Predictors of Readmissions and Length of Stay for Diabetes Related Patients

Abstract

Unplanned readmissions and long stays are becoming a major concern of hospitals and healthcare providers as an indicator for quality of service. Predicting readmissions and long hospital stays in early stages allows extensive attention to patients identified with higher risks which leverages the community s healthcare and saves healthcare expenses and resources. This paper aims to investigate the predictive factors and develop a robust risk prediction framework, by combining feature engineering and machine learning algorithms. An actual data set has been used with various levels of routinely collected data that includes demographics, admission information, diagnosis, medications, tests, and service utilization information. The accuracy of our model using the readmission prediction achieved 94.8% with random forests. The support vector machine (SVM) gives the highest area under curve (AUC) statistic (i.e. 0.97) in readmission prediction, and it stands out as an efficient algorithm in predicting length of stay with prediction accuracy of 78.5%. Finally, analysis of features can be used to determine at-risk patients and target the delivery of early resource-intensive interventions.