acceptodds
Under review as a conference paper at ICLR 2027

Hypergraph-Based Doubly Robust Estimation for Causal Inference of Drug Combination Effects in Heart Failure Treatment

11 views

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 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 multi–drug interactions, strong confounding factors, and patient heterogeneity across sex, age, and other characteristics. Existing causal machine learning methods mostly focus on comparing treatments with pairwise drug combinations, while techniques for multi-drug regimens are less studied. To fill this gap, we propose a Hypergraph–based Doubly Robust framework (HyperDR), which represents 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. Our framework jointly trains a propensity score model and an outcome model using cross-entropy losses and an inverse-probability-weighted residual regularizer. Treatment effects are then estimated using an augmented inverse-probability-weighted (AIPW) estimator, which is consistent under standard causal and regularity assumptions if either the propensity scores or conditional outcome means are consistently estimated. On a semi-synthetic benchmark with known ground-truth treatment effects, HyperDR demonstrates improved causal effect recovery by achieving the lowest ATE and CATE RMSE among the evaluated methods. Experiments on two real-world HFpEF cohorts further demonstrate strong factual outcome prediction, while exploratory case studies characterize subgroup-specific treatment rankings.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Hypergraph, Causal Machine Learning, Treatment Efficacy, Heart Failure

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_a5B4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a new model for modelling hospitalization risk for patients with regard to heart failure (HFpEF). The model is split into a propensity model and an outcome model. The models are trained using jointly with a loss that combines both individual losses with a regularizer. The authors especially highlight, that their model uses a hypergraph to encode multi-treatment effects.

Critical strengths
  • The paper is well presented, clearly structured and easy to follow.
  • The application of machine learning to heart failure treatment is relevant and of practical clinical interest.
  • The paper adresses the fact that patients are usually treated with mutliple treatments simultaneously.
Critical weaknesses
  • Higher-order modeling is not actually used. The proposed higher-order graph model is a clique expansion of the hypergraph. It is therefore equivalent to a standard GNN applied to a different (reweighted) pairwise graph. The ablation compares a GNN on the original pairwise graph with this reweighted GNN, but it does not compare against a genuine higher-order model, such as a hypergraph neural network or a simplicial or message-passing model operating on hyperedges. Because higher-orderedness is framed as the paper’s central contribution, this gap is significant. I would strongly encourage the authors to add an ablation with a true higher-order architecture.
  • Marginal empirical improvements. The reported performance is very close to that of competing methods, with only about a 1% improvement. The paper offers no theoretical advances, and its individual components have all been used before, so its contribution rests entirely on empirical evaluation. Under these conditions, the small margin is not sufficient to demonstrate a clear benefit of the proposed approach.
  • Missing runtime analysis. No runtimes or computational costs are reported. Given the small performance gains, a comparison of training and inference time against the baselines is needed to assess the practical trade-off. If this method is to aid in screening, it must be incredibly efficient and locally runnable (privacy constraints, limited compute capacity).
Questions for authors

-Since the clique expansion reduces the proposed model to a standard GNN on a reweighted pairwise graph, in what sense does the model capture higher-order interactions that a pairwise GNN cannot? Could you include a comparison with a genuine higher-order architecture, such as a hypergraph neural network operating directly on hyperedges? -What are the training and inference runtimes of the proposed method compared to the baselines? How does the computational overhead of the clique expansion scale with the size and number of hyperedges? Can you run your model on limited hardware (i.e. without a GPU)?

2 comments
AuthorsOpenReview ↗

Response to Reviewer a5B4

We sincerely thank the reviewer for highlighting the clinical relevance of combination therapy in HFpEF.

W1/Q1. Higher-order interactions. We disagree that HyperDR models only pairwise dependencies. The weighted-graph equivalence applies to the node update in Eq. (1). HyperDR pools each prescribed regimen as a distinct hyperedge (§3.1), a design none of the seven baselines in Tables 1–2 uses. OneFlorida contains all eight combinations of β-blockers, loop diuretics, and MRAs on a RAAS-inhibitor background (Table 6). Eq. (2) allows the β-blocker–diuretic interaction to change when an MRA is added.

