A mixed-ensemble model for hospital readmission.

Area of Science:

• Healthcare Informatics
• Machine Learning in Medicine
• Predictive Analytics

Background:

• Hospital readmissions, particularly within 30 days, pose a significant challenge in healthcare.
• Predicting readmissions is difficult due to the imbalanced nature of the data, where most patients are not readmitted.

Purpose of the Study:

• To develop a hospital readmission predictive model for Congestive Heart Failure (CHF) patients.
• To create a model that balances reasoning transparency with predictive accuracy.
• To improve the prediction of positive readmission instances.

Main Methods:

• Developed a mixed-ensemble model combining a boosted C5.0 tree with a Support Vector Machine (SVM).
• Validated the model using administrative records of 20,321 inpatient admissions for 4840 CHF patients.
• Utilized data from Veterans Health Administration (VHA) hospitals between fiscal years 2006-2014.

Main Results:

• The ensemble model achieved a total accuracy ranging from 81% to 85%.
• SVM predictions demonstrated higher sensitivity (true positive rates) compared to C5.0 predictions across various ROC curve cut-off values.
• Identified diverse roles of predictors like comorbidities, lab values, and vitals in both models.

Conclusions:

• The mixed-ensemble model enhances exploratory knowledge discovery and controls classification errors for positive readmission instances.
• This ensembling method overcomes limitations of individual classifiers and traditional methods for predicting all-cause CHF readmissions.
• The approach improves classification accuracy for positive readmissions, especially when predictive factors are limited.