acceptodds
Under review as a conference paper at ICLR 2027

SOAR-SSM: SOBOLEV-REGULARIZED STATE SPACE MODEL WITH ARMA NEURAL NETWORKS FOR TEMPORAL HYPERGRAPH REASONING IN HEALTH CARE

Abstract

Medical knowledge graphs (MKGs) provide a structured mechanism for incorporating clinical knowledge into electronic health record (EHR) modeling. However, existing MKG-based approaches remain limited in three important aspects. First, they predominantly encode pairwise relations or collapse temporal records into static graphs, thereby overlooking higher-order clinical interactions and their evolution across visits. Second, graph representations are commonly learned using message-passing operators that exhibit a strong low-frequency bias and are susceptible to oversmoothing, potentially suppressing localized and clinically salient variations distributed across different spectral regimes. Third, existing approaches often lack an explicit mechanism for preserving long-range patient history. To address these limitations, we propose SoAR-SSM, a Sobolev-regularized State Space Model with Auto Regressive Moving Average (ARMA) hypergraph neural networks for frequency-aware temporal hypergraph reasoning. SoAR-SSM represents each clinical visit as a knowledge-infused hypergraph to capture higher-order interactions among conditions, procedures, and medications. We then derive a Sobolev-regularized temporal SSM from first principles, yielding a topology-aware spectral operator that governs how structural information is injected into long-range temporal memory. To obtain a more expressive spectral realization, we approximate this operator using parallel ARMA hypergraph neural networks, whose rational responses enable the model to capture complementary information across multiple frequency regimes while mitigating oversmoothing. We evaluate SoAR-SSM on MIMIC-III and MIMIC-IV across multiple clinical prediction tasks. The results demonstrate the effectiveness of jointly modeling higher-order clinical interactions, multi-frequency structural information, and long-range temporal dependencies, while maintaining parameter-efficient EHR representation learning.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Sobolev regularizer, State space models, Temporal knowledge graphs

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_5nfLOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper explores how to jointly model the high-order clinical relationships within a single visit in electronic health records and the long-term dependencies across visits. The authors first construct an ultra-knowledge graph using the semantic grouping of medical concepts, and based on this, generate a patient-level ultra-graph. Subsequently, the model extracts structural representations within the visit through multiple parallel ARMA ultra-graph neural networks, and then inputs the representations of visit nodes into a state space model based on HiPPO to form the patient's longitudinal representation, which is used for tasks such as death prediction, hospital stay duration prediction, and drug recommendation. The main value of this paper lies in integrating the encoding of ultra-graph structure and HiPPO temporal modeling into a framework applicable to longitudinal EHR. However, there are still some theoretical correspondence relationships that need to be further clarified between Sobolev projection, memory update under dynamic topology, and the final nonlinear ARMA encoder and the patient-level state.

Critical strengths
  1. The paper distinguishes between two complementary modeling problems: the joint relationships among multiple clinical concepts within a single visit, and the longitudinal dependencies between different visits of patients. Representing the former with hypergraphs and the latter with state space models is a reasonable and understandable design. The paper further focuses on the potential over-smoothing problem caused by conventional hypergraph convolutions, and attempts to retain the distinguishable node information through more flexible spatial encoding.
  2. The method has a certain degree of theoretical organization. Starting from the HiPPO projection problem with Sobolev regularity terms, a closed-form solution is provided to establish the connection between the hypergraph structure and the historical memory coefficients.
  3. The experiment covered multiple prediction tasks and achieved significant performance improvements. The main results are highly appealing.
  4. The paper not only compares the complete model with external baselines, but also includes comparisons between ordinary graphs and hypergraphs, hypergraph convolution and ARMA, SSM and LSTM, as well as analyses for longer medical history. The appendix further examines the medical ontology, co-occurrence clustering, and the hypergraph construction methods of different LLMs.
Critical weaknesses
  1. There is still a lack of clear boundary connection between the state update in the dynamic topology and the Sobolev regularized projection solution. The author has already explained in Remark that the derivative of the topological change boundary is not calculated, and a strategy of "first diffusing the current input and then updating the memory" is adopted. This explanation can define a reasonable piecewise state space model, but it has not fully explained whether this model still corresponds to the projection optimal solution of Theorem 3.1 at each moment.
  2. The evidence of long-term reliance on modeling is still relatively limited, and it cannot fully explain the gains on the overall dataset. The author compared SSM, LSTM, and the model without time-series aggregation on a subset of patients who had more than 10 visits. The results supported the role of SSM in this setting. However, this experiment did not fully report the subset size, uncertainty estimation, and the threshold settings corresponding to sensitivity, making it difficult to judge the improved stability. Additionally, the overall cohort contains a large number of short visit sequences. The larger overall improvement on MIMIC-IV cannot be attributed solely to the dataset containing more long-term patients, due to the long-term reliance on modeling. More direct evidence should include results stratified by historical length and the performance changes after gradually truncating the history.
  3. The relationship between the Sobolev filter and the final nonlinear ARMA-SSM implementation is not yet fully established. The direct correspondence established in the appendix mainly pertains to the linear recursion when β=1\beta=1, and relies on appropriate sharing coefficients and convergence conditions. However, the actual model uses finite-layer recursion, nonlinear activation, and learnable feature transformation matrices.
Questions for authors
  1. When Rk≠Rk−1R_k \neq R_{k-1}, how does Equation (5) maintain Sk=RkCkS_k = R_kC_k? Are there any implicit boundary state transformations?
  2. After adopting the readout at the treatment level and the finite-depth nonlinear ARMA, which Sobolev–HiPPO theoretical results still hold strictly?
  3. Under the same knowledge, residual structure and parameter budget, how much is the current ARMA module's advantage over other hypergraph encoders?
  4. Can the size of the long historical subset, confidence intervals, and historical truncation experiments be supplemented?
  5. Do all baselines use the same task-specific patient cohort, prediction time point, and input truncation rules?

If the author addresses my concerns and clarifies my doubts, I will consider upgrading to 3.

10 comments
AuthorsOpenReview ↗

Q1 and W1, On memory update at topology change

We thank the reviewer for the encouraging and insightful review. Theorem 3.1 characterizes instantaneous Sobolev-optimal coefficients; Eqs. (5) and (9) define causal memory of topology-filtered visits. Their state conventions differ at topology changes.

1. Topology switches and Eq. (5). At visit kk, Theorem 3.1 gives the exact optimizer of Eq. (2):

Skinst=RkCk,Rk=[I+λ(LˉH,k+ϵI)β]−1.S_k^{\mathrm{inst}}=R_kC_k,\qquad R_k=\left[I+\lambda(\bar L_{H,k}+\epsilon I)^\beta\right]^{-1}.

On aligned nodes, maintaining this identity across a switch with continuous CC requires

Sinst(tk−1+)=RkRk−1−1Sinst(tk−1−).S^{\mathrm{inst}}(t_{k-1}^{+})=R_kR_{k-1}^{-1}S^{\mathrm{inst}}(t_{k-1}^{-}).

SoAR-SSM instead carries accumulated memory forward unchanged and uses the current topology only to filter new information. Under zero-order hold, R(t)=RkR(t)=R_k and X(t)=XkX(t)=X_k on (tk−1,tk](t_{k-1},t_k], with

S˙(t)=S(t)A⊤+RkXkB⊤,S(tk−1+)=S(tk−1−).\dot S(t)=S(t)A^\top+R_kX_kB^\top,\qquad S(t_{k-1}^{+})=S(t_{k-1}^{-}).

Exact integration yields Eq. (5):

Sk=Sk−1Aˉk+RkXkB⊤Gk,S_k=S_{k-1}\bar A_k+R_kX_kB^\top G_k,

where Aˉk=eΔkA⊤\bar A_k=e^{\Delta_kA^\top} and Gk=∫0ΔkeuA⊤,duG_k=\int_0^{\Delta_k}e^{uA^\top}\\,du.

This matches S=RkCS=R_kC dynamics within constant-topology intervals but preserves memory across switches. Thus Eq. (5) is an exact ZOH discretization of the specified filtered-input dynamics, not a claim that Sk=RkCkS_k=R_kC_k continues to hold under arbitrary topology changes. This also follows from the diffuse-then-update recurrence structure of (Li et., al., 2024, ref line no:599), however adapted to our modelling strategy. Additionally, we also provide a variational characterization supporting the diffusion then update.

2. Sobolev/HiPPO variational meaning. Let Qk=(LˉH,k+ϵI)β⪰0Q_k=(\bar L_{H,k}+\epsilon I)^\beta\succeq0. Then each visit has the exact spatial Sobolev optimizer as

Zk⋆=arg⁡min⁡Z[12∥Z−Xk∥F2+λ2Tr⁡(Z⊤QkZ)]=RkXk.Z_k^{\star}=\arg\min_Z\left[\tfrac12\Vert Z-X_k\Vert_F^2+\tfrac{\lambda}{2}\operatorname{Tr}(Z^\top Q_kZ)\right]=R_kX_k.

