Abstract
Disease management for heart failure with preserved ejection fraction (HFpEF) requires understanding the comparative effectiveness of real-world drug combinations rather than single agents. Standard randomized controlled trials (RCTs) for multi-drug regimens are prohibitively expensive, slow, and often infeasible at scale, motivating the use of causal machine learning methods on large-scale electronic health records (EHRs). However, reliable estimation of treatment effects is challenging due to the high-order drug–drug interactions, strong confounding factors, and patient heterogeneity across sex, age, and other statuses. Existing causal machine learning methods mostly focus on comparing treatments with pairwise drug combinations. Techniques for multi-drug regimens are less studied. To fill this gap, we propose a Hypergraph–based Doubly Robust framework for HF-pEF (HyperDR), which represents six HF drug classes as nodes and observed multi-drug regimens as hyperedges, and uses a hypergraph neural network to learn shared representations for both drugs and combinations from cross-sectional EHR data. On top of these representations, we jointly train a propensity-score model and an outcome model with a doubly robust objective that combines cross-entropy losses with an augmented inverse-probability–weighted regularizer. This enables consistent treatment-effect estimation when either component is correctly specified while stabilizing learning under rare regimens. Experiments on a real-world HFpEF cohort show that HyperDR improves outcome prediction (hospitalization risk) compared with baseline methods. We also did case studies to interpret model results in treatment rankings across different patient subgroups.
Introduction
Heart failure with preserved ejection fraction (HFpEF) remains a clinical "black box" compared to its counterpart, HFrEF. While HFrEF has established guidelines, HFpEF treatment often involves "polypharmacy"—the use of multiple drug classes to manage various comorbidities such as obesity, diabetes, and hypertension. Determining the optimal combination of these drugs is difficult because the effect of a combination is often more than the sum of its parts.
Randomized Controlled Trials (RCTs) are the gold standard but are logistically and financially impractical for testing the vast space of multi-drug combinations. Observational data from Electronic Health Records (EHRs) offer a potential solution, but causal inference in this context is plagued by confounding by indication, high-dimensional treatment spaces, and complex high-order interactions between drugs. Current methods often struggle with these interactions or rare treatment combinations.
Methodology: The HyperDR Framework
To address these challenges, we developed HyperDR (Hypergraph-based Doubly Robust framework). This approach moves beyond treating drug combinations as simple categorical labels and instead models them using a hypergraph structure.
Hypergraph Representation and Learning
In our model, six primary drug classes are represented as nodes. Any specific multi-drug regimen prescribed to a patient is modeled as a hyperedge connecting the constituent drug nodes. The framework utilizes a Hypergraph Neural Network (HGNN) to learn shared representations. By sharing information across regimens that contain overlapping drug classes, the model can "borrow strength" from common treatments to improve predictions for rarer ones.
Doubly Robust Estimation
The system jointly trains two components: a Propensity Model (predicting the likelihood of a patient receiving a specific drug combination) and an Outcome Model (predicting the risk of hospitalization). These are unified through a Doubly Robust (DR) objective. By incorporating an Augmented Inverse-Probability-Weighted (AIPW) regularizer, the framework ensures that treatment effect estimates remain consistent if either the propensity or the outcome model is correctly specified.
Experimental Results
We evaluated HyperDR using a real-world HFpEF EHR cohort, processing data into 100-day temporal snapshots. The primary goal was to predict HFpEF-related hospitalization risk.
Performance Comparison
HyperDR demonstrated significant improvements over standard Multi-Layer Perceptron (MLP) baselines. As shown in the performance metrics, HyperDR achieved a superior F1-score and AUC for outcome prediction.
Notably, while predicting the exact treatment a physician would choose (propensity) remains difficult for all models (accuracy ~0.23), HyperDR's doubly robust framework allows it to maintain high performance in outcome prediction (AUC 0.93) and reliable treatment effect estimation despite this difficulty.
Clinical Insights
The model's rankings identified RAAS inhibitors and diuretics as highly effective across most patient subgroups. Furthermore, the analysis revealed that treatment effectiveness rankings vary by sex and age, highlighting the potential for HyperDR to inform personalized treatment strategies in heart failure care.
Conclusion
HyperDR provides a robust framework for causal inference in the presence of complex multi-drug interactions. By leveraging hypergraph structures and doubly robust estimation, it overcomes the limitations of traditional observational study methods. Future research will focus on extending this framework to dynamic treatment regimes and validating the model across diverse healthcare systems.
Summary updated