MUSE-Net: Missingness-aware mUlti-branching Self-attention Encoder for Irregular Longitudinal Electronic Health...

Area of Science:

• Machine learning applications in healthcare
• Data science in clinical decision support
• Biomedical informatics and data analytics

Background:

• Electronic health records (EHRs) offer vast clinical data for data-driven tools.
• Modeling EHRs presents challenges: irregular time series, missing data, and imbalance.
• Advanced analytical models are crucial for unlocking EHR data potential.

Purpose of the Study:

• To propose MUSE-Net, a novel model for longitudinal EHR modeling.
• To address challenges in data-driven disease prediction using EHRs.
• To enhance clinical decision-making through accurate predictions.

Main Methods:

• Developed MUSE-Net with four key modules: MGP for imputation, multi-branching for imbalance, time-aware self-attention for irregular intervals, and interpretable attention.
• Utilized multi-task Gaussian process (MGP) with missing value masks for imputation.
• Employed a time-aware self-attention encoder and a multi-branching architecture.

Main Results:

• MUSE-Net demonstrated superior performance compared to existing methods on synthetic and real-world datasets.
• The model effectively handles irregularly spaced time series and data imbalance.
• Experimental results confirm MUSE-Net's advantage in investigating longitudinal signals.

Conclusions:

• MUSE-Net offers a robust solution for disease prediction using complex EHR data.
• The model's interpretable attention mechanism aids clinical decision support.
• MUSE-Net advances the use of EHRs for enhanced healthcare.