The compressed memory follows zq(τ)=R(τ)xq(τ)z_q(\tau)=R(\tau)x_q(\tau), where R(τ)R(\tau) is the topology at observation time. For an orthonormal HiPPO basis ptKp_{t_K} and its measure μtK\mu_{t_K},

Mq,K⋆=arg⁡min⁡M12∫0tK∥zq(τ)−MptK(τ)∥22,dμtK(τ),M_{q,K}^{\star}=\arg\min_M\tfrac12\int_0^{t_K}\Vert z_q(\tau)-Mp_{t_K}(\tau)\Vert_2^2\\,d\mu_{t_K}(\tau),

which, by orthonormality, equals Mq,K⋆=∫0tKzq(τ)ptK(τ)⊤,dμtK(τ).M_{q,K}^{\star}=\int_0^{t_K}z_q(\tau)p_{t_K}(\tau)^\top\\,d\mu_{t_K}(\tau).

A matching HiPPO state-space realization has the filtered-input structure of Eq. (5). Hence, in the idealized linear setting, the two-stage interpretation is Sobolev-regularize each visit, then temporally compress it. Theorem 3.1 asserts the latter's instantaneous optimality, not equality of the two histories after topology changes.

3. Exact discrepancy characterization. On a common aligned node space with matching transitions, zero initialization and Uk=XkB⊤GkU_k=X_kB^\top G_k, unrolling gives

SK−RKCK=∑k=1K(Rk−RK)Uk(∏j=k+1KAˉj),S_K-R_KC_K=\sum_{k=1}^K(R_k-R_K)U_k\left(\prod_{j=k+1}^K\bar A_j\right),

and the resolvent identity gives

Rk−RK=λRk(QK−Qk)RK.R_k-R_K=\lambda R_k(Q_K-Q_k)R_K.

Since Qk⪰0Q_k\succeq0 implies ∥Rk∥2≤1\Vert R_k\Vert_2\leq1,

∥SK−RKCK∥F≤λ∑k=1K∥QK−Qk∥2,∥Uk∥F,∥∏j=k+1KAˉj∥2.\Vert S_K-R_KC_K\Vert_F\leq\lambda\sum_{k=1}^K\Vert Q_K-Q_k\Vert_2\\,\Vert U_k\Vert_F\\,\left\Vert\prod_{j=k+1}^K\bar A_j\right\Vert_2.

Thus the two states coincide for static topology; otherwise their difference is controlled by operator variation, input strength and temporal propagation.

4. Eq. (9) and actual implementation. Each visit is processed independently in its own hypergraph:

Z~k=fθ(Lk,Xk),uk=ev,k⊤Z~k∈R1×d,\widetilde Z_k=f_\theta(L_k,X_k),\qquad u_k=e_{v,k}^\top\widetilde Z_k\in\mathbb R^{1\times d}, where fθf_\theta is the ARMA encoder and

sk=sk−1Aˉk+ukB⊤Gk,s_k=s_{k-1}\bar A_k+u_kB^\top G_k, The model stores one fixed-dimensional patient state, not persistent states for individual hypergraph nodes. Variable node identities/counts do not require cross-topology basis transformations: all topology-specific reasoning occurs before the visit embedding enters patient memory. With aligned nodes, exact RkR_k and a fixed selector eve_v, Eq. (9) is the left projection sk=ev⊤Sks_k=e_v^\top S_k of Eq. (5); with changing node sets/selectors it directly defines a fixed-dimensional patient model rather than a lossless projection of node-wise memory. As mentioned, maintaining per node wise statistics is difficult due to changing concept sets. The design is motivated from downstream tasks where all our tasks are patient-level and would require a pooling step. As Linear pooling commutes with propagation, given the readout is fixed, performing pooling through a fixed visit coordinate is lossless (detailed derivation below) with respect to the propagated dynamics.

We made this distinction clear in Secs. 3.2, 3.4 of the revised draft. To summarize, Theorem 3.1 gives exact instantaneous spatial regularization; Eq. (5) gives exact linear ZOH carried-memory evolution; Eq. (9) implements the corresponding causal diffuse-then-update principle with shared patient-level memory.

AuthorsOpenReview ↗

W2 and Q4 Long range

Repeated-run robustness: We additionally report mean ± standard deviation across 10 runs on long-range mortality prediction task on MIMIC-IV dataset. Specifically, it includes 59,262 patients, 964 have >10 visits, yielding 5,912 eligible prediction samples with 208 positive outcomes. The patient-disjoint train/validation/test splits contain 670/153/141 patients, 3,970/977/965 samples, and 144/36/28 positive outcomes, respectively. Importantly, we included results on SSM variants for EHR data (EHRMAMBA), whereas graph based SSM results are already presented in Table 3 of the manuscript.

Mean and standard deviation results across multiple seed runs.

ModelAUROC (%)Macro-F1 (%)
Hyper-GNN60.6 ± 0.362.8 ± 1.2
SoAR-LSTM65.1 ± 0.568.9 ± 0.9
HypEHR63.2 ± 0.767.1 ± 1.4
EHRMamba67.6 ± 1.372.3 ± 1.7
SoAR-SSM69.3 ± 1.276.1 ± 1.4

SoAR-SSM outperforms SoAR-LSTM by +4.2 AUROC/+7.2 Macro-F1 points and EHRMamba by +1.7 AUROC/+3.8 Macro-F1 points, while requiring substantially fewer parameters. The improvements are consistent with the reported run-to-run variability, with particularly clear gains in Macro-F1.

Overall, these controlled, independently tuned comparisons demonstrate that SoAR-SSM achieves superior long-range prediction performance with lower model capacity, strengthening the empirical evidence for effective and parameter-efficient temporal modeling.

To directly investigate reliance on historical information, we conduct a progressive history-truncation experiment on the MIMIC-IV long-history subset (>10 visits), retaining only the most recent (K∈{1,5,10}K\in\lbrace 1,5,10\rbrace) visits or the full available history.

ModelK=1K=5K=10Full history
SoAR-SSM (AUROC)0.5860.6030.6320.693
SoAR-LSTM (AUROC)0.5230.5740.5920.651

AUROC improves monotonically with increasing historical context, from 58.6% (K=1) to 69.3% (full history), an absolute improvement of 10.7 percentage points. Notably, retaining visits beyond the most recent 10 provides an additional 6.1-point improvement, indicating that earlier clinical information contributes predictive value beyond recent observations. These findings complement the SSM-versus-LSTM ablation in above table and provide direct empirical evidence supporting the model's utilization of long-range patient history.

AuthorsOpenReview ↗

W3, Generalization to

We thank the Reviewer for the comment. The manuscript provides the connection of Sobolev filter to ARMA-1 filter and remarks the connection to ARMA-K filters. However, for completeness we present the result for any integer β\beta. Let LˉH\bar{\mathbf{L}}_ H be a symmetric positive semidefinite hypergraph Laplacian, and define

RH,β=[I+λ(LˉH+ϵI)β]−1,λ>0,ϵ≥0.\mathbf{R}_ {H,\beta}=\left[\mathbf{I}+\lambda\left(\bar{\mathbf{L}}_ H+\epsilon\mathbf{I}\right)^\beta\right]^{-1},\qquad \lambda>0,\quad \epsilon\ge 0 .

Recall, for an eigenvalue μ\mu of LˉH\bar{\mathbf{L}}_ H, its spectral response is

hβ(μ)=11+λ(μ+ϵ)β.h_\beta(\mu)=\frac{1}{1+\lambda(\mu+\epsilon)^\beta}.

Now we express the relation of β\beta Sobolev filter to generalized ARMA filter.

