EvolveFNN: An Interpretable Framework for Early Detection Using Longitudinal Electronic Health Record Data.

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems
• Explainable AI (XAI)

Background:

• Increasing use of AI in clinical decision support necessitates interpretable models.
• Current models often lack transparency, hindering clinical trust and adoption.
• Longitudinal electronic health records (EHR) data is complex and high-dimensional.

Purpose of the Study:

• Introduce EvolveFNN, an interpretable recurrent neural network model.
• Enable precise and understandable model training using longitudinal EHR data.
• Identify variable encoding functions and significant clinical rules.

Main Methods:

• Developed EvolveFNN by merging fuzzy logic principles with recurrent neural network units.
• Employed supervised learning for training on high-dimensional longitudinal EHR data.
• Validated performance on simulated datasets, a pilot cardiac event detection task, and the MIMIC-III benchmark dataset.

Main Results:

• EvolveFNN achieved superior performance on simulated data, with learned rules closely matching synthetic data generation.
• In cardiac event detection, EvolveFNN showed comparable performance to GRU models and stability across prediction window sizes.
• Extracted rules align with clinical knowledge and suggest novel potential risk factors.

Conclusions:

• EvolveFNN effectively trains accurate, interpretable, and reliable models from longitudinal EHR data.
• The model provides valuable, clinically relevant insights for healthcare professionals.
• Demonstrated generalizability across different datasets and applications.