Predicting heart failure in-hospital mortality by integrating longitudinal and category data in electronic health...

Area of Science:

• Cardiology
• Medical Informatics
• Artificial Intelligence

Background:

• Heart failure is a critical condition affecting millions globally each year.
• Accurate prediction of in-hospital mortality is crucial for timely intervention.
• Electronic health records (EHRs) contain valuable data for predictive modeling.

Purpose of the Study:

• To develop and evaluate a deep fusion learning model (DFL-IMP) for predicting in-hospital mortality in heart failure patients.
• To leverage both time series and categorical data from EHRs for enhanced prediction accuracy.
• To assess the model's performance across different temporal observation and prediction windows.

Main Methods:

• A deep fusion learning model (DFL-IMP) was designed, integrating 41 time series features and 17 category features from EHRs.
• Data from 7696 heart failure patients were analyzed in an observational study.
• The model's predictive performance was evaluated using the Area Under the Curve (AUC) metric across various time windows.

Main Results:

• The DFL-IMP model achieved a high AUC of 0.914 with a 5-day observation window and a 30-day prediction window.
• This performance significantly outperformed established baseline models like Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), LSTM, GRU, GRU-U, and MTSSP.
• The model demonstrated robust predictive capabilities using readily available EHR data.

Conclusions:

• The DFL-IMP model offers a powerful tool for predicting in-hospital mortality in heart failure patients.
• Early prediction enables timely clinical interventions, potentially improving patient outcomes and life expectancy.
• This approach highlights the potential of deep learning in analyzing complex EHR data for critical care.