Proposition C.1 (Integer-order Sobolev–ARMA decomposition). For every positive integer β\beta, the operator RH,β\mathbf{R}_ {H,\beta} admits an exact decomposition into K=βK=\beta first-order rational branches i.e,

 RH,β=∑j=1βbj(I−ajLH)−1 ,(C.1)\boxed{\ \mathbf{R}_ {H,\beta}=\sum_{j=1}^{\beta}b_j\left(\mathbf{I}-a_j\mathbf{L}_ H\right)^{-1}\ }, \qquad\text{(C.1)}

where LH=I−LˉH\mathbf{L}_ H=\mathbf{I}-\bar{\mathbf{L}}_ H, aj=−1cj,bj=rjβcja_j=-\frac{1}{c_j},\qquad b_j=\frac{r_j}{\beta c_j} with cj=rj−(1+ϵ)c_j=r_j-(1+\epsilon) and rj=λ−1/βexp⁡((2j−1)πiβ),∀j=1,…,β,r_j=\lambda^{-1/\beta}\exp\left(\frac{(2j-1)\pi i}{\beta}\right),\qquad \forall j=1,\ldots,\beta,

Proof. Consider the polynomial as p(y)=1+λyβp(y)=1+\lambda y^\beta. Observe the roots of p(y)p(y) i.e., r1,…,rβr_1,\ldots,r_\beta are distinct and nonzero, and p′(rj)=λβrjβ−1≠0p'(r_j)=\lambda\beta r_j^{\beta-1}\neq0, so every root is simple, as λrjβ=−1\lambda r_j^\beta=-1 and 1p′(rj)=−rjβ.\frac{1}{p'(r_j)}=-\frac{r_j}{\beta}.

The partial-fraction expansion gives

11+λyβ=∑j=1β1/p′(rj)y−rj=1β∑j=1βrjrj−y.(C.2)\frac{1}{1+\lambda y^\beta}=\sum_{j=1}^{\beta}\frac{1/p'(r_j)}{y-r_j}=\frac{1}{\beta}\sum_{j=1}^{\beta}\frac{r_j}{r_j-y}.\qquad\text{(C.2)}

Let M=LˉH+ϵI\mathbf{M}=\bar{\mathbf{L}}_ H+\epsilon\mathbf{I}. Each rjr_j is either nonreal or negative, so every rjI−Mr_j\mathbf{I}-\mathbf{M} is invertible. Applying (C.2) yields

RH,β=1β∑j=1βrj(rjI−M)−1.(C.3)\mathbf{R}_ {H,\beta}=\frac{1}{\beta}\sum_{j=1}^{\beta}r_j\left(r_j\mathbf{I}-\mathbf{M}\right)^{-1}.\qquad\text{(C.3)}

Since M=(1+ϵ)I−LH\mathbf{M}=(1+\epsilon)\mathbf{I}-\mathbf{L}_ H, we obtain rjI−M=cjI+LH=cj(I−ajLH),r_j\mathbf{I}-\mathbf{M}=c_j\mathbf{I}+\mathbf{L}_ H=c_j\left(\mathbf{I}-a_j\mathbf{L}_ H\right), and in particular I−ajLH\mathbf{I}-a_j\mathbf{L}_ H is invertible. Substituting into (C.3) gives 1βrj(rjI−M)−1=bj(I−ajLH)−1\frac{1}{\beta}r_j\left(r_j\mathbf{I}-\mathbf{M}\right)^{-1}=b_j\left(\mathbf{I}-a_j\mathbf{L}_ H\right)^{-1}, which proves (C.1).

For β≥2\beta\ge2 the roots occur in conjugate pairs (rβ+1−j=rˉjr_{\beta+1-j}=\bar r_j, hence aβ+1−j=aˉja_{\beta+1-j}=\bar a_j, bβ+1−j=bˉjb_{\beta+1-j}=\bar b_j), so each pair sums to a real second-order rational branch and, for odd β\beta, the real root −λ−1/β-\lambda^{-1/\beta} gives one real first-order branch; thus RH,β\mathbf{R}_ {H,\beta} is realizable with real weights.

Connection to the rational ARMA family. For integer β\beta,

hβ(μ)=11+λ∑m=0β(βm)ϵβ−mμm.h_\beta(\mu)=\frac{1}{1+\lambda\sum_{m=0}^{\beta}\binom{\beta}{m}\epsilon^{\beta-m}\mu^m}.

Dividing the numerator and denominator by 1+λϵβ1+\lambda\epsilon^\beta obtains the rational ARMA family.

Proposition C.2 (Recursive realization and truncation error). For an input X\mathbf{X}, initialize Zj(0)=0\mathbf{Z}_ j^{(0)}=\mathbf{0} consider

Zj(ℓ+1)=ajLHZj(ℓ)+bjX.(C.4)\mathbf{Z}_ j^{(\ell+1)}=a_j\mathbf{L}_ H\mathbf{Z}_ j^{(\ell)}+b_j\mathbf{X}.\qquad\text{(C.4)}

Further if γj=∥ajLH∥2<1\gamma_j=\lVert a_j\mathbf{L}_ H\rVert_2<1 for every jj, then

lim⁡L→∞∑j=1βZj(L)=RH,βX.\lim_{L\to\infty}\sum_{j=1}^{\beta}\mathbf{Z}_ j^{(L)}=\mathbf{R}_ {H,\beta}\mathbf{X}.

Whereas the error between the representation are bounded as

 ∥∑j=1βZj(L)−RH,βX∥F≤∑j=1β∣bj∣ γjL1−γj ∥X∥F. (C.5)\boxed{\ \left\lVert\sum_{j=1}^{\beta}\mathbf{Z}_ j^{(L)}-\mathbf{R}_ {H,\beta}\mathbf{X}\right\rVert_F\le\sum_{j=1}^{\beta}\frac{|b_j|\thinspace\gamma_j^{L}}{1-\gamma_j}\thinspace\lVert\mathbf{X}\rVert_F .\ }\qquad\text{(C.5)}

Proof. Unrolling (C.4) gives

Zj(L)=bj∑ℓ=0L−1(ajLH)ℓX.\mathbf{Z}_ j^{(L)}=b_j\sum_{\ell=0}^{L-1}\left(a_j\mathbf{L}_ H\right)^{\ell}\mathbf{X}.

Since γj<1\gamma_j<1, the Neumann series converges to Zj⋆=bj(I−ajLH)−1X\mathbf{Z}_ j^\star=b_j\left(\mathbf{I}-a_j\mathbf{L}_ H\right)^{-1}\mathbf{X}, and satisfies

∥Zj(L)−Zj⋆∥F≤∣bj∣∑ℓ=L∞γjℓ ∥X∥F=∣bj∣ γjL1−γj ∥X∥F.\left\lVert\mathbf{Z}_ j^{(L)}-\mathbf{Z}_ j^\star\right\rVert_F\le|b_j|\sum_{\ell=L}^{\infty}\gamma_j^{\ell}\thinspace\lVert\mathbf{X}\rVert_F=\frac{|b_j|\thinspace\gamma_j^{L}}{1-\gamma_j}\thinspace\lVert\mathbf{X}\rVert_F ..

Summing these bounds and applying Proposition C.1 proves the result.

Recovery of the first-order construction. For β=1\beta=1, r1=−1/λr_1=-1/\lambda, which gives

a1=λ1+λ(1+ϵ),b1=11+λ(1+ϵ).a_1=\frac{\lambda}{1+\lambda(1+\epsilon)},\qquad b_1=\frac{1}{1+\lambda(1+\epsilon)} .

Thus RH,1=b1(I−a1LH)−1\mathbf{R}_ {H,1}=b_1\left(\mathbf{I}-a_1\mathbf{L}_ H\right)^{-1}, recovering the original first-order Sobolev–ARMA connection. Here 0<a1<10<a_1<1, so the recursion converges for every λ>0\lambda>0.

For completeness, detailed connection of β\beta to generalized ARMA filter is included in Appendix C

Reviewer_hKJFOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SoAR-SSM for clinical prediction from longitudinal electronic health records. It combines knowledge-infused visit-level hypergraphs, ARMA hypergraph encoders, and HiPPO-based temporal memory. A Sobolev-regularized formulation motivates the structural filtering, while gradient attribution and LLM verbalization provide explanations. Experiments on MIMIC-III and MIMIC-IV report strong results across several prediction tasks, supported by architectural and hypergraph-construction ablations.

Critical strengths
  1. The framework addresses a meaningful combination of challenges: intra-visit clinical interactions and dependencies across visits. Constructing a reusable global hypergraph also reduces repeated LLM calls.

  2. The evaluation covers multiple clinical tasks and includes comparisons with non-LLM hypergraph construction. These experiments support the usefulness of the architecture beyond a particular knowledge-generation method.

Critical weaknesses
  1. The theoretical-to-implementation connection is incomplete. Equation (3) defines memory as (S_k=R_{H,k}C_k), but Equation (5) propagates the previous state without an explicit transformation when (R_{H,k}) changes. The visit-level pooling in Equation (9) further departs from the node-wise derivation. This makes the scope of the claimed Sobolev-regularized memory formulation unclear.

  2. The evidence for the temporal component is limited by the comparisons. The long-history ablation uses LSTM and non-temporal aggregation, while existing SSM-based EHR methods discussed in the paper are not evaluated directly. This limits attribution of the gains to the proposed temporal design.

Questions for authors

When topology changes, is a state transformation applied to preserve (S_k=R_{H,k}C_k)? How does this interpretation extend to Equation (9)?

