiming to predict whether a patient is likely to be readmitted within 30 days. The model’s training on the Canadian Institute for Health Information’s Discharge Abstract Database yielded a precision score of of 74% and an f1-score of 46%. The ensemble prediction model resulted in being more effective than previous submodels because it minimized both variance and bias. This suggests the model is a viable candidate. To further analyze the resource trade-offs of prolonged stay, the expected length of Stay (ELOS) and resource intensity weight value (RIW) columns of the dataset were graphed then clustered using k-means. Based on the final model, the construction proved comparable to previous methods implemented in the literature, making it an overall efficient and robust model. Hence, quality data collection and construction of an ensemble readmission collection model should be a major area of focus and resource allocation for healthcare institutes worldwide.