acceptodds
Under review as a conference paper at ICLR 2027

Clinical Irregular Time-Series Classification via OPERA: Observation-Process Encoding and Residual Adaptation

Abstract

In clinical irregular multivariate time series (IMTS), how often measurements are taken, when they occur, and which variables are recorded provide predictive information beyond measured values. However, existing approaches often couple observation modeling with specific architectures, limiting the flexibility to combine monitoring information with the distinct modeling strengths of different IMTS backbones. We propose Observation-Process Encoding and Residual Adaptation (OPERA), a lightweight plug-in module for clinical IMTS classification. Its Observation-Process Encoder uses observation times and variable identities, but not measured values, to construct human-readable descriptors of monitoring amount, timing, and variable selection, making monitoring patterns explicit and analyzable. The Residual Observation Gate then adaptively adds an observation-based correction to the backbone's patient representation, guided by both the monitoring pattern and the backbone representation. This patient-level interface preserves the original encoding and aggregation mechanisms, allowing the same design to complement heterogeneous IMTS backbones. Experiments on public intensive care unit (ICU) benchmarks show that OPERA consistently improves all five evaluated backbones at only marginal additional computational cost and achieves the best aggregate performance among representative methods.

Primary area:
learning on time series and dynamical systems
Keywords:
Irregular Multivariate Time Series, Clinical Prediction, Observation-Process Modeling, Informative Missingness

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_V5grOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

OPERA adds an observation-process branch to existing clinical irregular time-series classifiers. It constructs deterministic descriptors of monitoring amount, timing, and variable selection, embeds them with a small network, and applies a gated residual correction after the backbone's native readout. The paper evaluates three ICU cohorts and five neural backbones, with higher mean AUROC and AUPRC in all 15 paired backbone-dataset comparisons. Observation-only gradient boosting, ablations, and runtime measurements provide additional evidence. The practical compatibility result is useful. The evidence for the specific descriptor and gate design is less conclusive.

Critical strengths
  1. The patient-level interface is useful and tested beyond a single architecture. OPERA preserves each backbone's encoder and native readout, and Table 1 reports gains across five heterogeneous backbones. Appendix C.1 matches inputs, preprocessing, backbone settings, training schedules, and seeds within each pair. This makes the compatibility claim more convincing than a comparison against unrelated models alone.

  2. The observation branch is inspectable. Appendix A defines the descriptors precisely, and HGB-Obs shows predictive information without measured values. HGB-Joint improves on HGB-Value in all six dataset-metric comparisons. The paper also acknowledges in Appendix E that these are predictive monitoring associations and that transfer across institutions remains untested.

Critical weaknesses
  1. The experiments establish the benefit of adding monitoring summaries more convincingly than the advantage of the proposed design. Table 3 does not sufficiently specify the Learnable Encoder or Direct Concat, including event aggregation, dimensions, classifier construction, parameter counts, and selection procedures. HGB-Value's measured-value summaries are also not enumerated. The weaker results of these controls therefore do not establish that explicit descriptors or jointly conditioned gating are preferable to adequately configured alternatives. The ungated comparison changes both capacity and conditioning; no control isolates the contribution of conditioning the gate on both inputs. Some differences are small. On MIMIC-III, Full OPERA exceeds Direct Concat by only 0.10 AUPRC points; on P12, removing intensity changes AUPRC by 0.02 points. Three-seed standard deviations on a fixed test split do not quantify uncertainty in these paired differences. I would need reproducible, capacity-comparable controls and paired uncertainty, or a narrower interpretation of the design-specific conclusions (Section 4.3, Table 3, Appendix C).

  2. The mortality evaluation needs a clearer definition of prediction time and patient independence. Appendix B defines MIMIC-III samples by HADM_ID, while Appendix C.1 calls the split patient-level without explaining whether admissions belonging to the same person are grouped. It also starts the input window at the first selected event. For both mortality cohorts, the manuscript does not state how death, discharge, or truncated records before the end of the 48-hour window are handled. These details matter particularly because OPERA uses observation amount and last-observed time. If early-ending records are retained, the model may partly classify record termination rather than predict a later outcome at a common landmark. The manuscript does not demonstrate that this occurs, but the present description does not resolve the distinction. Matching the Native-OPERA split ensures a fair comparison within that split; it does not establish person-level independence or temporal eligibility. Explicit cohort rules and an overlap audit are needed before the mortality results can be interpreted with stronger confidence (Appendices A, B, and C.1).

  3. The descriptor groups overlap in information, which limits what the group ablations establish. Appendix A retains the mean active-variable fraction in the variable group. This is proportional to the average measurement density in the intensity group, apart from the numerical-stability convention. Removing the intensity group therefore leaves some intensity information available. The early-allocation fraction and half-window contrast are also algebraically related. Please acknowledge these dependencies and interpret Table 3 as the marginal effect of removing the listed coordinates. If the intent is to establish the independent contribution of an information category, the corresponding ablation should remove its overlapping representations as well (Section 4.3, Appendix A).

  4. Figure 2 does not isolate why the predictions change. Section 3.5 states that the backbone and classifier are jointly trained with OPERA, so the native and adapted predictions come from models whose value representations and prediction heads may also differ. The displayed variable-selection differences cannot by themselves attribute the risk changes to those descriptors. Please state how the pair was selected, including whether adapted predictions were used in selection. The example should remain an illustration of contrasting monitoring profiles unless a within-model descriptor intervention, with the trained backbone representation and model parameters fixed, supports a stronger attribution (Section 4.4, Figure 2).

