TL;DR for operators
Historical outcomes from similar ICU patients can improve how predictive models rank patients at risk of newly recorded abnormalities. NECA demonstrates this across four EHR-defined endpoints and two retrospective datasets, with the largest average improvements in neurological event prediction.
The qualification is that most of the observed improvement in one key XGBoost comparison comes from a single historical event rate, not the complete four-horizon curve. For neurological prediction on PhysioNet 2012, average PR-AUC increases from 0.331 without augmentation to 0.417 with one target-horizon rate and 0.443 with all four rates.
That additional 0.026 is measurable, but the experiments do not establish that event timing independently causes the improvement. Clinical risk-scoring teams should compare simpler augmentation against the full curve using matched inputs and leakage-resistant evaluation. Neither approach has demonstrated calibrated clinical risk, improved monitoring decisions, or patient benefit.
Six hours of observations, four possible prediction windows
Six hours after ICU admission, a hospital analytics system has an initial collection of vital signs, laboratory results, and other clinical measurements. It may need to estimate which patients will subsequently develop neurological, renal, metabolic, or inflammatory abnormalities.
Historical ICU records offer additional information. Rather than relying exclusively on the current patient’s measurements, the system can identify earlier patients with similar initial observations and examine how frequently those patients subsequently experienced qualifying events.
The design question is how much of that historical outcome information the model needs.
A single event rate can summarize the fraction of comparable historical patients with a recorded abnormality within the next 24 hours, for example. A sequence of cumulative rates can additionally describe how that fraction changes over several future horizons.
The richer representation might capture clinically informative differences in event onset. It might also add little beyond the event rate already available for the prediction horizon. Choosing between these alternatives requires separating the value of historical outcomes from the incremental value of their timing.
NECA turns historical ICU outcomes into prediction features
Chenyang Wang and colleagues introduce Neighbor Event-Curve Augmentation (NECA), a model-agnostic method for enriching existing ICU prediction models with historical event information.1
The researchers evaluate two sources: 4,000 ICU records from PhysioNet 2012 Set A and a selected 4,000-stay subset of MIMIC-IV. Eligibility differs by endpoint because records already satisfying the relevant abnormality rule during the first six hours are excluded.
The endpoints are operational definitions based on recorded measurements. For example, the neurological rule uses a Glasgow Coma Scale score of 12 or lower, while renal events can be triggered by elevated creatinine or blood urea nitrogen. These labels represent newly recorded qualifying abnormalities, not necessarily clinically confirmed organ failure.
NECA uses the following procedure:
- Extract clinical information from the first six hours after ICU admission.
- Identify up to 100 similar historical training records using Euclidean distance over 27 standardized abnormality indicators.
- Calculate the cumulative fraction of those neighbors with a newly recorded qualifying event within 6, 12, 24, and 42 hours after the observation window.
- Append those four values to the current patient’s predictive representation.
For tabular models, the four rates become additional input features. For sequence models, they are appended to the final encoder representation.
The resulting vector summarizes historical event occurrence across four horizons. It is not itself a calibrated patient-level probability. The downstream classifier learns how to combine this information with the patient’s own measurements.
The researchers evaluate seven learners, including logistic regression, XGBoost, LightGBM, TabPFN, LSTM, ConCare, and an APRICOT-M-style state-space model. Separate classifiers are trained for each endpoint and prediction horizon.
The aggregate improvements are largest for neurological events
The main experiment compares models using their original inputs, labeled Raw, against models augmented with the complete NECA curve.
One reported metric is PR-AUC, the area under the precision-recall curve. It summarizes how well a model balances finding positive cases against the precision of its positive predictions across classification thresholds. It is especially informative when qualifying events are relatively uncommon, but does not measure probability calibration or clinical benefit.
The paper reports the following average test PR-AUC results.
| EHR-defined endpoint | PhysioNet 2012 Raw → NECA | MIMIC-IV Raw → NECA |
|---|---|---|
| Neurological | 0.461 → 0.557 | 0.419 → 0.520 |
| Renal | 0.353 → 0.387 | 0.319 → 0.345 |
| Electrolyte/metabolic | 0.357 → 0.378 | 0.340 → 0.357 |
| Inflammatory | 0.514 → 0.538 | 0.276 → 0.289 |
Source: Section 4.1, Table 5. Test results are averaged across four horizons, seven learners, and random seeds.
Mean ROC-AUC also increases for every endpoint on both datasets. For neurological prediction, it rises from 0.689 to 0.766 on PhysioNet and from 0.749 to 0.809 on MIMIC-IV.
These are consistent aggregate improvements in retrospective discrimination, not evidence that every model improves at every horizon. The paper also provides no paired confidence intervals for determining the statistical precision of the differences.
Event prevalence matters when interpreting the results. In MIMIC-IV, the recorded neurological event rate increases from 7.3% at the six-hour prediction horizon to 13.9% at 42 hours. Comparisons of PR-AUC across datasets therefore reflect different underlying event frequencies as well as model performance.
The next question is which component of NECA accounts for the reported improvement.
One historical event rate provides much of the gain
The paper’s single-horizon ablation directly examines whether a complete event curve improves prediction beyond the cumulative event rate at the target horizon.
