Interpretable time-aware and co-occurrence-aware network for medical prediction.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Computational Biology

Background:

• Electronic Health Records (EHRs) are crucial for personalized healthcare but present challenges due to their hierarchical structure and the need for interpretable prediction models.
• Existing methods often fail to fully capture the complexities of EHR data, such as patient admissions and co-occurring diagnoses, limiting their interpretability for non-specialists.

Purpose of the Study:

• To develop an interpretable deep learning network, TCoN (Time-aware and Co-occurrence-aware Network), specifically designed to handle the hierarchical structure of EHR data.
• To enhance disease prediction accuracy and provide clear interpretations of the prediction results for healthcare professionals.

Main Methods:

• The proposed TCoN model utilizes a co-occurrence-aware self-attention (CS-attention) mechanism and a time-aware gated recurrent unit (T-GRU) to model multilevel relationships within EHR data.
• An interpretation path and a diagnosis graph are integrated into the model to ensure the results are understandable and actionable.

Main Results:

• TCoN was evaluated on a real-world dataset for various prediction tasks, including mortality, readmission, and disease prediction, achieving 2.01% higher accuracy than baseline methods.
• The model successfully provided interpretable causal relationships and generated patient-specific diagnosis graphs, demonstrating its ability to explain predictions.

Conclusions:

• TCoN is a novel, interpretable, and effective deep learning method capable of modeling hierarchical medical data and predicting medical events, outperforming existing state-of-the-art approaches.
• Future research directions include leveraging graph embedding techniques with additional data sources like clinical notes to further enhance model capabilities.