Predicting 30-Day Readmissions in Patients With Heart Failure Using Administrative Data: A Machine Learning Approach

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Predictive Analytics

Background:

• Heart failure readmissions pose a significant burden on healthcare systems.
• Accurate prediction of readmission risk is crucial for effective patient management.
• Existing risk scores, like the LaCE score, may have limitations in predicting heart failure readmissions.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting heart failure readmission risk using administrative data.
• To compare the performance of ML models against the established LaCE score.
• To assess the clinical utility of ML-based predictions for unplanned hospital readmissions.

Main Methods:

• A prognostic study involving 9845 heart failure patients in Alberta, Canada (2012-2019).
• Development of ML models, including XGBoost and neural networks, using 80% of administrative data.
• Independent validation of ML models and a modified LaCE score using the remaining 20% of data.
• Performance evaluation using c-statistics, likelihood ratios, and positive predictive values.

Main Results:

• The XGBoost model achieved a higher c-statistic (0.65) compared to neural networks (0.58) and the modified LaCE score (0.57).
• The XGBoost model demonstrated a positive predictive value range of 21%-62%, while the LaCE score ranged from 21%-24%.
• Despite outperforming the LaCE score, the XGBoost model showed only a moderate shift in probability, indicating limited predictive power.

Conclusions:

• Machine learning models trained on administrative data show improved, yet moderate, prediction of heart failure readmissions compared to the LaCE score.
• Current ML models using administrative data alone may not be sufficiently informative for clinical deployment.
• Integrating additional data sources, such as clinical notes via natural language processing, could enhance ML model performance for readmission prediction.