Multi-task heterogeneous graph learning on electronic health records.

Area of Science:

• Computational medicine
• Health informatics
• Machine learning in healthcare

Background:

• Electronic Health Records (EHRs) offer rich data for medical diagnosis but suffer from heterogeneity, sparsity, and complexity.
• Existing EHR modeling methods often focus on single tasks, limiting generalizability and performance across diverse analytical problems.
• Graph-based approaches show promise for EHRs due to complex entity interactions, but challenges remain in handling data characteristics.

Purpose of the Study:

• To propose a novel framework, MulT-EHR (Multi-Task EHR), for enhanced EHR modeling.
• To address the heterogeneity and complexity of EHR data using a heterogeneous graph representation.
• To overcome limitations of single-task learning by enabling simultaneous multi-task prediction and improving generalizability.

Main Methods:

• Leveraging a heterogeneous graph to capture complex relationships and model EHR heterogeneity.
• Incorporating a denoising module based on causal inference to mitigate confounding effects and reduce data noise.
• Implementing a multi-task learning module within a single graph neural network to facilitate knowledge sharing and regularize training.

Main Results:

• MulT-EHR consistently outperformed state-of-the-art methods on MIMIC-III and MIMIC-IV datasets.
• The framework demonstrated superior performance in four key EHR analysis tasks: drug recommendation, length of stay prediction, mortality prediction, and readmission prediction.
• Ablation studies confirmed the robustness and effectiveness of MulT-EHR's core components and hyperparameter settings.

Conclusions:

• The proposed MulT-EHR framework offers a robust and effective solution for modeling complex EHR data.
• Multi-task learning on heterogeneous graphs significantly enhances predictive accuracy and generalizability in EHR analysis.
• MulT-EHR provides a promising direction for advancing computational approaches in clinical informatics and medical diagnosis.