Disease Prediction Using Graph Machine Learning Based on Electronic Health Data: A Review of Approaches and Trends

Haohui Lu, Shahadat UddinHealthcare 2023 · DOI

TL;DR

This paper reviews the application of graph machine learning (GML) techniques, particularly Graph Neural Networks (GNNs), for disease prediction using electronic health data. It highlights that GNNs are outperforming traditional ML methods and are poised to become a significant tool in medical diagnosis and prognosis, despite challenges in interpretability and handling dynamic data.

What Problem Does This Paper Address?

Electronic health records (EHRs) contain a wealth of complex, interconnected information about patients. Traditional machine learning (ML) methods often struggle to effectively leverage this relational data, as they typically assume data is structured in a grid-like (Euclidean) format. However, patient relationships, disease comorbidities, and treatment pathways can be naturally represented as graphs, which are non-Euclidean. This paper addresses the gap in understanding how graph machine learning, a field that excels at processing graph-structured data, is being applied to disease prediction using EHRs, a domain that has seen limited comprehensive review.

Methods

The paper categorizes graph machine learning approaches into two main classes:

  • Shallow Embedding Methods: These methods learn low-dimensional representations (embeddings) of nodes in a graph.
  • Hand-crafted Features: These rely on manually engineered graph properties like centrality, clustering coefficient, etc., to represent nodes.
  • Random Walk-Based Methods: Techniques like DeepWalk and node2vec generate sequences of nodes by performing random walks on the graph, which are then used to learn embeddings.
  • Graph Neural Network (GNN)-Based Methods: These are deep learning models designed to operate directly on graph structures.
  • Graph Convolutional Networks (GCNs): Extend convolutional operations from Euclidean data to graphs, learning node representations by aggregating information from neighbors. The paper discusses spectral-based (e.g., ChebNet) and spatial-based GCNs.
  • Graph Attention Networks (GATs): Incorporate an attention mechanism to weigh the importance of different neighbors during message passing, allowing the model to focus on more relevant information.
  • Graph Auto-Encoders (GAEs): Unsupervised learning models that use GNNs as encoders to learn latent representations and then decode them to reconstruct the graph structure.

The review focuses on two primary tasks within disease prediction: node classification (predicting a label for an unclassified node, e.g., disease presence) and link prediction (predicting the existence of an edge between two nodes, e.g., disease comorbidity).

Datasets

The paper does not introduce new datasets. Instead, it reviews existing studies that utilize real-world electronic health records. These include:

  • Administrative claim data from Australian insurance companies.
  • Real-world electronic health records over several years.
  • Medical Information Mart for Intensive Care (MIMIC-III) and eICU databases.
  • Electronic healthcare records from various medical centers and regions (e.g., Hong Kong, China, US).

The datasets vary significantly in size, from a few thousand records to hundreds of thousands.

Key Insights and Ideas

  • GNNs are the future: The review strongly suggests that GNN-based models are outperforming traditional ML and shallow embedding methods in disease prediction tasks using EHRs, indicating a shift towards deep learning on graph structures.
  • EHRs are inherently graph-like: The complex interdependencies within EHRs (patient-disease, disease-disease, patient-treatment) make graph representations a natural and powerful way to model this data.
  • Node classification is dominant: The majority of reviewed studies focus on node classification tasks, aiming to predict specific diseases or health outcomes for individual patients.
  • Interpretability and dynamic graphs are key challenges: Despite their superior performance, GNNs still face hurdles in explaining their predictions (black box problem) and adapting to evolving, dynamic patient data over time.
  • Lack of standardized benchmarks: The reliance on diverse, real-world EHR datasets makes direct comparison of model performance challenging, highlighting a need for standardized benchmarks in this domain.

Key Results

  • The review identified 18 articles that applied graph ML to disease prediction using EHRs, with a significant increase in GNN-based studies from 2020 onwards.
  • GNN-based models consistently showed superior predictive performance compared to shallow embedding approaches across various disease prediction tasks.
  • Node Classification: Studies reported high accuracies, AUCs, and F1-scores, with specific examples including:
  • Accuracy of 76% for Congestive Heart Failure (CHF) prediction (Liu et al., 2015).
  • AUC of 0.79-0.91 for Type 2 Diabetes risk prediction (Lu et al., 2021).
  • Accuracy of 92.80% for invasive breast cancer prediction (Wang et al., 2020).
  • Link Prediction: Studies also demonstrated strong performance, with AUCs ranging from 0.74 to 0.8969.
  • Hand-crafted features were the most frequently used method among shallow embedding approaches, while GCNs were the most common GNN variant.

Limitations and Future Work

  • Interpretability: The "black box" nature of GNNs remains a significant limitation, hindering clinical adoption where understanding the reasoning behind a prediction is crucial. Future work should focus on developing more interpretable GNN models.
  • Dynamic Graphs: Most current GNN models are designed for static graphs. However, patient data and disease progression are dynamic. Developing GNN approaches that can effectively handle temporal dependencies and evolving network structures is a critical area for future research.
  • Standardized Benchmarks: The lack of standardized datasets and evaluation protocols makes it difficult to compare different GML models rigorously. Creating and utilizing such benchmarks would accelerate progress.
  • Incorporating Node and Edge Features: While GNNs can leverage graph structure, more emphasis is needed on effectively incorporating rich node (patient attributes) and edge (relationship types) features into models.
  • Data Privacy and Availability: While EHR data is less prone to privacy issues due to de-identification, the limited availability of open-source models and datasets hinders reproducibility and further research.
  • Broader Disease Scope: Expanding the application of GML to a wider range of diseases, including emerging ones like COVID-19 and chronic conditions like thyroid diseases, is a promising avenue.

Summary updated

Abstract

Graph machine-learning (ML) methods have recently attracted great attention and have made significant progress in graph applications. To date, most graph ML approaches have been evaluated on social networks, but they have not been comprehensively reviewed in the health informatics domain. Herein, a review of graph ML methods and their applications in the disease prediction domain based on electronic health data is presented in this study from two levels: node classification and link prediction. Commonly used graph ML approaches for these two levels are shallow embedding and graph neural networks (GNN). This study performs comprehensive research to identify articles that applied or proposed graph ML models on disease prediction using electronic health data. We considered journals and conferences from four digital library databases (i.e., PubMed, Scopus, ACM digital library, and IEEEXplore). Based on the identified articles, we review the present status of and trends in graph ML approaches for disease prediction using electronic health data. Even though GNN-based models have achieved outstanding results compared with the traditional ML methods in a wide range of disease prediction tasks, they still confront interpretability and dynamic graph challenges. Though the disease prediction field using ML techniques is still emerging, GNN-based models have the potential to be an excellent approach for disease prediction, which can be used in medical diagnosis, treatment, and the prognosis of diseases.
BibTeX
@inproceedings{lu2023disease,
  author = {Haohui Lu and Shahadat Uddin},
  title = {{Disease Prediction Using Graph Machine Learning Based on Electronic Health Data: A Review of Approaches and Trends}},
  booktitle = {Healthcare},
  year = {2023},
  doi = {10.3390/healthcare11071031},
}
Figure