Multi-task heterogeneous graph learning on electronic health records

Tsai Hor Chan, Guosheng Yin, Kyongtae Bae, Lequan Yu

Electronic Health Records (EHRs) are a goldmine for medical AI, but they are notoriously "messy." They are sparse, highly heterogeneous (mixing medications, diagnoses, and lab results), and riddled with noise. Most existing models treat EHRs as simple sequences or homogeneous graphs, often focusing on a single task like mortality prediction. However, this approach ignores the complex, multi-dimensional relationships that define a patient's journey.

In a recent paper titled "Multi-task Heterogeneous Graph Learning on Electronic Health Records," researchers introduce MulT-EHR, a framework designed to navigate this complexity by treating the hospital database as a living, breathing knowledge graph.

TL;DR

MulT-EHR is a novel framework that models Electronic Health Records as heterogeneous graphs to capture complex medical relationships. By combining causal denoising to strip away data "shortcuts" and multi-task learning to share knowledge across clinical predictions, it sets a new state-of-the-art for tasks ranging from drug recommendation to mortality prediction.

What Problem Does This Paper Address?

The authors identify three primary hurdles in EHR analysis:

  1. Relational Complexity: Standard models often miss the "meta-relations" (e.g., a patient linked to a treatment via a specific visit).
  2. Data Noisiness and Confounding: EHR data contains "backdoor paths"—shortcuts where a model might predict an outcome based on a hospital's specific administrative protocol rather than the patient's actual health status.
  3. Task Isolation: Most models are trained for one specific goal. In reality, predicting a patient's length of stay is deeply related to their risk of readmission; ignoring these overlaps leads to poor generalizability.

Methods

The MulT-EHR framework introduces three key innovations:

  • Heterogeneous Graph Construction: Instead of a flat list, the model builds a graph with six node types (Patients, Visits, Diagnoses, Prescriptions, Procedures, and Lab Events) and five distinct connection types.
  • Causal Denoising Module: Drawing from causal inference, the authors disentangle "causal" features from "trivial" noise. They use a uniform loss (based on Jensen-Shannon divergence) to ensure that trivial features don't provide a predictive "shortcut," forcing the model to rely on robust, invariant medical features.
  • Multi-task Learning with Environment-Invariant Objective: The model uses a single shared-weight GNN (specifically a Heterogeneous Graph Transformer) to predict four tasks simultaneously. To prevent the model from overfitting to one task at the expense of others, they implement a loss function that minimizes the variance between task losses, encouraging the learning of task-invariant representations.

Datasets

The researchers evaluated their model on the two "gold standard" benchmarks in the field:

  • MIMIC-III: A large, freely available database comprising de-identified health-related data from over 40,000 patients.
  • MIMIC-IV: An updated version providing more recent data and a larger patient cohort (over 180,000 patients).

Key Insights and Ideas

  • Relational > Temporal: While Recurrent Neural Networks (RNNs) are the traditional choice for EHRs, this paper suggests that the structural connections between medical entities are often more informative than the mere chronological order of events.
  • The "Shortcut" Problem: The paper highlights how models can be "lazy." By using causal denoising, they prove that explicitly preventing the model from using non-causal correlations (like specific imaging protocols) actually improves its ability to generalize to new patients.
  • Synergy in Prediction: Training on mortality, readmission, length of stay, and drug recommendation at the same time makes the model better at each individual task. The shared knowledge acts as a form of regularization.

Key Results

MulT-EHR consistently outperformed state-of-the-art (SOTA) models (including GRU, Transformer, and specialized EHR models like GraphCare) across all four clinical tasks:

  • Mortality Prediction: Achieved an AUROC of 71.28% on MIMIC-III, significantly higher than the 70.00% of the previous leader, GraphCare.
  • Drug Recommendation: In this highly complex multi-label task, MulT-EHR reached a Jaccard index of 52.20% on MIMIC-III, a substantial jump over the 47.20% achieved by Safedrug.
  • Robustness: Ablation studies showed that removing either the causal denoising or the multi-task module led to a significant drop in performance, validating that both are essential for handling "messy" medical data.

Limitations and Future Work

While powerful, the authors note a few areas for growth:

  • Causal Interpretation: While the model uses causal principles for denoising, the "causal features" are identified via attention weights. The authors acknowledge that more rigorous theoretical work is needed to provide formal causal explanations.
  • Data Modalities: The current model focuses on structured/tabular data. Future iterations could integrate unstructured data like clinical notes (text) or radiology scans (images) to create an even more holistic patient representation.
  • Generalization: While tested on the MIMIC datasets, the framework's performance on data from different hospital systems with different "noise" profiles remains an open question for future research.

Summary updated

Abstract

Learning electronic health records (EHRs) has received emerging attention because of its capability to facilitate accurate medical diagnosis. Since the EHRs contain enriched information specifying complex interactions between entities, modeling EHRs with graphs is shown to be effective in practice. The EHRs, however, present a great degree of heterogeneity, sparsity, and complexity, which hamper the performance of most of the models applied to them. Moreover, existing approaches modeling EHRs often focus on learning the representations for a single task, overlooking the multi-task nature of EHR analysis problems and resulting in limited generalizability across different tasks. In view of these limitations, we propose a novel framework for EHR modeling, namely MulT-EHR (Multi-Task EHR), which leverages a heterogeneous graph to mine the complex relations and model the heterogeneity in the EHRs. To mitigate the large degree of noise, we introduce a denoising module based on the causal inference framework to adjust for severe confounding effects and reduce noise in the EHR data. Additionally, since our model adopts a single graph neural network for simultaneous multi-task prediction, we design a multi-task learning module to leverage the inter-task knowledge to regularize the training process. Extensive empirical studies on MIMIC-III and MIMIC-IV datasets validate that the proposed method consistently outperforms the state-of-the-art designs in four popular EHR analysis tasks — drug recommendation, and predictions of the length of stay, mortality, and readmission. Thorough ablation studies demonstrate the robustness of our method upon variations to key components and hyperparameters.
BibTeX
@article{chan2024multi,
  author = {Tsai Hor Chan and Guosheng Yin and Kyongtae Bae and Lequan Yu},
  title = {{Multi-task heterogeneous graph learning on electronic health records}},
  journal = {Neural Networks},
  year = {2024},
  doi = {10.1016/j.neunet.2024.106644},
}

Loading related content...

Figure