Explainable ICD multi-label classification of EHRs in Spanish with convolutional attention

Area of Science:

• Computational linguistics
• Medical informatics
• Artificial intelligence

Background:

• Electronic Health Records (EHRs) are coded using the International Classification of Diseases (ICD), presenting a multi-label classification challenge.
• Existing approaches often function as black boxes, lacking transparency in their predictions.
• Explainable Artificial Intelligence (XAI) offers methods to interpret model decision-making processes.

Purpose of the Study:

• To generate explainable predictions for diseases and procedures within EHRs.
• To visualize attention mechanisms within models to understand prediction drivers.
• To develop a prototype Decision Support System (DSS) that highlights EHR text influencing ICD code assignments.

Main Methods:

• Convolutional Neural Networks (CNNs) integrated with attention mechanisms were employed.
• Attention mechanisms were utilized to identify input (EHR) segments influencing output (medical codes).
• The approach was validated on a Spanish corpus.

Main Results:

• The application of explainable deep learning models to predict medical codes yielded significant results.
• Attention mechanisms successfully identified key textual elements in EHRs driving ICD code predictions.
• A preliminary method for extracting the chronological order of ICD codes within EHRs was demonstrated.

Conclusions:

• Explainable deep learning models applied to EHRs store valuable information that can support medical experts.
• Attention mechanisms provide insights for developing Decision Support Systems (DSS) and understanding diagnostic timelines.
• The study highlights the potential of XAI in enhancing the interpretability and utility of EHR analysis.