PaperlayerView exact revision on arXiv

Source-linked AI summary

NECA: Neighbor Event Curve Augmentation for Multi-Horizon ICU Deterioration Prediction

Chenyang Wang, Jiawei Theresa Tang, Xinran Liu, Xinyu Lu, Xulei Yang

arXiv:2610.40420v1cs.LGcs.AI

TL;DR

Early ICU prediction must estimate organ-specific event risk across future horizons, but existing supervised models generally do not use similar patients’ subsequent event timing. NECA adds cumulative event curves from retrieved training patients to standard predictors; these features complement supervised prediction, especially for neurological dysfunction.

  • Problem

    Existing supervised ICU models generally treat patients independently and provide limited case-based evidence about when organ dysfunction events may occur across future horizons.

    Source: §1 · p. 2

  • Method

    NECA retrieves training patients using early-window information and concatenates their cumulative event rates at 6h, 12h, 24h, and 42h with the original patient representation.

    Source: §1 · p. 2

  • Results

    Event-curve features consistently complement supervised predictors, especially for neurological dysfunction; the target-horizon neighbor rate explains much of the improvement, while other curve points add smaller increments.

    Source: §1 · p. 2, §5 · p. 12

  • Takeaways & Limitations

    Similar-patient event timing can serve as an inspectable, structured temporal risk signal for standard predictors.

    Source: §1 · p. 2, §5 · p. 12

    The findings remain preliminary and do not support claims about event-time specificity, generalization, or clinical benefit.

    Source: §5 · p. 12

Abstract

from arXiv · show

Early prediction of ICU organ dysfunction events supports timely clinical monitoring and intervention, but estimating short-term organ-specific risk from early ICU records remains challenging due to irregular measurements, heterogeneous clinical variables, and the need to predict risk across multiple future horizons. Existing supervised models usually treat each patient as an isolated sample and do not explicitly use the observed event timing of similar historical patients. We propose Neighbor Event-Curve Augmentation (NECA), a model-agnostic method that retrieves training-set patients with similar early-window profiles and converts their subsequent organ dysfunction event onset times into cumulative event-curve features over 6h, 12h, 24h, and 42h. These features are concatenated with the original patient representation and used by standard downstream predictors. On PhysioNet 2012 and a 4,000-stay MIMIC-IV subset, NECA improves average PR-AUC and ROC-AUC across four organ dysfunction event tasks and seven supervised learners. The largest gain appears for neurological dysfunction events, where task-level average PR-AUC increases from 0.461 to 0.557 on PhysioNet 2012. Ablations show that clinically meaningful retrieval and horizon-specific neighbor risk drive most of the improvement, while the full cumulative curve adds task-dependent temporal context. These results support similar-patient event timing as a structured temporal risk signal for ICU organ dysfunction event prediction. The source code will be publicly released upon acceptance of this paper.

1 Introduction

Existing ICU predictors generally do not use similar historical patients’ subsequent event timing to provide case-based evidence across multiple horizons. NECA adds cumulative event curves from retrieved training patients to standard predictors, and these features complement supervised prediction, especially for neurological dysfunction.

  • Existing supervised models typically predict fixed-horizon risk without explicitly using subsequent trajectories of clinically similar historical patients.This limits case-based evidence about whether events occur within 6, 12, 24, or 42 hours.Source: §1 · p. 2
  • NECA retrieves training-set patients using early-window information and adds their cumulative event-onset rates at 6h, 12h, 24h, and 42h to the original patient representation.The downstream architecture remains unchanged, and each feature corresponds to a concrete horizon.Source: §1 · p. 2
  • The task covers four EHR-defined organ dysfunction event categories and predicts across 6h, 12h, 24h, and 42h horizons.Source: §1 · p. 2
  • Event-curve features consistently complement supervised predictors, especially for neurological dysfunction events.Source: §1 · p. 2

2 Related Work

Prior ICU prediction work includes global acuity, mortality, and organ-specific risk models, while retrieval work draws on text, clinical corpora, knowledge graphs, and patient similarity. NECA distinguishes itself by providing predictors with inspectable event rates from retrieved historical ICU records.