We tested direct hyperedge propagation by applying ReLU within each regimen before passing messages back to its drugs. Across 20 additional paired OneFlorida Polynomial runs with a shared evaluation propensity per run, this variant and Eq. (1) gave nearly identical mean errors. Eq. (1) reduced mean CATE RMSE by 8.1% versus the binary GCN.

Treatment encoderAIPW ATE RMSE ↓Outcome CATE RMSE ↓
Binary GCN + attention0.067 ± 0.0140.200 ± 0.041
HGNN + attention (Eq. 1)0.065 ± 0.0130.184 ± 0.034
Direct hyperedge message passing0.105 ± 0.0230.285 ± 0.033

W2/Q2. Effect recovery. HyperDR improved factual F1 by 1.7 percentage points on OneFlorida (80.5% to 82.2%) and 1.0 point on MIMIC-IV (82.0% to 83.0%) over the strongest baseline in each cohort (Table 1). Table 2 evaluates recovery of known treatment effects across all 378 regimen pairs on the OneFlorida Polynomial benchmark. Against NCoRE, the strongest baseline on both metrics, HyperDR reduced ATE RMSE by 36.7% (0.061 to 0.039) and CATE RMSE by 44.2% (0.149 to 0.083) across three runs.

W3/Q3. Time complexity and CPU runtime. We added a time-complexity analysis of the treatment encoder.

A regimen eje_j containing ∣ej∣|e_j| drug classes contributes (∣ej∣2)\binom{|e_j|}{2} distinct drug pairs and ∣ej∣|e_j| diagonal entries to the weighted drug graph. Summing over the KcombK_{\mathrm{comb}} regimens gives O(∑j=1Kcomb∣ej∣2)O(\sum_{j=1}^{K_{\mathrm{comb}}}|e_j|^2) operations to rebuild the clique expansion when the learned hyperedge weights change.

Eq. (1) can also be evaluated directly from the sparse incidence matrix H\mathbf H. First, XW\mathbf X\mathbf W projects KdrugK_{\mathrm{drug}} drug embeddings from dimension dd to dhd_h, costing O(Kdrug d dh)O(K_{\mathrm{drug}}\,d\,d_h). Next, H⊤\mathbf H^\top pools the drug features within each regimen, and H\mathbf H sends the weighted regimen features back to their member drugs. Each pass visits the ∑j∣ej∣\sum_j |e_j| drug–regimen memberships once and processes dhd_h features per membership. The resulting cost per layer is O(Kdrug⋅d⋅dh+dh∑j=1Kcomb∣ej∣).O\left(K_{\mathrm{drug}} \cdot d \cdot d_h + d_h \sum_{j=1}^{K_{\mathrm{comb}}} |e_j|\right). The drug–regimen incidence matrix is 6×286\times28 in OneFlorida and 6×426\times42 in MIMIC-IV. Summing co-membership weights across regimens gives a 6×66\times6 drug–drug matrix in each cohort, with one row and one column per drug class. Thus, for a fixed regimen size, graph construction scales linearly with the number of regimens, while the matrix dimensions depend only on the number of drug classes.

As requested by the reviewer, we also ran CPU-only benchmarks for both training and inference on an Apple M5 Max using one PyTorch thread. Values are medians over three OneFlorida Polynomial runs (~3,900 training snapshots each, dh=32d_h=32, batch size 256). Training time includes validation through early stopping and excludes preprocessing and any separate propensity fit. Inference time covers all 28 regimens for one patient without cached regimen embeddings.

ModelTraining time (s)Inference time (ms/patient)
HyperDR (Eq. 1)0.4340.049
Binary GCN0.3670.038
Direct hyperedge message passing0.4650.060
NCoRE (core adaptation)0.8330.247
GraphITE (core adaptation)0.9930.082

We will include these measurements and additional training and inference benchmarks in the revised paper.

I thank the authors for the detailed response. However, some things have remained unclear.

W1/Q1 Indeed the hyperedges are aggregated in a truly higher-order fashion. However, the message passing is not higher order. I thank the authors for conducting the ablation, but it is unclear to me what "direct hyperedge message passing" is. Could the authors provide a reference? Other than this model, the ablation seems to contain only pairwise models.

