Deep learning prediction models based on EHR trajectories: A systematic review.

Area of Science:

• Health Informatics
• Artificial Intelligence in Medicine
• Computational Health

Background:

• Electronic Health Records (EHRs) generate vast amounts of temporal data, crucial for predicting patient health risks and enabling proactive healthcare.
• Deep learning techniques excel at analyzing complex EHR trajectories, offering significant potential for early disease identification and primary prevention strategies.

Approach:

• A systematic review was conducted, searching major databases (Scopus, PubMed, IEEE Xplore, ACM) from January 2016 to April 2022.
• 63 relevant studies were analyzed, focusing on publication trends, research objectives, and solutions for challenges like data complexity, insufficiency, and model explainability.

Key Points:

• Recurrent neural networks, attention mechanisms, and graph neural networks are frequently employed for modeling long-term dependencies and relationships within EHR trajectories.
• Common prediction targets include general disease onset for the next visit and specific conditions like cardiovascular diseases.
• Techniques like self-attention and attention scores are utilized for enhancing model explainability.

Conclusions:

• Deep learning has significantly advanced the modeling of EHR trajectories, with notable progress in graph neural networks and attention mechanisms.
• Further research is needed to address data insufficiency and develop models capable of handling the multifaceted nature of EHR trajectory data.
• Increasing the availability of public EHR trajectory datasets is crucial for standardized model comparison and validation.