Questions for authors
  1. What are the architecture, pooling, parameter count, and validation-selection procedure for Learnable Encoder and Direct Concat? Please also define the HGB-Value summaries. Do comparable controls and gates conditioned on only one input preserve the claimed design advantage?

  2. Were all admissions belonging to one MIMIC-III person assigned to the same partition? Please provide unique-person counts and a cross-partition overlap audit. If overlap exists, how do the paired gains change under person-grouped splitting?

  3. At what time is each mortality prediction intended to be made? How are deaths and discharges before the completed input window handled, and how many such records are present? If the intended use is prediction after 48 hours, do the gains persist among patients eligible at that landmark?

  4. Can you report seed-level paired differences and uncertainty from resampling test patients for the main gains and the fusion ablations? This would help distinguish a consistent increase in reported means from a reliably established advantage of the particular design.

Reviewer_h2sTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes OPERA, a add-on module for clinical irregular time-series classification. The motivation is that how often, when, and which variables are measured carries information beyond the measured values. OPERA computes a fixed set of hand-crafted statistics from observation times and variable IDs, without using values. A 2-layer MLP embeds them, and the result is added to the backbone’s patient-level representation through a sigmoid-gated residual. The method is evaluated on P12, P19, and MIMIC-III, all using the first 48 hours and random in-domain splits. OPERA improves all five backbones it is attached to. The paper also includes ablations and a case study.

Critical strengths
  • I think the motivation is reasonable and the module is easy to use. Informative observation patterns in ICU data are a real phenomenon, and a backbone-agnostic way to add them is practically useful.
  • The HGB diagnostics results is quite interesting, especially the performance of HGB-Obs. It shows that observation statistics alone carry predictive signal (HGB-Obs). This is a simple and clear piece of evidence for the paper's premise.
Critical weaknesses
  • Although the motivation is reasonable, the method relies heavily on hand-crafted observation statistics and standard gated residual fusion, limiting its methodological novelty. The choice of statistics is largely heuristic, and the resulting representation cannot capture information discarded by these fixed summaries. While the subsequent MLP is learned, the sampling process itself is not explicitly modeled. More importantly, the "Learnable Encoder" ablation lacks sufficient implementation details, and the gains over direct concat are small relative to the reported variability. The tabular baselines further weaken the justification for the neural module: HGB-Value outperforms all neural models on P12, while HGB-Joint performs comparably to RNN-Decay+OPERA on MIMIC3.
  • OPERA’s gains are shown only on in-domain splits, so it is unclear whether the descriptors capture transferable monitoring signal or site-specific artifacts. This matters most for P19, which pools two hospitals, uses source-stratified splits, and shows the largest gains; variable-selection descriptors could simply encode hospital identity. Cross-site evaluation (e.g., hospital A to B in P19, hospital-wise splits on eICU, or CareVue to MetaVision in MIMIC-III) would address this. In addition, all tasks are single-shot predictions from a fixed 48-hour window, so the method is not tested under monitoring patterns that evolve over time.
Questions for authors
  1. What is the architecture and training objective of the "Learnable Encoder" in Table 3?
  2. Can you report cross-site results, for example hospital-wise splits on eICU or with benchmark such as YAIB?
  3. Given that HGB-Value outperforms all neural models on P12 and HGB-Joint is on par with RNN-Decay+OPERA on MIMIC-III, what is the practical advantage of the neural plug-in over simply using a tabular model with the same descriptors?
  4. On P19, do the OPERA gains persist when evaluated within each hospital separately?
Reviewer_fuXcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Existing approaches often couple observation modeling with specific architectures (e.g., tabular form or event-based tokenization), limiting the flexibility to combine monitoring information with the distinct modeling strengths of different IMTS (irregular multivariate time series) backbones. Hence, the authors proposed a two-branch architecture for improvement: (1) an existing IMTS backbone that encodes the measured clinical series and an Observation-Process Encoder that summarizes monitoring patterns. (2) the Residual Observation Gate aggregates the encoded monitoring summary to the IMTS backbone’s original representation for a better context for downstream tasks. Although the experiments were limited to ICU studies, the proposed method shows potential for more comprehensive medical event prediction by considering observation patterns.

Critical strengths
  1. Based on recent studies and clinical practice, it is crucial to consider the observation patterns when designing medical representation construction for downstream tasks.
Critical weaknesses
  1. Current experimental evidence is primarily limited to ICU settings, which involve critically ill patients and intensive monitoring. The proposed method has not been thoroughly evaluated in broader clinical settings, particularly those involving chronic diseases and less acute conditions, where differences in measurement frequency and variable types may significantly affect its applicability.
  2. The performance gains appear marginal on certain datasets. At a minimum, statistical significance testing should be conducted to determine whether the observed improvements are statistically significant.
Questions for authors
  1. Please further justify the choice of datasets for experiments (only on ICU cases).
  2. The authors are encouraged to conduct statistical significance tests to further substantiate the effectiveness of the proposed method.
  3. Some baselines appear to exhibit substantial performance discrepancies compared to their originally reported results (e.g., STraTS on the P12 and MIMIC-III datasets). The authors should clarify the possible reasons for these discrepancies.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.