CATNet: Cross-event attention-based time-aware network for medical event prediction

Area of Science:

• Biomedical Informatics
• Artificial Intelligence in Healthcare
• Machine Learning for Clinical Prediction

Background:

• Electronic Health Record (EHR) data presents challenges due to its heterogeneous and irregular temporal nature.
• Existing medical event prediction (MEP) models often treat heterogeneous and temporal information separately, neglecting crucial correlations between different event types.
• Accurate prediction of future medical events is vital for proactive patient care and healthcare management.

Purpose of the Study:

• To develop a novel neural network, the Cross-event Attention-based Time-aware Network (CATNet), for improved medical event prediction.
• To address the limitations of existing models by unifying the modeling of heterogeneous and temporal EHR data.
• To enhance the understanding of correlations among different medical event types for more accurate predictions.

Main Methods:

• Proposed CATNet, a neural network incorporating an attention mechanism designed to be time-aware, event-aware, and task-adaptive.
• Developed a unified approach to model both heterogeneous medical information and temporal dynamics, considering local and global irregular temporal characteristics.
• Implemented cross-event attention to effectively capture and utilize correlations between various types of medical events.

Main Results:

• CATNet demonstrated superior performance compared to state-of-the-art methods across various medical event prediction tasks.
• Experiments conducted on the MIMIC-III and eICU public datasets validated the effectiveness of the proposed model.
• The model successfully integrated heterogeneous and temporal information, outperforming methods that handle these aspects separately.

Conclusions:

• CATNet offers a significant advancement in medical event prediction by effectively integrating heterogeneous and temporal EHR data.
• The cross-event attention mechanism is crucial for capturing inter-event correlations, leading to improved predictive accuracy.
• The proposed model provides a robust and adaptable framework for various clinical prediction tasks, with potential for real-world healthcare applications.