6 comments
AuthorsOpenReview ↗

On memory update at topology change

We thank the reviewer for requesting this clarification. Theorem 3.1 characterizes instantaneous Sobolev-optimal coefficients; Eqs. (5) and (9) define causal memory of topology-filtered visits. Their state conventions differ at topology changes.

1. Topology switches and Eq. (5). At visit kk, Theorem 3.1 gives the exact optimizer of Eq. (2):

Skinst=RkCk,Rk=[I+λ(LˉH,k+ϵI)β]−1.S_k^{\mathrm{inst}}=R_kC_k,\qquad R_k=\left[I+\lambda(\bar L_{H,k}+\epsilon I)^\beta\right]^{-1}.

On aligned nodes, maintaining this identity across a switch with continuous CC requires

Sinst(tk−1+)=RkRk−1−1Sinst(tk−1−).S^{\mathrm{inst}}(t_{k-1}^{+})=R_kR_{k-1}^{-1}S^{\mathrm{inst}}(t_{k-1}^{-}).

SoAR-SSM instead carries accumulated memory forward unchanged and uses the current topology only to filter new information. Under zero-order hold, R(t)=RkR(t)=R_k and X(t)=XkX(t)=X_k on (tk−1,tk](t_{k-1},t_k], with

S˙(t)=S(t)A⊤+RkXkB⊤,S(tk−1+)=S(tk−1−).\dot S(t)=S(t)A^\top+R_kX_kB^\top,\qquad S(t_{k-1}^{+})=S(t_{k-1}^{-}).

Exact integration yields Eq. (5):

Sk=Sk−1Aˉk+RkXkB⊤Gk,S_k=S_{k-1}\bar A_k+R_kX_kB^\top G_k,

where Aˉk=eΔkA⊤\bar A_k=e^{\Delta_kA^\top} and Gk=∫0ΔkeuA⊤,duG_k=\int_0^{\Delta_k}e^{uA^\top}\\,du.

This matches S=RkCS=R_kC dynamics within constant-topology intervals but preserves memory across switches. Thus Eq. (5) is an exact ZOH discretization of the specified filtered-input dynamics, not a claim that Sk=RkCkS_k=R_kC_k continues to hold under arbitrary topology changes. This also follows from the diffuse-then-update recurrence structure of (Li et., al., 2024, ref line no:599), however adapted to our modelling strategy. Additionally, we also provide a variational characterization supporting the diffusion then update.

2. Sobolev/HiPPO variational meaning. Let Qk=(LˉH,k+ϵI)β⪰0Q_k=(\bar L_{H,k}+\epsilon I)^\beta\succeq0. Then each visit has the exact spatial Sobolev optimizer as

Zk⋆=arg⁡min⁡Z[12∥Z−Xk∥F2+λ2Tr⁡(Z⊤QkZ)]=RkXk.Z_k^{\star}=\arg\min_Z\left[\tfrac12\Vert Z-X_k\Vert_F^2+\tfrac{\lambda}{2}\operatorname{Tr}(Z^\top Q_kZ)\right]=R_kX_k.

The compressed memory follows zq(τ)=R(τ)xq(τ)z_q(\tau)=R(\tau)x_q(\tau), where R(τ)R(\tau) is the topology at observation time. For an orthonormal HiPPO basis ptKp_{t_K} and its measure μtK\mu_{t_K},

Mq,K⋆=arg⁡min⁡M12∫0tK∥zq(τ)−MptK(τ)∥22,dμtK(τ),M_{q,K}^{\star}=\arg\min_M\tfrac12\int_0^{t_K}\Vert z_q(\tau)-Mp_{t_K}(\tau)\Vert_2^2\\,d\mu_{t_K}(\tau),

which, by orthonormality, equals Mq,K⋆=∫0tKzq(τ)ptK(τ)⊤,dμtK(τ).M_{q,K}^{\star}=\int_0^{t_K}z_q(\tau)p_{t_K}(\tau)^\top\\,d\mu_{t_K}(\tau).

A matching HiPPO state-space realization has the filtered-input structure of Eq. (5). Hence, in the idealized linear setting, the two-stage interpretation is Sobolev-regularize each visit, then temporally compress it. Theorem 3.1 asserts the latter's instantaneous optimality, not equality of the two histories after topology changes.

3. Exact discrepancy characterization. On a common aligned node space with matching transitions, zero initialization and Uk=XkB⊤GkU_k=X_kB^\top G_k, unrolling gives

SK−RKCK=∑k=1K(Rk−RK)Uk(∏j=k+1KAˉj),S_K-R_KC_K=\sum_{k=1}^K(R_k-R_K)U_k\left(\prod_{j=k+1}^K\bar A_j\right),

and the resolvent identity gives

Rk−RK=λRk(QK−Qk)RK.R_k-R_K=\lambda R_k(Q_K-Q_k)R_K.

Since Qk⪰0Q_k\succeq0 implies ∥Rk∥2≤1\Vert R_k\Vert_2\leq1,

∥SK−RKCK∥F≤λ∑k=1K∥QK−Qk∥2,∥Uk∥F,∥∏j=k+1KAˉj∥2.\Vert S_K-R_KC_K\Vert_F\leq\lambda\sum_{k=1}^K\Vert Q_K-Q_k\Vert_2\\,\Vert U_k\Vert_F\\,\left\Vert\prod_{j=k+1}^K\bar A_j\right\Vert_2.

Thus the two states coincide for static topology; otherwise their difference is controlled by operator variation, input strength and temporal propagation.

4. Eq. (9) and actual implementation. Each visit is processed independently in its own hypergraph:

Z~k=fθ(Lk,Xk),uk=ev,k⊤Z~k∈R1×d,\widetilde Z_k=f_\theta(L_k,X_k),\qquad u_k=e_{v,k}^\top\widetilde Z_k\in\mathbb R^{1\times d}, where fθf_\theta is the ARMA encoder and

sk=sk−1Aˉk+ukB⊤Gk,s_k=s_{k-1}\bar A_k+u_kB^\top G_k, The model stores one fixed-dimensional patient state, not persistent states for individual hypergraph nodes. Variable node identities/counts do not require cross-topology basis transformations: all topology-specific reasoning occurs before the visit embedding enters patient memory. With aligned nodes, exact RkR_k and a fixed selector eve_v, Eq. (9) is the left projection sk=ev⊤Sks_k=e_v^\top S_k of Eq. (5); with changing node sets/selectors it directly defines a fixed-dimensional patient model rather than a lossless projection of node-wise memory. As mentioned, maintaining per node wise statistics is difficult due to changing concept sets. The design is motivated from downstream tasks where all our tasks are patient-level and would require a pooling step. As Linear pooling commutes with propagation, given the readout is fixed, performing pooling through a fixed visit coordinate is lossless (detailed derivation below) with respect to the propagated dynamics.

We made this distinction clear in Secs. 3.2, 3.4 of the revised draft. To summarize, Theorem 3.1 gives exact instantaneous spatial regularization; Eq. (5) gives exact linear ZOH carried-memory evolution; Eq. (9) implements the corresponding causal diffuse-then-update principle with shared patient-level memory.

AuthorsOpenReview ↗

Visit node characterization

Visit-level SSM is an exact row restriction. Eq. (8) selects a designated visit node, not an average pool. On the aligned node space of Sec. 3.2, with fixed selector eve_v, left-multiplication of Eq. (5) gives

ev⊤Sk=(ev⊤Sk−1)Aˉk+(ev⊤RH,kXk)B⊤Gk.e_v^{\top}S_k=(e_v^{\top}S_{k-1})\bar A_k+(e_v^{\top}R_{H,k}X_k)B^{\top}G_k.

Since Aˉk\bar A_k acts on state dimensions only, rows do not mix. Thus sk⋆=ev⊤Sks_k^{\star}=e_v^{\top}S_k and zk⋆=ev⊤RH,kXkWz_k^{\star}=e_v^{\top}R_{H,k}X_k W obey exactly Eq. (9)'s recurrence, with the implementation replacing zk⋆z_k^{\star} by z~k\tilde z_k. With matched SSM parameters/initialization and δk=∥z~k−zk⋆∥2\delta_k=\lVert\tilde z_k-z_k^{\star}\rVert_2,

∥s~K−sK⋆∥2≤∑k=1Kδk ∥B⊤Gk∥2∏j=k+1K∥Aˉj∥2.\lVert\tilde s_K-s_K^{\star}\rVert_2\le\sum\nolimits_{k=1}^{K}\delta_k\,\lVert B^{\top}G_k\rVert_2\prod\nolimits_{j=k+1}^{K}\lVert\bar A_j\rVert_2.

Hence visit selection adds no temporal approximation, and spatial error has bounded temporal propagation.

