arXiv News

First-hand information for everyone

Site

AboutPrivacy PolicyTerms of Use

© 2026 arXiv News

arXiv News

EnglishJapanese

Switch language

EnglishJapanese
Loading account…

First-hand information for everyone

Back to front page
Machine LearningEnglishPublished

Neighbor event‑curve augmentation boosts ICU organ‑dysfunction prediction

October 31, 2026arXiv: 2610.40420v1

Neighbor‑based augmentation (NECA) adds a cumulative curve of when similar patients develop organ problems to a patient’s early‑ICU data, letting standard predictors look beyond the immediate record. The method retrieves up to K training patients whose early‑window profiles match the query and, for each prediction horizon (6 h, 12 h, 24 h, 42 h), counts the fraction of those neighbors that experience the event within that horizon. These fractions become new features that are concatenated with the original representation and fed to any downstream model.

Diagram of the NECA framework: data preprocessing builds early‑window representations; a neighbor‐based event‑curve is derived and concatenated; the combined feature set feeds a model that outputs predictions for multiple horizons.
Figure 1. NECA pipeline showing early‑window representation, neighbor event‑curve augmentation, model integration, and multi‑horizon output.Wang, Chenyang, Tang, Jiawei Theresa, Liu, Xinran, Lu, Xinyu, Yang, Xulei · arXiv:2610.40420v1 · Figure 1: · CC BY 4.0

On PhysioNet 2012 the Curve‑NECA model outperformed the raw baseline across all four organ‑dysfunction endpoints. For neurological events the mean PR‑AUC rose from 0.461 to 0.557 and the mean ROC‑AUC from 0.689 to 0.766. Similar gains appeared for renal, electrolyte/metabolic and inflammatory events. The same pattern held on a 4,000‑stay subset of MIMIC‑IV.

Tabular summary of performance showing higher values for Curve‑NECA over Raw for each endpoint and horizon on PhysioNet 2012.
PR-AUCROC-AUC
DatasetEndpointRawCurve-NECARawCurve-NECA
PN12Neurological0.4610.5570.6890.766
Renal0.3530.3870.7110.740
Electrolyte/metabolic0.3570.3780.6710.692
Inflammatory0.5140.5380.6920.727
MIMICNeurological0.4190.5200.7490.809
Renal0.3190.3450.6680.693
Electrolyte/metabolic0.3400.3570.6620.674
Inflammatory0.2760.2890.6790.696
Table 5. Mean PR‑AUC and ROC‑AUC for the four organ‑dysfunction endpoints on PhysioNet 2012, comparing Raw to Curve‑NECA across the 6 h, 12 h, 24 h, and 42 h horizons.Wang, Chenyang, Tang, Jiawei Theresa, Liu, Xinran, Lu, Xinyu, Yang, Xulei · arXiv:2610.40420v1 · Table 5: · CC BY 4.0

The study used a record‑level split, so records from the same patient could fall in training and test sets, potentially inflating the reported gains. A subject‑disjoint split and cross‑fitted neighbor curves would be needed to confirm the robustness of the improvement.

Figure 1 illustrates the NECA pipeline, while Table 5 presents the quantified superiority of Curve‑NECA over the raw baseline.

Source paper

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

Machine LearningArtificial Intelligence
View on arXiv
Neighbor event‑curve augmentation boosts ICU organ‑dysfunction prediction | arXiv News