2.1 Early ICU Organ Dysfunction Event Prediction

  • Earlier ICU prediction models address mortality, decompensation, length of stay, phenotype, and acuity using benchmark, recurrent, personalized-embedding, and state-space approaches.Source: §2.1 · p. 3
  • Organ-level and multi-label studies address short-horizon dysfunction risks, while NECA adds recorded outcomes from retrieved historical records.Examples include SeqSNR’s six organ dysfunction events and IOC-MT’s graph attention for inter-organ correlations.Source: §2.1 · p. 3

2.2 Retrieval-Augmented and Patient-Similarity Clinical Prediction

  • Clinical retrieval methods use text, multi-source corpora, knowledge graphs, or community summaries, but their evidence units and downstream tasks differ from NECA’s historical ICU records.The cited examples include MedRetriever, RAM-EHR, GraphCare and KARE, and CLEAR.Source: §2.2 · p. 3
  • Patient-similarity methods use related trajectories or similarity graphs, whereas NECA supplies four cumulative neighbor event rates that make historical evidence inspectable [26].Source: §2.2 · p. 3

3 Methods

NECA adds similar training patients’ post-window event timing to early ICU records as cumulative, multi-horizon risk features. The methods define the event targets, representations, retrieval procedure, model integration, and evaluation protocol.

3 Methods

  • NECA constructs early-window representations, retrieves similar training records, summarizes their later event times, and adds the resulting event curve to a predictor.The framework outputs predictions over multiple horizons.Source: §3 · p. 3, Figure 1 · p. 4

3.1 Data Sources and Cohort Construction

  • The study evaluates 4,000 PhysioNet 2012 records and a 4,000-stay MIMIC-IV subset, with MIMIC stays restricted to adults and specified admission criteria.MIMIC stays are sampled without replacement using a fixed random seed.Source: §3.1 · p. 4

3.2 Multi-Horizon Organ Dysfunction Event Prediction

  • Targets are first newly recorded endpoint events after the six-hour early window, labeled separately at 6, 12, 24, and 42 hours.An event during the early window excludes a record for that endpoint; missing early measurements can still leave it eligible.Source: §3.1 · p. 4, §3.2 · pp. 5–6

3.3 Early-Window Feature Representation

  • Tabular inputs summarize early-window measurements, while sequence models use hourly bins of the same time series in temporal order.Tabular summaries include first, last, minimum, and maximum values; sequence bins use the last observed value.Source: §3.3 · p. 6

3.4 Neighbor Event-Curve Augmentation

  • For each endpoint, NECA converts retrieved neighbors’ recorded event times into a nondecreasing curve of weighted event fractions at four prediction horizons.The curve is appended to the tabular vector or to a sequence model’s final hidden representation.Source: §3.4 · p. 7

3.5 Neighbor Retrieval and Leakage Control

  • Main experiments retrieve K=100 training records using 27 standardized abnormality indicators, Euclidean distance, and equal weights.Record-level splits can place admissions from one subject in different partitions; training event curves are not cross-fitted, potentially making performance optimistic.Source: §3.5 · p. 7

3.6 Prediction Models

  • NECA is compared across seven supervised learners, with separate classifiers for each endpoint and prediction horizon.Sequence models append the curve after encoding; the APRICOT-M-style model is an adaptation, not a reproduction of the original multi-task system.Source: §3.6 · p. 8

3.7 Evaluation Protocol

  • Evaluation reports test ROC-AUC and PR-AUC, averaging across horizons, learners, and seeds for endpoint-level results.A KNN reference uses event rates among the same neighbors, but the comparison does not isolate neighbor outcomes from additional retrieval inputs.Source: §3.7 · p. 9

4 Results

Across PhysioNet 2012 and the 4,000-stay MIMIC-IV subset, Curve-NECA has higher endpoint-level mean PR-AUC and ROC-AUC than Raw. Ablations favor retrieval using 27 abnormality indicators and show that single-horizon neighbor risk already improves over Raw, while differing inputs and the shuffled-time procedure constrain attribution.

