MedGAITS: a graph autoencoder network for modeling irregular time series data in electronic medical records

Area of Science:

• Biomedical Informatics
• Machine Learning
• Clinical Data Analysis

Background:

• Electronic medical records (EMR) are crucial for predicting patient outcomes.
• Irregular sampling and missing data in EMR hinder accurate clinical time-series analysis.
• Existing methods struggle with the complexity of incomplete clinical data.

Purpose of the Study:

• To develop a robust framework for handling irregular and incomplete clinical time-series data.
• To effectively capture complex temporal patterns and feature interactions in EMR.
• To improve the prediction of patient prognosis and disease progression using EMR.

Main Methods:

• Proposed MedGAITS, a two-stage graph autoencoder framework for irregular and incomplete clinical time series.
• Employed a progressive learning strategy with dynamic graph learning for coarse-grained reconstruction.
• Utilized iterative dynamic graph construction and residual learning for refined feature extraction, learning uncertainty-aware representations directly from raw data.

Main Results:

• MedGAITS achieved competitive or superior performance on public datasets (PhysioNet 2012, COVID-19, eICU) for regression and classification tasks.
• Identified key biomarkers for COVID-19 progression, including neutrophils and LDH as early indicators and white blood cell count as a later-stage marker.
• Demonstrated effective handling of irregular and incomplete clinical time-series data.

Conclusions:

• MedGAITS offers an effective solution for analyzing challenging clinical time-series data with missing values.
• The framework enhances downstream predictive task performance and uncovers clinically meaningful, time-evolving features.
• Provides valuable insights for disease monitoring and biomarker discovery.