W2 I thank the authors for the clarification. However, I must maintain my stance that modest improvements on two real-world datasets does not demonstrate a clear benefit/justification of the approach.

W3 The runtime analysis and CPU evaluation has addressed this concern.

I maintain my score.

Reviewer_KRcjOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper propose a Hypergraph–based Doubly Robust framework (HyperDR), which represents 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.

Critical strengths

Most existing models represent drug combination therapies as simple lists, sequences, or pairwise graphs, while this paper uses a hypergraph.

Critical weaknesses

In my view, the problem setting has limited practical relevance, and the method lacks sufficient novelty.

Questions for authors

See weaknesses

1 comment
AuthorsOpenReview ↗

Response to Reviewer KRcj

We fully disagree with the reviewer KRcj's assessment, which gives no clear reason for either conclusion.

W1. Novelty. Under Strengths, the reviewer writes, “Most existing models represent drug combination therapies as simple lists, sequences, or pairwise graphs, while this paper uses a hypergraph.” Under Weaknesses, the reviewer says the method “lacks sufficient novelty.” The reviewer calls this a strength, then says the method lacks novelty without explaining why. The review doesn’t name a single earlier method that does the same thing.

W2. Practical relevance. We evaluated HyperDR on two real-world HFpEF cohorts. OneFlorida records 28 prescribed regimens among 5,697 patients, and MIMIC-IV records 42 among 16,432 patients (§4.1). HyperDR achieved the highest hospitalization-prediction F1 on both cohorts, 82.2% and 83.0%, respectively (Table 1). Reviewer xUHn calls this “a meaningful combination-treatment problem” and gives the paper a 3 for informativeness. Reviewer a5B4 likewise describes the application as “relevant and of practical clinical interest” and gives the paper a 3 for both soundness and presentation. By contrast, Reviewer KRcj simply states that the problem has “limited practical relevance” without giving any reason or supporting detail.

Reviewer_xUHnOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

HyperDR represents six drug classes as nodes and observed regimens as hyperedges, learning shared treatment representations for outcome prediction. A shared patient encoder supports outcome and propensity heads, trained with cross-entropy losses and squared inverse-propensity residual regularization. AIPW then estimates marginal treatment risks. Evaluation includes two HFpEF EHR cohorts and a polynomial semi-synthetic benchmark with known treatment effects.

Critical strengths
  1. The formulation addresses a meaningful combination-treatment problem. Sharing drug representations across overlapping regimens is a sensible inductive bias, and the nonlinear outcome head permits patient-dependent, nonadditive responses. The comparisons include combination-aware methods, not only binary-treatment baselines.
  2. The semi-synthetic results provide relevant positive evidence. Table 2 reports AIPW ATE RMSE of 0.039 +/- 0.005, versus 0.061 +/- 0.011 for NCoRE, and outcome-based CATE RMSE of 0.083 +/- 0.009, versus 0.149 +/- 0.022. Appendix D uses patient-disjoint holdout evaluation, keeps oracle errors out of model selection, and compares all treatment effects on the same test population. It appropriately distinguishes these results from identification of the original clinical effects.
Critical weaknesses
  1. The residual training objective does not generally preserve propensity consistency. Equation 3 and Appendix D.3, Equation 11, allow the squared residual term to backpropagate through the propensity prediction. Holding outcome predictions fixed, let ete_t be the true propensity, ptp_t its fitted value, and rt=E[(Y−μ^t(X))2∣X=x,T=t]r_t=E[(Y-\hat\mu_t(X))^2\mid X=x,T=t]. Away from clipping, the propensity-dependent conditional population objective is