4.1 Overall Performance of NECA

  • Across four endpoints, Curve-NECA exceeds Raw in mean PR-AUC and ROC-AUC; neurological PR-AUC rises from 0.461 to 0.557 on PhysioNet 2012 and from 0.419 to 0.520 on MIMIC-IV.Mean gains do not establish gains for every learner.Source: §4.1 · pp. 9–10, Table 5 · p. 10
  • MIMIC-IV results need subject-disjoint re-evaluation; the 4,000-stay, single-source analysis does not establish external validity.Records from one subject can fall in different partitions.Source: §4.1 · p. 10

4.2 Neighbor Retrieval Ablation

  • On neurological events, retrieval with 27 abnormality indicators ranks highest at PR-AUC 0.550 and ROC-AUC 0.744, but differing feature sets prevent isolating the similarity measure.Laboratory-variable retrieval ranks second.Source: §4.2 · p. 10, Table 6 · p. 10

4.3 Shuffled Event-Time Control

  • The shuffled-curve control changes event status as well as event time, so its score decrease cannot isolate the contribution of timing.A timing-only test would shuffle times among neighbors with recorded events by 42 hours while preserving event status.Source: §4.3 · p. 11, Table 7 · p. 11

4.4 Single-Horizon and Neighbor-Mean Ablations

  • The single-horizon rate already improves over Raw XGBoost; on neurological events, validation PR-AUC is 0.509 for neighbor-mean, 0.443 for Raw, and 0.550 for Curve-NECA.The full curve adds a smaller increment, especially for neurological events.Source: §4.4 · p. 11, Table 8 · p. 11

5 Discussion and Conclusion

NECA’s target-horizon neighbor rates explain much of its improvement, while other curve points add smaller increments. The findings remain preliminary because endpoint definitions, evaluation choices, and calibration leave important limits on interpretation.

  • Target-horizon neighbor rates explain much of NECA’s improvement, while other curve points add smaller increments.Source: §5 · p. 12
  • Independent horizon classifiers can yield decreasing risks at longer horizons, and the frequency of such predictions is unmeasured.Source: §5 · p. 12
  • The neurological-event difference is largest, but its relation to cross-system retrieval variables remains untested; low GCS may reflect sedation, intubation, or measurement practice.Source: §5 · p. 12
  • Endpoint rules lack clinical adjudication, and missing measurements or insufficient follow-up can be coded as no recorded event.Source: §5 · p. 12
  • Raw tabular models omit some retrieval variables; MIMIC-IV uses a record-level split, training curves were not cross-fitted, and K=100 was not validation-selected.Source: §5 · p. 12
  • The shuffle control changes event status and mean differences lack paired uncertainty estimates, limiting claims about event-time specificity, generalization, or clinical benefit.Source: §5 · p. 12

A Retrieval Indicator Definitions

The retrieval vector uses 27 entries from the first six ICU hours, combining basic clinical indicators, system-level and interaction indicators, and normalized summary fields.

  • Thirteen basic indicators encode first-six-hour findings such as age, hypotension, tachycardia, hypoxemia, laboratory abnormalities, ventilation, measurement count, and GCS.Source: A Retrieval Indicator Definitions · p. 12
  • Six system indicators summarize hemodynamic, respiratory, renal, inflammatory, hepatic, and coagulation findings; five interaction indicators capture combinations such as shock-like and cardio-respiratory patterns.Source: A Retrieval Indicator Definitions · p. 12
  • The final three entries normalize abnormal-system count, severe-signal count, and ICU type; missing measurements make binary rules false, and missing ICU type is zero.Some retrieval variables are unavailable to the endpoint-specific Raw tabular predictor.Source: A Retrieval Indicator Definitions · p. 12

Related papers

Loading 2610.40420v1…