Properties relaxed. The unrestricted encoder does not enforce the resolvent’s exact monotone response. Relaxing these restrictions enables more flexible feature-dependent representations.

Why the replacement is justified. As discussed in the paper Using RH,k\mathbf{R}_{H,k} directly has three costs:

  • It requires a dense O(N3)O(N^3) inversion per visit, which destroys sparsity.
  • It fixes (λ,β,ϵ)(\lambda,\beta,\epsilon), and hence the spectral emphasis, before training.

It costs O(L(Sd+Nd2))O(L(Sd+Nd^2)) with sparse propagation (Appendix D.4), contains the Sobolev filter exactly. Rational filters are also the natural family here: polynomial filters can only approximate a resolvent, whereas ARMA represents it exactly.

Beyond the exact correspondence established for (β=1)(\beta=1) in Appendix C, we propose evaluating operator approximation, structural robustness, and temporal stability jointly.

Experimental protocol. For each held-out visit hypergraph, we introduce controlled structural perturbations through degree-preserving incidence rewiring. Specifically, we randomly select two existing node–hyperedge memberships (u,ea)(u,e_a) and (v,eb)(v,e_b) and exchange them to (u,eb)(u,e_b) and (v,ea)(v,e_a), provided the new memberships do not already exist. This preserves node degrees, hyperedge cardinalities, clinical entity features, and visit-node connectivity. We recompute the normalized hypergraph operator and quantify perturbation magnitude as δ=∥LH′−LH∥2\delta=\lVert\mathbf{L}_ H'-\mathbf{L}_ H\rVert_2.

We evaluate three complementary properties:

  • Sobolev approximation: ESob=100 ∥ZARMA−R1XW∥F / ∥R1XW∥FE_{\mathrm{Sob}}=100\thinspace\lVert\mathbf{Z}_ {\mathrm{ARMA}}-\mathbf{R}_1\mathbf{X}\mathbf{W}\rVert_F\thinspace/\thinspace\lVert\mathbf{R}_1\mathbf{X}\mathbf{W}\rVert_F.
  • Structural sensitivity: Ksp=∥Z′−Z∥F / (δ ∥X∥F)K_{\mathrm{sp}}=\lVert\mathbf{Z}'-\mathbf{Z}\rVert_F\thinspace/\thinspace(\delta\thinspace\lVert\mathbf{X}\rVert_F).
  • Temporal drift: Dtemp=100 ∥sT′−sT∥2 / ∥sT∥2D_{\mathrm{temp}}=100\thinspace\lVert\mathbf{s}_ T'-\mathbf{s}_ T\rVert_2\thinspace/\thinspace\lVert\mathbf{s}_ T\rVert_2, after propagating perturbed visit representations.

Approximation and perturbation stability

ModelESobE_{Sob} (%) ↓KspK_{sp} ↓DtempD_{temp} (%) ↓
Exact Sobolev0.000.141.90
Linear ARMA (L=4)4.100.111.75
Linear ARMA (L=8)0.180.141.91
Nonlinear ARMA (F=4,L=4)7.020.192.42
Nonlinear ARMA (F=4,L=8)5.170.152.15

It can be observed that increasing recursion depth improves approximation to the analytical Sobolev operator. The nonlinear ARMA variants exhibit controlled structural sensitivity and temporal drift comparable to the analytical reference, consistent with Theorem 4.1 and Corollary 1. These are distinct from exact operator equivalence: finite-depth nonlinear ARMA provides a learnable generalization whose approximation capability and stability can be evaluated independently.

For completeness, detailed connection of β\beta to generalized ARMA filter is included in Appendix C

AuthorsOpenReview ↗

Long Range Results

We additionally report results comparing the proposed model against the Mamba variant on EHR data (EHR MAMBA), whereas graph based SSM variant is presented in Table 3. Specifically, we report results across 10 runs on long-range mortality prediction task on MIMIC-IV dataset. Specifically, it includes 59,262 patients, 964 have >10 visits, yielding 5,912 eligible prediction samples with 208 positive outcomes. The patient-disjoint train/validation/test splits contain 670/153/141 patients, 3,970/977/965 samples, and 144/36/28 positive outcomes, respectively.

Mean and standard deviation results across multiple seed runs.

ModelAUROC (%)Macro-F1 (%)
Hyper-GNN60.6 ± 0.362.8 ± 1.2
SoAR-LSTM65.1 ± 0.568.9 ± 0.9
HypEHR63.2 ± 0.767.1 ± 1.4
EHRMamba67.6 ± 1.372.3 ± 1.7
SoAR-SSM69.3 ± 1.276.1 ± 1.4

SoAR-SSM outperforms SoAR-LSTM by +4.2 AUROC/+7.2 Macro-F1 points and EHRMamba by +1.7 AUROC/+3.8 Macro-F1 points, while requiring substantially fewer parameters. The improvements are consistent with the reported run-to-run variability, with particularly clear gains in Macro-F1.

Overall, these controlled, independently tuned comparisons demonstrate that SoAR-SSM achieves superior long-range prediction performance with lower model capacity, strengthening the empirical evidence for effective and parameter-efficient temporal modeling.

To directly investigate reliance on historical information, we conduct a progressive history-truncation experiment on the MIMIC-IV long-history subset (>10 visits), retaining only the most recent (K∈{1,5,10}K\in\lbrace 1,5,10\rbrace) visits or the full available history.

ModelK=1K=5K=10Full history
SoAR-SSM (AUROC)0.5860.6030.6320.693
SoAR-LSTM (AUROC)0.5230.5740.5920.651

AUROC improves monotonically with increasing historical context, from 58.6% (K=1) to 69.3% (full history), an absolute improvement of 10.7 percentage points. Notably, retaining visits beyond the most recent 10 provides an additional 6.1-point improvement, indicating that earlier clinical information contributes predictive value beyond recent observations. These findings complement the SSM-versus-LSTM ablation in above table and provide direct empirical evidence supporting the model's utilization of long-range patient history.

Reviewer_iJSnOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SoAR-SSM for clinical prediction from longitudinal electronic health records, combining knowledge-infused visit-level hypergraphs, ARMA spatial encoders, and HiPPO-based temporal memory. It motivates the architecture through a Sobolev-regularized approximation objective and provides perturbation and permutation analyses, alongside gradient-based explanations verbalized by an LLM. Experiments on MIMIC-III and MIMIC-IV report improvements across mortality, length-of-stay, medication, phenotyping, and diagnosis prediction, with additional structural, temporal, and knowledge-source ablations.

Critical strengths

The architecture explicitly separates clinically grouped interactions within visits from temporal aggregation across visits, addressing a meaningful limitation of static patient representations.

Critical weaknesses
  1. The connection between the derived filter and the implemented model remains incomplete: Appendix C establishes an infinite-depth linear ARMA correspondence for (\beta=1), whereas Equation (6) uses finite-depth nonlinear transformations and Equation (9) propagates a pooled visit representation. Although the paper acknowledges this generalization, it does not establish which Sobolev properties survive or compare the learned encoder with the explicit resolvent, weakening the claimed theoretical basis of the implemented architecture.
  2. The claimed reasoning paths are sequences of highly salient entities rather than demonstrated relational or causal explanations, and their LLM verbalizations are not evaluated for faithfulness. For example, Figure 11 infers “stabilization of the septicemia” and a “narrowly averted” mortality event from entity lists that do not establish either conclusion.
Questions for authors
  1. At a topology change, is the stored memory transformed by (R_{H,k}R_{H,k-1}^{-1}) before propagation? If not, should Equation (5) be understood as a separately defined filtered-input SSM rather than the evolving optimizer of Equation (2)?
  2. Which properties of the Sobolev formulation are retained after replacing the resolvent with the nonlinear ARMA encoder and pooling to the visit node?
  3. For Table 4, how many patients and positive outcomes are included, and are the LSTM and SSM variants matched in capacity and tuning budget? Are repeated-run uncertainties available for the reported differences?
  4. Has explanation faithfulness been assessed through entity-removal tests or another existing evaluation? Does the verbalizing LLM receive information beyond the displayed entity trajectories that supports statements such as the stabilization described in Figure 11?
10 comments
AuthorsOpenReview ↗

On memory update at topology change

We thank the reviewer for requesting this clarification. Theorem 3.1 characterizes instantaneous Sobolev-optimal coefficients; Eqs. (5) and (9) define causal memory of topology-filtered visits. Their state conventions differ at topology changes.

1. Topology switches and Eq. (5). At visit kk, Theorem 3.1 gives the exact optimizer of Eq. (2):

Skinst=RkCk,Rk=[I+λ(LˉH,k+ϵI)β]−1.S_k^{\mathrm{inst}}=R_kC_k,\qquad R_k=\left[I+\lambda(\bar L_{H,k}+\epsilon I)^\beta\right]^{-1}.

On aligned nodes, maintaining this identity across a switch with continuous CC requires

Sinst(tk−1+)=RkRk−1−1Sinst(tk−1−).S^{\mathrm{inst}}(t_{k-1}^{+})=R_kR_{k-1}^{-1}S^{\mathrm{inst}}(t_{k-1}^{-}).

SoAR-SSM instead carries accumulated memory forward unchanged and uses the current topology only to filter new information. Under zero-order hold, R(t)=RkR(t)=R_k and X(t)=XkX(t)=X_k on (tk−1,tk](t_{k-1},t_k], with

S˙(t)=S(t)A⊤+RkXkB⊤,S(tk−1+)=S(tk−1−).\dot S(t)=S(t)A^\top+R_kX_kB^\top,\qquad S(t_{k-1}^{+})=S(t_{k-1}^{-}).

Exact integration yields Eq. (5):

Sk=Sk−1Aˉk+RkXkB⊤Gk,S_k=S_{k-1}\bar A_k+R_kX_kB^\top G_k,

where Aˉk=eΔkA⊤\bar A_k=e^{\Delta_kA^\top} and Gk=∫0ΔkeuA⊤,duG_k=\int_0^{\Delta_k}e^{uA^\top}\\,du.

This matches S=RkCS=R_kC dynamics within constant-topology intervals but preserves memory across switches. Thus Eq. (5) is an exact ZOH discretization of the specified filtered-input dynamics, not a claim that Sk=RkCkS_k=R_kC_k continues to hold under arbitrary topology changes. This also follows from the diffuse-then-update recurrence structure of (Li et., al., 2024, ref line no:599), however adapted to our modelling strategy. Additionally, we also provide a variational characterization supporting the diffusion then update.

2. Sobolev/HiPPO variational meaning. Let Qk=(LˉH,k+ϵI)β⪰0Q_k=(\bar L_{H,k}+\epsilon I)^\beta\succeq0. Then each visit has the exact spatial Sobolev optimizer as

Zk⋆=arg⁡min⁡Z[12∥Z−Xk∥F2+λ2Tr⁡(Z⊤QkZ)]=RkXk.Z_k^{\star}=\arg\min_Z\left[\tfrac12\Vert Z-X_k\Vert_F^2+\tfrac{\lambda}{2}\operatorname{Tr}(Z^\top Q_kZ)\right]=R_kX_k.

The compressed memory follows zq(τ)=R(τ)xq(τ)z_q(\tau)=R(\tau)x_q(\tau), where R(τ)R(\tau) is the topology at observation time. For an orthonormal HiPPO basis ptKp_{t_K} and its measure μtK\mu_{t_K},

Mq,K⋆=arg⁡min⁡M12∫0tK∥zq(τ)−MptK(τ)∥22,dμtK(τ),M_{q,K}^{\star}=\arg\min_M\tfrac12\int_0^{t_K}\Vert z_q(\tau)-Mp_{t_K}(\tau)\Vert_2^2\\,d\mu_{t_K}(\tau),

which, by orthonormality, equals Mq,K⋆=∫0tKzq(τ)ptK(τ)⊤,dμtK(τ).M_{q,K}^{\star}=\int_0^{t_K}z_q(\tau)p_{t_K}(\tau)^\top\\,d\mu_{t_K}(\tau).

A matching HiPPO state-space realization has the filtered-input structure of Eq. (5). Hence, in the idealized linear setting, the two-stage interpretation is Sobolev-regularize each visit, then temporally compress it. Theorem 3.1 asserts the latter's instantaneous optimality, not equality of the two histories after topology changes.

3. Exact discrepancy characterization. On a common aligned node space with matching transitions, zero initialization and Uk=XkB⊤GkU_k=X_kB^\top G_k, unrolling gives

SK−RKCK=∑k=1K(Rk−RK)Uk(∏j=k+1KAˉj),S_K-R_KC_K=\sum_{k=1}^K(R_k-R_K)U_k\left(\prod_{j=k+1}^K\bar A_j\right),

and the resolvent identity gives

Rk−RK=λRk(QK−Qk)RK.R_k-R_K=\lambda R_k(Q_K-Q_k)R_K.

Since Qk⪰0Q_k\succeq0 implies ∥Rk∥2≤1\Vert R_k\Vert_2\leq1,

∥SK−RKCK∥F≤λ∑k=1K∥QK−Qk∥2,∥Uk∥F,∥∏j=k+1KAˉj∥2.\Vert S_K-R_KC_K\Vert_F\leq\lambda\sum_{k=1}^K\Vert Q_K-Q_k\Vert_2\\,\Vert U_k\Vert_F\\,\left\Vert\prod_{j=k+1}^K\bar A_j\right\Vert_2.

Thus the two states coincide for static topology; otherwise their difference is controlled by operator variation, input strength and temporal propagation.

4. Eq. (9) and actual implementation. Each visit is processed independently in its own hypergraph:

Z~k=fθ(Lk,Xk),uk=ev,k⊤Z~k∈R1×d,\widetilde Z_k=f_\theta(L_k,X_k),\qquad u_k=e_{v,k}^\top\widetilde Z_k\in\mathbb R^{1\times d}, where fθf_\theta is the ARMA encoder and

sk=sk−1Aˉk+ukB⊤Gk,s_k=s_{k-1}\bar A_k+u_kB^\top G_k, The model stores one fixed-dimensional patient state, not persistent states for individual hypergraph nodes. Variable node identities/counts do not require cross-topology basis transformations: all topology-specific reasoning occurs before the visit embedding enters patient memory. With aligned nodes, exact RkR_k and a fixed selector eve_v, Eq. (9) is the left projection sk=ev⊤Sks_k=e_v^\top S_k of Eq. (5); with changing node sets/selectors it directly defines a fixed-dimensional patient model rather than a lossless projection of node-wise memory. As mentioned, maintaining per node wise statistics is difficult due to changing concept sets. The design is motivated from downstream tasks where all our tasks are patient-level and would require a pooling step. As Linear pooling commutes with propagation, given the readout is fixed, performing pooling through a fixed visit coordinate is lossless (detailed derivation below) with respect to the propagated dynamics.

We made this distinction clear in Secs. 3.2, 3.4 of the revised draft. To summarize, Theorem 3.1 gives exact instantaneous spatial regularization; Eq. (5) gives exact linear ZOH carried-memory evolution; Eq. (9) implements the corresponding causal diffuse-then-update principle with shared patient-level memory.

AuthorsOpenReview ↗

Generalization to

We thank the Reviewer for the comment. The manuscript provides the connection of Sobolev filter to ARMA-1 filter and remarks the connection to ARMA-K filters. However, for completeness we present the result for any integer β\beta. Let LˉH\bar{\mathbf{L}}_ H be a symmetric positive semidefinite hypergraph Laplacian, and define

RH,β=[I+λ(LˉH+ϵI)β]−1,λ>0,ϵ≥0.\mathbf{R}_ {H,\beta}=\left[\mathbf{I}+\lambda\left(\bar{\mathbf{L}}_ H+\epsilon\mathbf{I}\right)^\beta\right]^{-1},\qquad \lambda>0,\quad \epsilon\ge 0 .

Recall, for an eigenvalue μ\mu of LˉH\bar{\mathbf{L}}_ H, its spectral response is

hβ(μ)=11+λ(μ+ϵ)β.h_\beta(\mu)=\frac{1}{1+\lambda(\mu+\epsilon)^\beta}.

Now we express the relation of β\beta Sobolev filter to generalized ARMA filter.

Proposition C.1 (Integer-order Sobolev–ARMA decomposition). For every positive integer β\beta, the operator RH,β\mathbf{R}_ {H,\beta} admits an exact decomposition into K=βK=\beta first-order rational branches i.e,

 RH,β=∑j=1βbj(I−ajLH)−1 ,(C.1)\boxed{\ \mathbf{R}_ {H,\beta}=\sum_{j=1}^{\beta}b_j\left(\mathbf{I}-a_j\mathbf{L}_ H\right)^{-1}\ }, \qquad\text{(C.1)}

where LH=I−LˉH\mathbf{L}_ H=\mathbf{I}-\bar{\mathbf{L}}_ H, aj=−1cj,bj=rjβcja_j=-\frac{1}{c_j},\qquad b_j=\frac{r_j}{\beta c_j} with cj=rj−(1+ϵ)c_j=r_j-(1+\epsilon) and rj=λ−1/βexp⁡((2j−1)πiβ),∀j=1,…,β,r_j=\lambda^{-1/\beta}\exp\left(\frac{(2j-1)\pi i}{\beta}\right),\qquad \forall j=1,\ldots,\beta,

Proof. Consider the polynomial as p(y)=1+λyβp(y)=1+\lambda y^\beta. Observe the roots of p(y)p(y) i.e., r1,…,rβr_1,\ldots,r_\beta are distinct and nonzero, and p′(rj)=λβrjβ−1≠0p'(r_j)=\lambda\beta r_j^{\beta-1}\neq0, so every root is simple, as λrjβ=−1\lambda r_j^\beta=-1 and 1p′(rj)=−rjβ.\frac{1}{p'(r_j)}=-\frac{r_j}{\beta}.

The partial-fraction expansion gives

11+λyβ=∑j=1β1/p′(rj)y−rj=1β∑j=1βrjrj−y.(C.2)\frac{1}{1+\lambda y^\beta}=\sum_{j=1}^{\beta}\frac{1/p'(r_j)}{y-r_j}=\frac{1}{\beta}\sum_{j=1}^{\beta}\frac{r_j}{r_j-y}.\qquad\text{(C.2)}

Let M=LˉH+ϵI\mathbf{M}=\bar{\mathbf{L}}_ H+\epsilon\mathbf{I}. Each rjr_j is either nonreal or negative, so every rjI−Mr_j\mathbf{I}-\mathbf{M} is invertible. Applying (C.2) yields

RH,β=1β∑j=1βrj(rjI−M)−1.(C.3)\mathbf{R}_ {H,\beta}=\frac{1}{\beta}\sum_{j=1}^{\beta}r_j\left(r_j\mathbf{I}-\mathbf{M}\right)^{-1}.\qquad\text{(C.3)}

Since M=(1+ϵ)I−LH\mathbf{M}=(1+\epsilon)\mathbf{I}-\mathbf{L}_ H, we obtain rjI−M=cjI+LH=cj(I−ajLH),r_j\mathbf{I}-\mathbf{M}=c_j\mathbf{I}+\mathbf{L}_ H=c_j\left(\mathbf{I}-a_j\mathbf{L}_ H\right), and in particular I−ajLH\mathbf{I}-a_j\mathbf{L}_ H is invertible. Substituting into (C.3) gives 1βrj(rjI−M)−1=bj(I−ajLH)−1\frac{1}{\beta}r_j\left(r_j\mathbf{I}-\mathbf{M}\right)^{-1}=b_j\left(\mathbf{I}-a_j\mathbf{L}_ H\right)^{-1}, which proves (C.1).

For β≥2\beta\ge2 the roots occur in conjugate pairs (rβ+1−j=rˉjr_{\beta+1-j}=\bar r_j, hence aβ+1−j=aˉja_{\beta+1-j}=\bar a_j, bβ+1−j=bˉjb_{\beta+1-j}=\bar b_j), so each pair sums to a real second-order rational branch and, for odd β\beta, the real root −λ−1/β-\lambda^{-1/\beta} gives one real first-order branch; thus RH,β\mathbf{R}_ {H,\beta} is realizable with real weights.

Connection to the rational ARMA family. For integer β\beta,

hβ(μ)=11+λ∑m=0β(βm)ϵβ−mμm.h_\beta(\mu)=\frac{1}{1+\lambda\sum_{m=0}^{\beta}\binom{\beta}{m}\epsilon^{\beta-m}\mu^m}.

Dividing the numerator and denominator by 1+λϵβ1+\lambda\epsilon^\beta obtains the rational ARMA family.

Proposition C.2 (Recursive realization and truncation error). For an input X\mathbf{X}, initialize Zj(0)=0\mathbf{Z}_ j^{(0)}=\mathbf{0} consider

Zj(ℓ+1)=ajLHZj(ℓ)+bjX.(C.4)\mathbf{Z}_ j^{(\ell+1)}=a_j\mathbf{L}_ H\mathbf{Z}_ j^{(\ell)}+b_j\mathbf{X}.\qquad\text{(C.4)}

Further if γj=∥ajLH∥2<1\gamma_j=\lVert a_j\mathbf{L}_ H\rVert_2<1 for every jj, then

lim⁡L→∞∑j=1βZj(L)=RH,βX.\lim_{L\to\infty}\sum_{j=1}^{\beta}\mathbf{Z}_ j^{(L)}=\mathbf{R}_ {H,\beta}\mathbf{X}.

Whereas the error between the representation are bounded as

 ∥∑j=1βZj(L)−RH,βX∥F≤∑j=1β∣bj∣ γjL1−γj ∥X∥F. (C.5)\boxed{\ \left\lVert\sum_{j=1}^{\beta}\mathbf{Z}_ j^{(L)}-\mathbf{R}_ {H,\beta}\mathbf{X}\right\rVert_F\le\sum_{j=1}^{\beta}\frac{|b_j|\thinspace\gamma_j^{L}}{1-\gamma_j}\thinspace\lVert\mathbf{X}\rVert_F .\ }\qquad\text{(C.5)}

Proof. Unrolling (C.4) gives

Zj(L)=bj∑ℓ=0L−1(ajLH)ℓX.\mathbf{Z}_ j^{(L)}=b_j\sum_{\ell=0}^{L-1}\left(a_j\mathbf{L}_ H\right)^{\ell}\mathbf{X}.

Since γj<1\gamma_j<1, the Neumann series converges to Zj⋆=bj(I−ajLH)−1X\mathbf{Z}_ j^\star=b_j\left(\mathbf{I}-a_j\mathbf{L}_ H\right)^{-1}\mathbf{X}, and satisfies

∥Zj(L)−Zj⋆∥F≤∣bj∣∑ℓ=L∞γjℓ ∥X∥F=∣bj∣ γjL1−γj ∥X∥F.\left\lVert\mathbf{Z}_ j^{(L)}-\mathbf{Z}_ j^\star\right\rVert_F\le|b_j|\sum_{\ell=L}^{\infty}\gamma_j^{\ell}\thinspace\lVert\mathbf{X}\rVert_F=\frac{|b_j|\thinspace\gamma_j^{L}}{1-\gamma_j}\thinspace\lVert\mathbf{X}\rVert_F ..

Summing these bounds and applying Proposition C.1 proves the result.

Recovery of the first-order construction. For β=1\beta=1, r1=−1/λr_1=-1/\lambda, which gives

a1=λ1+λ(1+ϵ),b1=11+λ(1+ϵ).a_1=\frac{\lambda}{1+\lambda(1+\epsilon)},\qquad b_1=\frac{1}{1+\lambda(1+\epsilon)} .

Thus RH,1=b1(I−a1LH)−1\mathbf{R}_ {H,1}=b_1\left(\mathbf{I}-a_1\mathbf{L}_ H\right)^{-1}, recovering the original first-order Sobolev–ARMA connection. Here 0<a1<10<a_1<1, so the recursion converges for every λ>0\lambda>0.

For completeness we have included the proof in Appendix C of the revised draft

AuthorsOpenReview ↗

Retained properties after non-linear transformation

Learned nonlinear encoder: scope of guarantees. Eq. (6) (of main paper) extends the recursion with trainable matrices and activation σ\sigma. Thus the resolvent lies in the realizable limiting linear ARMA family, not necessarily in the response of an arbitrary nonlinear model. The implemented encoder nevertheless retains: (i) Generalized frequency response and permutation-equivariant node features and invariant visit-node readouts (Thm. 4.2); (ii) hypergraph-perturbation stability under the Lipschitz/weight assumptions of Thm. 4.1, with temporal accumulation in Cor. 1; and (iii) bounded input sensitivity. Specifically, for fixed shift, CC-Lipschitz σ\sigma, ∥Wr∥2≤Θ\|W_r\|_2\le\Theta, ∥Wi∥2≤Ω\|W_i\|_2\le\Omega, and ∥LH∥2≤ρ\|L_H\|_2\le\rho, unrolling Eq. (6) yields

∥Z~(X)−Z~(X′)∥F≤κ∥X−X′∥F,κ=CΩ∑ℓ=0L−1(CρΘ)ℓ.\lVert\tilde Z(X)-\tilde Z(X')\rVert_F\le\kappa\lVert X-X'\rVert_F,\quad \kappa=C\Omega\sum\nolimits_{\ell=0}^{L-1}(C\rho\Theta)^{\ell}.

This analog of bounded Sobolev gain is nonexpansive if κ≤1\kappa\le1. The tied linear subfamily also has rational spectral response H(ν)=Wi(I−νWr)−1H(\nu)=W_i(I-\nu W_r)^{-1}.

Visit-level SSM is an exact row restriction. Eq. (8) selects a designated visit node, not an average pool. On the aligned node space of Sec. 3.2, with fixed selector eve_v, left-multiplication of Eq. (5) gives

ev⊤Sk=(ev⊤Sk−1)Aˉk+(ev⊤RH,kXk)B⊤Gk.e_v^{\top}S_k=(e_v^{\top}S_{k-1})\bar A_k+(e_v^{\top}R_{H,k}X_k)B^{\top}G_k.

Since Aˉk\bar A_k acts on state dimensions only, rows do not mix. Thus sk⋆=ev⊤Sks_k^{\star}=e_v^{\top}S_k and zk⋆=ev⊤RH,kXkz_k^{\star}=e_v^{\top}R_{H,k}X_kW obey exactly Eq. (9)'s recurrence, with the implementation replacing zk⋆z_k^{\star} by z~k\tilde z_k. With matched SSM parameters/initialization and δk=∥z~k−zk⋆∥2\delta_k=\lVert\tilde z_k-z_k^{\star}\rVert_2,

∥s~K−sK⋆∥2≤∑k=1Kδk ∥B⊤Gk∥2∏j=k+1K∥Aˉj∥2.\lVert\tilde s_K-s_K^{\star}\rVert_2\le\sum\nolimits_{k=1}^{K}\delta_k\,\lVert B^{\top}G_k\rVert_2\prod\nolimits_{j=k+1}^{K}\lVert\bar A_j\rVert_2.

Hence visit selection adds no temporal approximation, and spatial error has bounded temporal propagation.

Properties relaxed. The unrestricted encoder does not enforce the resolvent’s exact monotone response. Relaxing these restrictions enables more flexible feature-dependent representations.

Why the replacement is justified. As discussed in the paper using RH,k\mathbf{R}_{H,k} directly has three costs:

  • It requires a dense O(N3)O(N^3) inversion per visit, which destroys sparsity.
  • It fixes (λ,β,ϵ)(\lambda,\beta,\epsilon), and hence the spectral emphasis, before training.

It costs O(L(Sd+Nd2))O(L(Sd+Nd^2)) with sparse propagation (Appendix D.4), contains the Sobolev filter exactly. Rational filters are also the natural family here: polynomial filters can only approximate a resolvent, whereas ARMA represents it exactly.

Beyond the exact correspondence established for β\beta in Appendix C, we propose evaluating operator approximation, structural robustness, and temporal stability jointly.

Experimental protocol. For each held-out visit hypergraph, we introduce controlled structural perturbations through degree-preserving incidence rewiring. Specifically, we randomly select two existing node–hyperedge memberships (u,ea)(u,e_a) and (v,eb)(v,e_b) and exchange them to (u,eb)(u,e_b) and (v,ea)(v,e_a), provided the new memberships do not already exist. This preserves node degrees, hyperedge cardinalities, clinical entity features, and visit-node connectivity. We recompute the normalized hypergraph operator and quantify perturbation magnitude as δ=∥LH′−LH∥2\delta=\lVert\mathbf{L}_ H'-\mathbf{L}_ H\rVert_2.

We evaluate three complementary properties:

  • Sobolev approximation: ESob=100 ∥ZARMA−R1XW∥F / ∥R1XW∥FE_{\mathrm{Sob}}=100\thinspace\lVert\mathbf{Z}_ {\mathrm{ARMA}}-\mathbf{R}_1\mathbf{X}\mathbf{W}\rVert_F\thinspace/\thinspace\lVert\mathbf{R}_1\mathbf{X}\mathbf{W}\rVert_F.
  • Structural sensitivity: Ksp=∥Z′−Z∥F / (δ ∥X∥F)K_{\mathrm{sp}}=\lVert\mathbf{Z}'-\mathbf{Z}\rVert_F\thinspace/\thinspace(\delta\thinspace\lVert\mathbf{X}\rVert_F).
  • Temporal drift: Dtemp=100 ∥sT′−sT∥2 / ∥sT∥2D_{\mathrm{temp}}=100\thinspace\lVert\mathbf{s}_ T'-\mathbf{s}_ T\rVert_2\thinspace/\thinspace\lVert\mathbf{s}_ T\rVert_2, after propagating perturbed visit representations.

Approximation and perturbation stability

ModelESobE_{Sob} (%) ↓KspK_{sp} ↓DtempD_{temp} (%) ↓
Exact Sobolev0.000.141.90
Linear ARMA (L=4)4.100.111.75
Linear ARMA (L=8)0.180.141.91
Nonlinear ARMA (F=4,L=4)7.020.192.42
Nonlinear ARMA (F=4,L=8)5.170.152.15

It can be observed that increasing recursion depth improves approximation to the analytical Sobolev operator. The nonlinear ARMA variants exhibit controlled structural sensitivity and temporal drift comparable to the analytical reference, consistent with Theorem 4.1 and Corollary 1. These are distinct from exact operator equivalence: finite-depth nonlinear ARMA provides a learnable generalization whose approximation capability and stability can be evaluated independently.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Summary of Reviewer Concerns and Rebuttal Responses

Dear Area Chair,

We sincerely thank you for handling our manuscript and overseeing the review process. We have submitted a detailed, point-by-point rebuttal addressing the concerns raised by all three reviewers. Below, we summarize the principal concerns and the corresponding clarifications, theoretical extensions, and additional experimental evidence provided in our response.

Overall, the three reviewers raised largely overlapping concerns regarding (i) state updates at temporal boundaries, (ii) the generalization of the Sobolev–ARMA connection beyond β=1\beta=1, and (iii) the characterization and evaluation of long-range datasets. We have carefully addressed each concern and incorporated the relevant clarifications and extensions into the revised manuscript.

1. State updates at temporal boundaries: The original manuscript introduced the diffuse-then-update principle for handling state transitions at temporal boundaries. However, recognizing that this terminology may have led to ambiguity, we have provided a more explicit mathematical characterization of the underlying mechanism, along with an alternative formulation that clarifies its operation. We have also explained how the visit-node-based implementation preserves the propagation of temporal information without introducing additional information loss at the boundaries. These clarifications have been incorporated into the revised manuscript. Importantly, this concern primarily pertains to the exposition of the proposed mechanism rather than a limitation of the underlying formulation.

2. Generalization of the Sobolev–ARMA connection: For analytical simplicity, the original manuscript established the connection between Sobolev regularization and ARMA filtering for β=1\beta=1, while noting its extension to general β\beta. To address the reviewers' concerns and for completeness due to availability of space, we have now provided a formal derivation connecting the general Sobolev formulation to generalized ARMA filters. Crucially, this extension preserves the validity of the theoretical inferences and generalization bounds, leaving the principal conclusions of the manuscript unchanged.

3. Long-range datasets and experimental validation: While the reviewers' concerns primarily focused on the reporting of cohort statistics and the characterization of long-range dependencies in the evaluated datasets, we have gone beyond these requests to further strengthen the empirical evaluation. Specifically, we have included additional comparisons against SSM-based baselines, a perturbation analysis examining the robustness of the proposed formulation, and further studies comparing different filtering mechanisms and relaxation strategies. These supplementary experiments provide additional evidence supporting the effectiveness, robustness, and theoretical motivation of the proposed approach.

In summary, we believe that the reviewers' principal concerns are mostly the clarifications and have been comprehensively addressed through clearer mathematical exposition, a more general theoretical characterization, and substantially strengthened experimental validation. We hope that these clarifications and new results will help resolve the reviewers' concerns and facilitate a favorable assessment of our manuscript.

We sincerely appreciate your time and consideration. We would also be grateful for your assistance in encouraging the reviewers to actively participate in the discussion phase. Given that we have provided detailed clarifications, theoretical extensions, and additional experimental evidence addressing their concerns, we believe that further engagement would facilitate a constructive exchange and allow the reviewers to reassess their evaluations in light of our responses.

AuthorsOpenReview ↗

Summary of Author-Reviewer Discussion

Dear Area Chair,

Thank you again for handling our manuscript. We would like to briefly summarize the reviewer feedback and the outcome of the discussion phase.

As noted in our earlier messages, all the three reviewers raised essentially the same concerns, which was primarily a request for clarification regarding the update at the boundaries and generalization of Sobolev filters.

In our response, we addressed this concerns with a detailed mathematical clarification, including a variational characterization that supports the claims made in the original manuscript. All the reviewers engaged in the discussion and confirmed that the additional explanation fully resolved their concerns. Importantly, reviewers also expressed appreciation for the alternative characterization, and all the reviewers have raised their scores, and indicated support for acceptance.

Thank you for your time again.

Sorry for my delayed reply. Thanks for your detailed reponse. My major concerns have been alleviated, and I raised my recommendation to a weak accept.

AuthorsOpenReview ↗

Thanks for raising the score

We thank the Reviewer for the response. We are happy to hear that all the concerns has been resolved. Thank you again for raising the score and for the feedback

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.