Analysing repeated hospital readmissions using data mining techniques.

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Predictive Analytics

Background:

• Identifying patients at high risk for hospital readmission, particularly those with frequent emergency department (ED) visits, remains a challenge.
• Effective prediction models are needed to reduce readmission rates and improve patient outcomes.

Purpose of the Study:

• To explore 30-day readmission risk prediction for frequent ED patients.
• To compare the performance of five machine learning classification methods: Logistic Regression, Boosted Decision Trees (BDTs), Support Vector Machine (SVM), Bayes Point Machine (BPM), and Two-Class Neural Network (TCNN).

Main Methods:

• Utilized Microsoft Azure machine learning software for predictive modeling.
• Extracted data from 8455 penultimate visits of frequent ED patients from electronic health records.
• Compared classification algorithms based on their ability to predict 30-day readmission.

Main Results:

• Boosted Decision Trees (BDTs) demonstrated marginally superior performance (AUC) compared to Logistic Regression and Bayes Point Machine (BPM).
• Two-Class Neural Network (TCNN) and Support Vector Machine (SVM) showed comparatively lower predictive accuracy.
• Analysis revealed similarities and differences in significant predictors identified by BDT and Logistic Regression.

Conclusions:

• Machine learning models, particularly BDTs, show promise in predicting 30-day readmissions for frequent ED users.
• Further research incorporating time-varying covariates could enhance predictive accuracy and identify longitudinal factors for readmission risk reduction.