Exploiting complex medical data with interpretable deep learning for adverse drug event prediction.

Area of Science:

• Medical informatics
• Computational biology
• Artificial intelligence in healthcare

Background:

• Deep learning models are increasingly used for predictive modeling and knowledge extraction from medical records.
• Temporal attention mechanisms and decay factors are crucial for incorporating time-sensitive information and enhancing interpretability in medical event prediction.

Purpose of the Study:

• To empirically evaluate state-of-the-art medical-code based models for adverse drug event (ADE) prediction.
• To augment an interpretable deep learning architecture with numerical risk and clinical text features for improved ADE prediction.
• To assess the utility of attention mechanisms for medical code-level and text-level interpretability in understanding ADEs.

Main Methods:

• Utilized a large Electronic Patient Record (EPR) dataset including diagnoses, medications, and clinical text.
• Employed and evaluated two state-of-the-art medical-code based deep learning models for ADE prediction.
• Augmented an interpretable deep learning architecture with numerical risk and clinical text features.

Main Results:

• Empirical evaluation of two baseline models on diagnosis and medication data for ADE prediction.
• Demonstrated improved predictive performance of the augmented interpretable deep learning architecture compared to baselines.
• Assessed the importance of attention mechanisms for medical code and text interpretability in ADE prediction.

Conclusions:

• Augmenting interpretable deep learning models with clinical text and numerical risk features enhances ADE prediction accuracy.
• Attention mechanisms are valuable for achieving medical code-level and text-level interpretability, offering insights into ADE occurrence.
• The study provides a framework for leveraging diverse data types within EPRs for robust ADE prediction and understanding.