Designing a clinical decision support system to predict readmissions for patients admitted with all-cause conditions

Abstract

Readmission of patients is generally caused by inappropriate inpatient care, unexpected disease deterioration, inadequate health care plans for transition to community care, or premature discharge, and may be manifested by a degradation of healthcare quality and an increase of healthcare costs. In this study, we aimed to design a clinical decision support system (CDSS) based on patient data collected in Taiwan National Health Insurance Research Database in the prediction of 30-day all-cause readmissions for patients admitted with all-cause conditions. In order to select significant variables and adjust appropriate model parameters, we applied a method integrating genetic algorithm and support vector machine to design CDSSs based on 3 different cost-sensitive objective functions. The imbalanced data of patients admitted with all-cause conditions, including 6864 cases of admission and 48,374 cases of non-admission in 2011 as well as 6923 admissions and 49,521 non-admissions in 2012 were retrieved and used for training and testing the CDSSs, respectively. Each patient data consist of 31 candidate variables, including age, gender, number of comorbidities, Charlson comorbidity index (CCI), presence of comorbidities for calculating CCI, events within 1 year before admission, inpatient intervention, category of admitted hospital, length of admission, and discharge status. In the training phase, cluster-based under-sampling method (kNN) was used to prepare the balanced dataset for training and validating the CDSS. In contrast, the imbalanced testing dataset was applied to test the predictive performance of the trained CDSS model. The accuracy, sensitivity, specificity, and area under ROC curve (AUC) of the CDSSs designed with tenfold cross validation are 69.0–69.4%, 69.3.9–69.7%, 68.8–69.4%, and 0.74–0.76, respectively, in the training phase, and 62.3–64.3%, 68.0–69.5%, 61.3–63.8%, and 0.70–0.72, respectively, in the testing phase. Compared with the model in Maali et al. (BMC Med Inform Decis Making 18(1):1, 2018) (training phase: 0.74; testing phase: 0.71), all three models exhibit higher sensitivity and greater AUC. Among the 31 variables, 10, 18, and 17 salient variables, respectively, were selected for designing the models with three different objective functions. An online web service system designed based on Model 2 is provided for assisting physicians to detect potential patients having higher risk of all-cause readmissions after discharge and, if necessary, give them appropriate interventions to reduce their morbidities and mortalities as well as to reduce their healthcare costs.