Interpretable disease prediction using heterogeneous patient records with self-attentive fusion encoder

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Cardiovascular Disease Research

Background:

• Accurate prediction of cardiovascular disease (CVD) events is crucial for public health.
• Existing models often struggle to effectively integrate diverse patient data types.
• There is a need for interpretable models that can identify key factors in disease prediction.

Purpose of the Study:

• To develop and evaluate an interpretable disease prediction model using a self-attentive fusion encoder.
• To assess the model's performance in predicting cardiovascular disease events.
• To compare the proposed model against state-of-the-art methods.

Main Methods:

• Utilized a large dataset from South Korea, including medical codes and patient characteristics.
• Developed a self-attentive fusion encoder to combine sequential medical codes and patient data.
• Compared the model's prediction performance against recurrent neural network-based and other machine learning approaches.

Main Results:

• The proposed model achieved a superior area under the curve (0.839) compared to existing methods (0.741-0.830).
• Demonstrated consistent outperformance in predicting events from nonobvious and diverse patient factors.
• Attention weights provided interpretability, highlighting informative parts of patient history.

Conclusions:

• The developed interpretable disease prediction model effectively fuses heterogeneous patient records.
• The self-attentive fusion encoder offers superior performance in cardiovascular disease event prediction.
• The model's interpretability enhances understanding of patient history's role in disease risk.