WaveGNN: Integrating Graph Neural Networks and Transformers for Decay-Aware Classification of Irregular Clinical...

Area of Science:

• Biomedical Informatics
• Machine Learning
• Time Series Analysis

Background:

• Clinical time series data often exhibit irregularities like missing values and variable sampling rates.
• Existing methods for handling irregular time series can introduce bias or uninterpretable relationships.

Purpose of the Study:

• To introduce WaveGNN, a novel model designed to process irregular multivariate time series directly.
• To overcome limitations of interpolation and conversion methods in time series analysis.

Main Methods:

• WaveGNN employs a decay-aware Transformer for intra-series dynamics.
• It integrates a sample-specific graph neural network for inter-sensor relationships.
• The model generates sparse, interpretable graphs directly from irregular data.

Main Results:

• WaveGNN demonstrated consistent and robust performance across diverse benchmark datasets (P12, P19, MIMIC-III, PAM).
• It outperformed or matched state-of-the-art baselines in various settings.
• Learned graphs showed alignment with known physiological structures, improving interpretability.

Conclusions:

• WaveGNN offers a robust and interpretable solution for analyzing irregular multivariate clinical time series.
• Its direct processing approach avoids interpolation-induced biases.
• The model supports enhanced clinical decision-making through interpretable graph representations.