Deep Dynamic Patient Similarity Analysis: Model Development and Validation in ICU.

Area of Science:

• Artificial Intelligence
• Biomedical Informatics
• Machine Learning

Background:

• Personalized medicine necessitates patient similarity analysis for tailored treatments.
• Challenges in patient similarity analysis include heterogeneous Electronic Health Records (EHRs) and complex disease progression.
• Dynamic retrieval of similar patient sequences is difficult due to diverse disease states over time.

Purpose of the Study:

• To propose a novel dynamic patient similarity analysis model using deep learning.
• To address challenges posed by heterogeneous EHR data and sequential disease progression.
• To validate the model's effectiveness in clinical tasks like diagnosis prediction and medication recommendation.

Main Methods:

• Developed a deep learning model with an embedding and attention module for heterogeneous EHR data.
• Implemented dynamic retrieval of similar patient sequences based on learned representations.
• Integrated a drug-drug interaction (DDI) knowledge graph for safer medication recommendations.
• Evaluated the model on the MIMIC-III critical care database.

Main Results:

• The model significantly outperformed state-of-the-art methods in diagnosis prediction (0.6200 PR-AUC vs. 0.2497–0.5407).
• Achieved superior performance in medication recommendation (0.6682 PR-AUC) compared to the K-nearest model (0.3805 PR-AUC).
• Demonstrated a reduction in adverse drug-drug interactions.

Conclusions:

• The proposed dynamic patient similarity analysis model is effective for clinical decision support systems.
• The model can aid in diagnosis prediction, surgical procedure selection, and medication recommendation.
• The model offers an explainable approach analogous to clinical reasoning for empirical diagnosis and treatment.