Effective hospital readmission prediction models using machine-learned features.

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Predictive Analytics

Background:

• Hospital readmissions pose a significant financial burden on healthcare systems.
• Conventional readmission prediction models often lack accuracy and rely on labor-intensive manual features.
• There is a need for more effective models utilizing automatically generated features from longitudinal data.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting 30-day hospital readmissions.
• To compare the predictive performance of models using automatically generated features against traditional manually-engineered features.
• To identify high-risk patients for potential readmission prevention interventions.

Main Methods:

• Utilized population-level linked administrative hospital data from Alberta, Canada (2011-2017).
• Trained prediction models on data from 428,669 patients discharged in 2015-2016, excluding maternity and psychiatric admissions.
• Employed machine learning techniques to derive features automatically from observational data, alongside manually-derived features.

Main Results:

• The machine learning model achieved an Area Under the Curve (AUC) of 0.83 ± 0.0045, outperforming the LACE model (AUC 0.66 ± 0.0064).
• Identified key predictors of readmission including higher hospital utilization, more physician visits, increased prescriptions, chronic conditions, and age (≥65).
• The study included 428,669 patients, with a 30-day readmission rate of 5.83%.

Conclusions:

• Machine learning models integrating automatically generated and manual features enhance prediction accuracy for hospital readmissions.
• The developed model effectively identifies patients at high risk of readmission.
• This improved prediction capability can facilitate targeted interventions to reduce costly readmissions.