Addressing Cancer Readmission Prediction Model Drift: A Case Study

ABSTRACT

The cancer readmission prediction model classifies patients as high-risk or low-risk for readmittance. Consequently, intervention strategies focus on high-risk patients. Nevertheless, the performance of machine learning models generally degrades over time due to changes in the environment that violates models' assumptions, which include statistical data changes and process changes. This research introduces a framework that improves the sensitivity of the cancer readmission prediction model by identifying new features of cancer readmission, such as Diabetes and Anti-Nausea, which potentially cause the model's sensitivity to drift. The proposed model considers these 20 new factors with the 35 original factors that use the most recent dataset to predict cancer readmissions. Recursive feature elimination was used to identify key features. Some of the most popular classification algorithms, which include logistic regression and adaptive boosting, were used to retrain and classify cancer readmissions. The best algorithm was validated on a new dataset that was collected over 11 months, which covered three different waves of Covid-19. The results suggested K-Nearest Neighbors (KNN) algorithm performs the best among all eight studied algorithms. The KNN model incorporated new dominant features that did not exist in the original Random Forest (RF) model. The KNN model has an improvement of 8.05% in sensitivity compared to the RF model. The presence of Covid-19 does not have a significant impact on the performance of the KNN model. The suggested framework identifies potential admitted patients for intervention actions, helps reduce cancer readmission rates, costs, and improves the quality of care for cancer patients. © 2022 IISE Annual Conference and Expo 2022. All rights reserved.