On PhysioNet 2012 using XGBoost, the comparison is:
| Endpoint | Raw | Single-horizon | Full curve | Full-curve increment |
|---|---|---|---|---|
| Neurological | 0.331 | 0.417 | 0.443 | +0.026 |
| Renal | 0.328 | 0.358 | 0.366 | +0.008 |
| Electrolyte/metabolic | 0.326 | 0.331 | 0.341 | +0.010 |
| Inflammatory | 0.490 | 0.507 | 0.513 | +0.006 |
Source: Section 4.4, Table 8. Values are average PR-AUC across horizons for XGBoost, not the seven-learner averages in the main experiment.
For neurological prediction, the first historical event rate adds 0.086 PR-AUC over Raw. The remaining three event-rate features add another 0.026.
This supports a narrower interpretation than the aggregate results alone might suggest: retrieval-derived event occurrence information is valuable in this comparison, while the additional discrimination from the full temporal representation is smaller.
The same pattern appears across the other three endpoints, although the absolute improvements are smaller.
For a clinical analytics developer, this creates a legitimate model-selection question. Both variants require historical neighbor retrieval, and adding three further cumulative rates does not necessarily require additional retrieval searches. The paper does not measure their relative computational or maintenance costs.
The decision should therefore depend on whether the incremental discrimination survives stronger validation and changes operationally relevant predictions, rather than assuming that four horizons necessarily justify a more complex representation.
The controls do not isolate event timing
Two additional experiments help explain why the full-curve improvement should not automatically be attributed to clinically meaningful event timing.
First, a retrieval-feature ablation tests how the clinical representation used to identify neighbors affects prediction. In neurological validation, retrieval using 27 abnormality indicators achieves average PR-AUC of 0.550, compared with 0.534 for laboratory-variable retrieval and 0.446 for random neighbors.
This establishes sensitivity to the retrieval representation. However, the tested feature sets contain different clinical information, so the experiment does not independently isolate the quality of the similarity calculation.
A separate neighbor-mean augmentation comparison also indicates that historical feature information can improve prediction without the complete event curve.
Second, the event-time shuffle control is intended to examine dependence on historical event information. Neurological PR-AUC falls from 0.519 to 0.355 on PhysioNet and from 0.493 to 0.342 on MIMIC-IV under the reported shuffle procedure.
Those decreases are substantial. But the shuffle assigns finite event times even to historical records without recorded events. It therefore changes both whether neighbors are represented as experiencing an event and when that event occurs.
Consequently, the performance decline does not isolate the predictive contribution of timing alone.
A further attribution problem affects the main Raw-versus-NECA comparison: the 27 cross-system retrieval indicators include clinical information unavailable as direct inputs to some Raw tabular models. Improvements can therefore reflect additional clinical variables, historical event status, event timing, or combinations of these signals.
A timing-specific conclusion would require controls that preserve event occurrence and available clinical information while selectively altering temporal relationships.
Better retrospective discrimination is not yet a clinical workflow improvement
Before an ICU analytics team uses NECA to prioritize monitoring or generate clinical alerts, several evaluation issues need resolution.
The MIMIC-IV analysis uses record-level rather than patient-disjoint partitions, allowing different admissions from the same individual potentially to appear across training and evaluation sets. Training event curves are also not cross-fitted. Although each training query excludes itself from retrieval, other records from the same patient are not necessarily excluded.
These design choices create potential optimism in the reported performance.
The event labels introduce another uncertainty. Missing measurements, ICU discharge, incomplete follow-up, and differences in testing frequency can affect whether an abnormality becomes visible in the EHR. Neurological measurements may additionally reflect sedation or intubation rather than deterioration alone.
The system’s predicted probabilities also remain unvalidated for clinical use. The paper does not assess probability calibration, and independently trained horizon-specific classifiers are not constrained to produce nondecreasing cumulative risks as the prediction horizon increases.
Cognaptus inference: A clinical risk-scoring team evaluating this approach should establish information-matched Raw baselines, patient-disjoint testing, and cross-fitted retrieval features before investing in operational integration. It should then compare target-horizon augmentation against the complete curve, assess event-label reliability and calibration, and test whether improved rankings translate into acceptable alert burdens and better clinical decisions.
These are proposed validation requirements, not demonstrated outcomes of NECA.
Prospective evaluation remains necessary before either augmentation method influences monitoring priorities, treatment decisions, or clinical escalation.
The useful result is retrieval enrichment; the timing claim remains open
NECA provides an interpretable way to connect a patient’s early measurements with observed outcomes from similar historical ICU records. Its retrospective benchmarks show improved average discrimination across multiple endpoints, models, and prediction horizons.
The more specific proposition—that representing the complete temporal distribution of historical events is responsible for a substantial improvement—receives weaker support. A single target-horizon event rate already captures much of the observed XGBoost gain, and the existing controls cannot cleanly separate event timing from other sources of predictive information.
For teams building clinical prediction systems, NECA is a candidate augmentation to test against simpler alternatives under stronger evaluation controls. The evidence supports further predictive development, not yet a claim of improved ICU care.
Cognaptus: Automate the Present, Incubate the Future.
-
Chenyang Wang and Theresa Jiawei Tang and Xinran Liu and Xinyu Lu and Xulei Yang (2026). NECA: Neighbor Event Curve Augmentation for Multi-Horizon ICU Deterioration Prediction. arXiv:2610.40420. https://arxiv.org/abs/2610.40420 ↩︎