Machine Learning For Risk Prediction After Heart Failure Emergency Department Visit or Hospital Admission Using...

Area of Science:

• Cardiology
• Data Science
• Health Informatics

Background:

• Patients with heart failure (HF) face high risks of adverse events after emergency department (ED) visits or hospitalizations.
• Accurate risk stratification for these patients remains a significant clinical challenge.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for identifying patients at high risk of adverse outcomes following ED visits or HF hospitalizations.
• To compare the performance of an ML approach against traditional logistic regression for HF risk prediction.

Main Methods:

• Utilized a large regional administrative healthcare database from Alberta, Canada (2002-2016).
• Applied deep feature synthesis for feature extraction from diverse health data sources.
• Trained and evaluated a gradient boosting algorithm (CatBoost) against logistic regression.
• Assessed model performance using the area under the receiver operating characteristic curve (AUC-ROC).

Main Results:

• The CatBoost model demonstrated superior predictive performance compared to logistic regression for combined endpoints (HF ED visit, HF readmission, or death) at both 30-day and 1-year follow-up.
• At 30 days, CatBoost achieved an AUC-ROC of 74.16% versus 62.25% for logistic regression.
• At 1 year, CatBoost achieved an AUC-ROC of 76.80% versus 69.52% for logistic regression. For all-cause mortality alone, CatBoost achieved AUC-ROC values of 83.21% (30-day) and 85.73% (1-year).

Conclusions:

• Machine learning-based modeling, enhanced by deep feature synthesis, offers superior risk stratification for heart failure patients.
• This approach effectively identifies patients at risk of adverse outcomes after ED visits or hospitalizations.
• The findings highlight the potential of ML in improving care for heart failure patients within large healthcare systems.