acceptodds
Under review as a conference paper at ICLR 2027

Hierarchical Relational Trajectory Modeling for Longitudinal Clinical Prognosis

Abstract

Prognosis from longitudinal electronic health records requires integrating evidence that is generated across hospitalizations and days, yet most existing models do not align their modeling of such records with this process, so a prediction cannot be traced back to the evidence it rests on, and whether a model actually relies on the trajectory cannot be checked. To address this mismatch, we propose **HiReT**, a hierarchical relational trajectory model that not only reconstructs patient records into a `visit-day-node` hierarchy, but also explicitly models the dynamic dependencies among heterogeneous evidence. **HiReT** expresses the dependencies among evidence types with a typed relation map within each visit, and uses hierarchical aggregation to follow the clinical data flow, combining days into visits and visits into a patient representation. Experiments on a natural-distribution six-class mortality benchmark from MIMIC-IV show that **HiReT** achieves relative improvements of 19.95%, 27.81%, and 47.39% in Balanced Accuracy, Macro-F1, and quadratic weighted kappa (QWK), respectively, over the strongest baseline, and is the best performer on all five metrics of a prolonged length-of-stay (LOS) benchmark. Furthermore, ablations over the model's structure and evidence types demonstrate that the model's predictions are genuinely grounded in specific clinical evidence. Overall, these results indicate that aligning a model with how clinical evidence is generated improves prognosis accuracy and makes the basis of its predictions checkable.

Primary area:
learning on time series and dynamical systems
Keywords:
electronic health records, clinical prognosis, longitudinal modeling, hierarchical representation learning, relational modeling

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_n3QQOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HiReT for longitudinal clinical prognosis using EHR data. It combines a visit-day-node hierarchy, typed relational message passing, and hierarchical GRUs. Experiments on MIMIC-IV show improvements in six-class mortality and prolonged length-of-stay prediction, particularly on class-balanced and ordinal metrics.

Critical strengths

The paper addresses a useful clinical prediction problem and reports promising results. The experiments include established baselines, a same-encoder flat control, component ablations, and patient-level bootstrap analysis, which provide some support for the proposed approach.

Critical weaknesses
  1. The methodological novelty is limited. MedHMP [1] already explores hierarchical multimodal EHR modeling, while TRANS [2] integrates temporal dependencies and heterogeneous graphs. These closely related methods are missing from the experimental comparisons. Removing Node Interaction reduces Balanced Accuracy from 27.00 to 26.42 and QWK from 43.85 to 43.32 (Table 3). The Typed Node Encoding ablation also removes multiple components simultaneously, making their individual contributions unclear.

  2. There are concerning inconsistencies in the reported results. Table 9 reports 94.57% survival recall, whereas Table 17 reports survival specificity of 92.62% for females and 90.58% for males. Under the stated definitions and test split, the overall recall cannot exceed both subgroup values. Table 16 reports the seed-42 HiReT results from Table 8 while labeling the row as a three-seed mean. These discrepancies raise concerns about the reliability of the evaluation.

  3. Some implementation details remain unclear. The construction of daily typed nodes does not sufficiently explain how multiple measurements, medications, and reports are aggregated. Baseline tokenization and context-length handling also differ without adequate evidence-coverage details. Furthermore, in Eq. (8), the directly concatenated background and availability terms appear to cancel during masked-softmax normalization under the shared linear DayFuse scorer, making their intended contribution unclear.

[Related Works]

[1] Hierarchical Pretraining on Multimodal Electronic Health Records. Xiaochen Wang, Junyu Luo, Jiaqi Wang, Ziyi Yin, Suhan Cui, Yuan Zhong, Yaqing Wang, and Fenglong Ma. EMNLP 2023.

[2] Predictive Modeling with Temporal Graphical Representation on Electronic Health Records. Jiayuan Chen, Changchang Yin, Yuanlong Wang, and Ping Zhang. IJCAI 2024.

Questions for authors
  1. What is the main methodological contribution beyond MedHMP and TRANS? Could the authors provide a closer comparison?

  2. How can the survival metrics in Tables 9 and 17 be reconciled? Why does Table 16 report seed-42 HiReT results as a three-seed mean?

  3. How are multiple events aggregated into daily typed nodes, and how is comparable evidence coverage ensured across baselines?

  4. Does the DayFuse implementation follow Eq. (8)? If so, how do the background and availability terms affect the normalized fusion weights?

Reviewer_1dZ1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces HiReT, a hierarchical relational model for predicting clinical outcomes from longitudinal electronic health records. Instead of flattening a patient’s history into one long sequence, HiReT organizes it into visits, days within each visit, and five evidence types: background information, measurements, treatments, careflow, and reports. Each type is encoded separately, contextualized over time, passed through a directed relation map, fused into daily representations, and then aggregated across days and visits. The authors evaluate the model on two patient-disjoint MIMIC-IV tasks: six-class mortality-window prediction and binary prediction of prolonged length of stay. HiReT improves Balanced Accuracy, Macro-F1, and QWK over the strongest reported mortality baseline, and it performs best on all five reported LOS metrics. The paper also includes component ablations, evidence-removal experiments, pathway interventions, and trajectory perturbations intended to show that the model uses both typed clinical evidence and the order of the patient’s history

Critical strengths
  1. The core modeling idea makes clinical sense. EHRs are naturally organized around hospitalizations, days, and different kinds of evidence. HiReT preserves that structure instead of forcing everything into a single mixed sequence. The architecture also follows the representation closely: type-specific encoders handle each evidence stream, the relation map lets evidence types interact, and the day- and visit-level GRUs preserve temporal order.

  2. The cohort construction is more careful than what is often seen in this area. Patients do not cross between the training and test partitions, even when they contribute several admissions. The authors also remove fields that are only known after the prediction landmark. For the mortality task, a missing death record is not automatically treated as survival; the patient must have enough follow-up to establish survival through one year.

  3. The same-encoder flat baseline is a useful control. Both models use the same cached Qwen3-Embedding-0.6B features and the same underlying examples. That makes the comparison much more informative than a table containing only older models with different encoders and preprocessing pipelines. The patient-clustered bootstrap analysis also shows clear gains over this control in Balanced Accuracy, Macro-F1, Macro-AUROC, and QWK.

  4. The evaluation is unusually broad. The authors do not stop at one main-results table. They report class-wise recall, per-seed results, architectural ablations, evidence-type removal, encoder controls, transfer to the LOS task, relation-path interventions, perturbation tests, and a subgroup audit. That gives the reader several ways to inspect what the model is doing.

Critical weaknesses
  1. My biggest concern is that the reported test set appears to have been used for model selection. The paper describes only Train and Test partitions. Appendix A.3 then says that several propagation depths and readout strategies were compared before the final configuration was chosen, while also stating that all reported results come from Test-set evaluation. The selected seed-42 configuration in Table 8 nearly matches the seed-42 Test result in Table 10. If those are the same evaluations, the final test numbers are no longer unbiased estimates of generalization. The clean fix is to create patient-disjoint training, validation, and test sets. Architecture choices, loss weights, stopping rules, and baseline tuning should happen on validation data. The final test set should be evaluated only after the full protocol has been locked.

  2. The baseline setup is not described well enough to judge whether the comparison is fair. The paper says the trainable baselines receive the same examples and evidence, but it does not provide complete architecture-specific hyperparameters, pretrained initialization details, trainable parameter counts, or tuning ranges. It is also unclear whether compatible baselines use HiReT’s ordinal-aware objective or ordinary cross-entropy. That distinction matters because QWK is one of the headline metrics. Prompted models are limited to 8,192 tokens, while HiReT appears to process the full available trajectory, yet the paper does not report how much evidence is truncated for those models.

  3. The ablations do not cleanly isolate the relational contribution. The “w/o Typed Node Encoding” variant removes several things at once: type-specific encoding, node interaction, and DayFuse. A large drop in that row cannot be attributed to typed encoding alone. By contrast, removing Node Interaction by itself changes Balanced Accuracy by only 0.58 points and QWK by 0.53 points. The test-time removal of dynamic-node interaction also produces a fairly small QWK drop of 0.60 points. Those results suggest that the hierarchy and typed representations may be doing more of the work than the relation map. A factorial ablation would be much clearer: shared versus type-specific projectors, simple typed pooling versus DayFuse, and DayFuse with versus without relation propagation

  4. The novelty claim feels broader than the evidence supports. Earlier work has already used hierarchical EHR representations, heterogeneous graphs, and relation-aware within-visit aggregation. MiME, TRANS, and HieraMed are especially relevant comparisons. HiReT’s more distinctive pieces seem to be the explicit day level, the background-conditioned cross-type relation map, and the particular mortality-window benchmark—not the general idea of representing EHRs hierarchically and relationally.

  5. The ordinal loss assumes a class geometry that is not obviously justified. The six outcomes are represented by consecutive class indices, so the loss treats neighboring classes as equally spaced. Clinically, they are not. The post-discharge windows cover 30, 60, 90, and 185 days, while in-hospital death has variable duration and uses a different temporal reference point. A two-class error therefore does not correspond to the same amount of time everywhere in the label space. Calibration and Brier scores would make the evaluation more clinically meaningful.

  6. The mortality cohort is highly selected. Every index admission must have at least one earlier hospitalization. Patients also need enough follow-up to establish a one-year outcome, which may favor people with more recorded healthcare contact. The paper acknowledges this, but the practical effect is not quantified. On top of that, one patient may contribute several examples, so frequent hospital users receive more weight than patients with fewer admissions.

  7. The interpretability claims are not supported. The ablations and perturbations show that the model is sensitive to certain structures and evidence types. They do not prove that an individual prediction can be faithfully traced back to specific clinical evidence, and they do not establish causal importance. Some interventions, such as replacing evidence with cross-patient noise, also push the input outside the training distribution. Mean relation weights are difficult to interpret because source availability is uneven; reports appear on only about 5% of days, so low average report weights are partly built into the data.

Questions for authors
  1. Can you clarify whether the reported Test set influenced architecture or hyperparameter selection? Appendix A.3 says that several propagation depths and readout strategies were evaluated before the final model was chosen, and it also says that all reported results come from Test-set evaluation. Were Table 8 or any other test results used to choose the architecture, loss weights, number of epochs, or checkpoint? If so, can you rerun selection on a patient-disjoint validation set and reserve the test set for one final evaluation?

  2. How were the baselines tuned and matched to HiReT? Please report the hyperparameter search, pretrained initialization, trainable parameter count, training objective, and available context for each baseline. In particular, did compatible mortality baselines use the same ordinal loss, and what fraction of each trajectory was truncated for models limited to 8,192 tokens?

  3. How much of the improvement actually comes from the relation map? The “w/o Typed Node Encoding” ablation removes several components at once, while removing Node Interaction alone has a relatively small effect. Could you provide parameter-matched factorial ablations that separate type-specific encoding, DayFuse, temporal hierarchy, and relation propagation?

  4. How dependent are the results on the selected cohort and on MIMIC-IV specifically? The mortality task requires a prior hospitalization and enough follow-up to establish a one-year outcome, and each patient may contribute several examples. Could you report the exclusion rate, test one-example-per-patient or patient-weighted evaluation, and either add temporal/external validation or narrow the target population stated in the paper?

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

This paper studies clinical prognosis from longitudinal electronic health records and argues that existing models are not aligned with how clinical evidence is generated. The authors propose HiReT, which rebuilds each patient record into a visit, day, and evidence-node hierarchy, encodes the nodes with a frozen text embedding model, exchanges information among evidence types through a typed relation map, and aggregates days into visits and visits into a patient representation. Structural and evidence-type ablations and test-time perturbations are used to probe which evidence the model relies on. Experimental results show that HiReT outperforms selected baselines on the class-balanced and ordinal metrics of a six-class mortality benchmark and on all five metrics of a prolonged length-of-stay benchmark, both built from MIMIC-IV.

Critical strengths
  1. A careful prediction setup. Evidence after the landmark is removed and all examples of a patient stay in one partition, which addresses the usual leakage routes in longitudinal EHR prediction.

  2. Transparent reporting around the main comparison. A same-encoder control, per-seed results, class-wise recall, a patient-level bootstrap, and a subgroup audit let the reader see where the gains come from and where they stop.

Critical weaknesses
  1. The experiments do not support the motivation. Tracing a prediction to its evidence is never tested at the prediction level, the paper's own audit runs on a baseline it says cannot be checked, and the relation map and day level barely move the primary results.

  2. The main comparison is not sound. The configuration is chosen on the Test set, the so-called natural benchmark drops survivors without later hospital contact, the gains barely show in ranking metrics, the hierarchy is confounded with the text encoder, and recent baselines are missing.

Questions for authors
  1. In most existing models, the abstract says, a prediction "cannot be traced back to the evidence" and reliance on the trajectory "cannot be checked" (lines 012 to 015). HiReT's evidence is population-level, and Table 4(B) runs the same model-agnostic check on Hi-BEHRT (reversal costs 8.08 QWK points against 5.57). Which experiment traces one prediction to its evidence, and what can be checked only in HiReT?

  2. Line 411 names typed organization "the principal source" of the gains, yet the untyped hierarchy gains 3.38 Balanced Accuracy and 11.64 QWK points over Flat (Table 16), against 2.23 and 6.28 added by typed modeling. Of total gains of 5.61 and 17.92, removing the relation map costs 0.58 and 0.53 (Table 3), and shuffling days 0.03 and 0.62 (Table 13). What supports this attribution and the relational design?

  3. The split has only Train and Test (line 353), and line 783 fixes depth and readout from Table 8, whose row equals the seed-42 Test results of Table 10 on four metrics. So the configuration was selected on Test. What are the results with selection on a separate validation set?

  4. Class F needs a later hospitalization or a recorded death after day 365 (lines 670 to 672), and MIMIC-IV censors deaths beyond one year after the last discharge, so F requires later hospital contact while the death classes do not. Keeping 133,737 of 213,971 Train examples (lines 355 to 358) on this rule is not "natural-distribution", even unresampled. How do results change under a censoring-aware analysis?

  5. Against Flat, Macro-AUROC moves by 0.62 points and Accuracy by -0.29 (Table 11) while F recall falls from 97.87 to 94.57 (Table 9), which looks like a shifted decision boundary. Qwen3-4B SFT beats HiReT on both metrics (Table 2). Do the Balanced Accuracy, Macro-F1, and QWK gains survive at matched F recall or after logit adjustment for all methods?

  6. Flat reaches Macro-AUROC 76.86 while every structured baseline stays at or below 62.34 (Table 2), so that gap comes from the frozen text encoder, not the hierarchy, and the "strict" control (line 368) is only global pooling. How do the existing baselines or a serialized Transformer perform on the same node embeddings?

  7. The newest non-LLM baseline in Table 2 is Hi-BEHRT (2022), and among newer LLMs only Qwen3-4B is trained. Section 2's relational models are not compared, and DuETT [1], HEART [2], the temporal graph model [3], REMed [4], and ProtoEHR [5] are neither compared nor cited. How does HiReT compare with them under the same splits, seeds, and tuning?

[1] Alex Labach, Aslesha Pokhrel, Xiao Shi Huang, Saba Zuberi, Seung Eun Yi, Maksims Volkovs, Tomi Poutanen, and Rahul G. Krishnan. DuETT: Dual event time transformer for electronic health records. In MLHC, 2023.

[2] Tinglin Huang, Syed Asad Rizvi, Rohan Krishna Thakur, Vimig Socrates, Meili Gupta, David van Dijk, R. Andrew Taylor, and Rex Ying. HEART: Learning better representation of EHR data with a heterogeneous relation-aware transformer. Journal of Biomedical Informatics, 159:104741, 2024.

[3] Jiayuan Chen, Changchang Yin, Yuanlong Wang, and Ping Zhang. Predictive modeling with temporal graphical representation on electronic health records. In IJCAI, 2024.

[4] Junu Kim, Chaeeun Shim, Bosco Seong Kyu Yang, Chami Im, Sung Yoon Lim, Han-Gil Jeong, and Edward Choi. General-purpose retrieval-enhanced medical prediction model using near-infinite history. In MLHC, 2024.

[5] Zi Cai, Yu Liu, Zhiyao Luo, and Tingting Zhu. ProtoEHR: Hierarchical prototype learning for EHR-based healthcare predictions. In CIKM, 2025.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.