Interpretable Representation Learning for Healthcare via Capturing Disease Progression through Time.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Computational Medicine

Background:

• Electronic Health Records (EHR) are increasingly utilized for predictive modeling.
• Medical claims data represents patients as sequences of healthcare visits with associated medical codes.
• Existing deep learning models face challenges in capturing the temporal dynamics of patient conditions.

Purpose of the Study:

• To introduce Timeline, a novel interpretable deep learning model for predictive modeling of EHR data.
• To incorporate learnable time decay factors and attention mechanisms to improve prediction accuracy and interpretability.
• To analyze the temporal progression of patient conditions and understand the impact of medical codes over time.

Main Methods:

• Developed Timeline, a deep learning architecture with a time decay factor mechanism for medical codes.
• Integrated an attention mechanism to enhance visit vector embeddings.
• Evaluated Timeline on two large-scale real-world medical claims datasets for next hospital visit diagnosis prediction.

Main Results:

• Timeline achieved higher accuracy in predicting the primary diagnosis category for future hospital visits compared to state-of-the-art RNN-based models.
• Learned time decay factors demonstrated alignment with medical knowledge regarding the persistence of chronic versus acute conditions.
• Attention weights and disease progression functions provided interpretable insights into prediction reasoning.

Conclusions:

• Timeline offers a significant advancement in the interpretable deep learning for EHR predictive modeling.
• The model's ability to learn temporal code impacts enhances prediction accuracy and clinical relevance.
• Timeline provides valuable insights into disease progression and future healthcare risks, aiding clinical decision-making.