    +\lambda_2\sum_t\frac{e_t r_t}{p_t^2}.$$ At $p=e$, the first-order conditions generally fail unless $r_t/e_t^2$ is constant across treatments. Thus, even a correctly specified propensity model need not estimate the true assignment probabilities under this objective. The standard AIPW identity remains valid conditional on consistent nuisance estimation; the gap concerns the proposed training procedure. Validation checkpointing and temperature calibration may help but do not establish consistency. Moreover, Appendix D.3 uses a residual weight of $10^{-4}$, whereas the clinical sensitivity analysis explores $0.01$-$0.4$ (Section 4.5); the final clinical weight is not clearly specified. The reported causal benchmark therefore leaves the penalty's contribution and behavior at the clinical settings unresolved.
  2. The specific causal benefit of hypergraph encoding is not isolated. Equation 1 is normalized graph convolution on the weighted co-membership matrix A=HWeDe−1H⊤A=H W_e D_e^{-1}H^\top. Attention aggregation and the nonlinear outcome head can still model interactions, but this propagation alone does not establish an advantage over matched graph or set encoders. The MLP ablation treats regimens as unrelated categories, and the GCN's graph construction and pooling are not specified. Table 3 evaluates these ablations only on factual classification. Table 2 compares complete estimation procedures, including different propensity estimators, as Appendix D.3 acknowledges. Consequently, the strong semi-synthetic results cannot yet be attributed specifically to the hypergraph component.

Questions for authors
  1. Does the residual penalty improve ATE estimation without degrading propensity calibration? Please compare the current method with zero residual penalty and a separately fitted propensity model, including a setting with a correctly specified propensity model and a misspecified outcome model. Report ATE error, propensity calibration, and the actual clinical loss weight.
  2. Do the causal-recovery gains persist when the existing encoder ablations and a simple drug-membership encoder are evaluated on the same semi-synthetic repetitions with common propensity estimates? Please specify the initial drug features, attention pooling, and how the GCN differs from Equation 1's induced weighted graph.
1 comment
AuthorsOpenReview ↗

Response to Reviewer xUHn

We sincerely thank the reviewer for highlighting our treatment-effect recovery results (Table 2).

W1. Propensity consistency. We agree that Eq. (3) can shift the fitted propensity even when the model family contains the truth. Our abstract states that AIPW in Eq. (4) is consistent if either the propensity or the outcome regression is consistently estimated, under the usual causal and regularity assumptions. We will use “consistently estimated” in §3.3 as well.

Q1. Residual penalty. With the jointly fitted propensity, the residual weight used in Appendix D.3 (10−410^{-4}) lowered AIPW ATE RMSE relative to zero penalty in all three additional paired runs on the OneFlorida Polynomial semi-synthetic benchmark. Mean RMSE fell by 9.6%, while mean classwise expected calibration error (ECE) rose by 0.005 percentage points.

We reimplemented Eqs. (1)–(4) using the same data and initialization across weights within each run. At each weight, we evaluated the same outcome predictions with either the joint propensity or a separately fitted propensity held fixed across weights. Values are mean ± sample SD.

Residual weightJoint ATE RMSE ↓Joint ECE (pp) ↓Separate ATE RMSE ↓Separate ECE (pp) ↓
00.066 ± 0.0110.724 ± 0.0310.062 ± 0.0080.680 ± 0.034
10−410^{-4}0.060 ± 0.0140.729 ± 0.0190.062 ± 0.0080.680 ± 0.034
0.10.076 ± 0.0121.314 ± 0.0680.061 ± 0.0080.680 ± 0.034

In 20 paired simulations with a correctly specified propensity and an outcome model that omitted drug interactions, increasing the residual weight from 0 to 0.1 raised mean ATE RMSE from 0.013 to 0.022 with the joint propensity. At 0.1, switching to the separate propensity brought it back to 0.013.

W2/Q2. Encoder ablation. We compared four treatment encoders on the OneFlorida Polynomial semi-synthetic benchmark over 20 runs, using the same evaluation propensity within each run. Both graph encoders used trainable drug embeddings and softmax attention over each regimen’s drugs. Eq. (1) uses the induced weighted co-membership graph. The GCN uses binary co-membership edges with self-loops and symmetric normalization.

Treatment encoderAIPW ATE RMSE ↓Outcome CATE RMSE ↓
Regimen-ID MLP0.078 ± 0.0110.201 ± 0.014
Six-bit membership MLP0.080 ± 0.0100.205 ± 0.018
Binary GCN + attention0.067 ± 0.0140.200 ± 0.041
HGNN + attention (Eq. 1)0.065 ± 0.0130.184 ± 0.034

HGNN reduced mean CATE RMSE by 8.1% relative to the binary GCN, with lower error in 19 of 20 runs.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.