Early warning of late-onset sepsis at a fixed alarm budget: a benchmark of prediction models in 3,829 very-low-birth-weight infants from four NICUs

Objective: To measure what neonatal early-warning models detect, and how early, at a fixed alarm budget. Materials and Methods: On the public PreMo cohort (3,829 very-low-birth-weight infants, four US NICUs), a benchmark scored detection and warning time at fixed alarm budgets over monitored time for four outcomes. Population (engineered-feature) and deep sequence models, frozen on selection infants, were scored once on 747 test infants. Results: At 10 alarms per 100 infant-days, the registered comparisons' four detection points detected 0.430 to 0.574 of monitored episodes of treated late-onset infection (162 of 225 evaluable test episodes had warning-window monitoring), against 0.056 for random alarms and 0.414 for baseline risk (floors chosen post hoc by the selection rule). Deep models did not detect more: the first deep arm trailed the unsearched population model (-0.144; 97.5% interval -0.233 to -0.051), and an extended search, a declared secondary, was inconclusive against a searched population model falling from 0.597 (selection) to 0.438 (test); every deep point estimate for the primary outcome trailed the unsearched model's (post hoc, extended search: -0.089; 95% interval -0.169 to -0.011). Conformal thresholds held the budget at three of four points as registered (two strictly). Among equally accurate models, higher detection came with shorter median warning (0.51 at 7.6 h; 0.30 at 16.4 h; H1 not supported). Discussion: Baseline covariates did not raise physiology's best value (0.574, chosen post hoc). Conclusion: Neonatal early-warning studies should report detection and warning time at fixed alarm budgets over monitored time, against random and baseline-risk comparators. Files: the manuscript (40 pages) and Supplementary Files 1 to 5 (169 pages). This is a preprint and has not been peer reviewed.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-12
DOI
https://doi.org/10.5281/zenodo.24173151
Primary Topic
Neonatal and Maternal Infections
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Early warning of late-onset sepsis at a fixed alarm budget: a benchmark of prediction models in 3,829 very-low-birth-weight infants from four NICUs

Ryan F. McCullough
Zenodo (CERN European Organization for Nuclear Research)
Neonatal and Maternal Infections
preprint

Early warning of late-onset sepsis at a fixed alarm budget: a benchmark of prediction models in 3,829 very-low-birth-weight infants from four NICUs

Ryan F. McCullough
preprint en

Abstract

Objective: To measure what neonatal early-warning models detect, and how early, at a fixed alarm budget. Materials and Methods: On the public PreMo cohort (3,829 very-low-birth-weight infants, four US NICUs), a benchmark scored detection and warning time at fixed alarm budgets over monitored time for four outcomes. Population (engineered-feature) and deep sequence models, frozen on selection infants, were scored once on 747 test infants. Results: At 10 alarms per 100 infant-days, the registered comparisons' four detection points detected 0.430 to 0.574 of monitored episodes of treated late-onset infection (162 of 225 evaluable test episodes had warning-window monitoring), against 0.056 for random alarms and 0.414 for baseline risk (floors chosen post hoc by the selection rule). Deep models did not detect more: the first deep arm trailed the unsearched population model (-0.144; 97.5% interval -0.233 to -0.051), and an extended search, a declared secondary, was inconclusive against a searched population model falling from 0.597 (selection) to 0.438 (test); every deep point estimate for the primary outcome trailed the unsearched model's (post hoc, extended search: -0.089; 95% interval -0.169 to -0.011). Conformal thresholds held the budget at three of four points as registered (two strictly). Among equally accurate models, higher detection came with shorter median warning (0.51 at 7.6 h; 0.30 at 16.4 h; H1 not supported). Discussion: Baseline covariates did not raise physiology's best value (0.574, chosen post hoc). Conclusion: Neonatal early-warning studies should report detection and warning time at fixed alarm budgets over monitored time, against random and baseline-risk comparators. Files: the manuscript (40 pages) and Supplementary Files 1 to 5 (169 pages). This is a preprint and has not been peer reviewed.

Zenodo (CERN European Organization for Nuclear Research)
Good health and well-being
Neonatal and Maternal Infections
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon