The four late-onset sepsis models, compared across eight dimensions.

Why this page exists.

Episode 03 of Road to CEPAS 2026 is a close read of the four anchor papers in Branch A — continuous-physiology machine learning for late-onset sepsis (LOS) in preterm infants. The Substack post argues that the headline AUCs are similar but not the same, and that the two decisions which matter most — signal stack and alarm policy — are the two the literature documents least clearly.

This page is the working file behind that argument: the full eight-dimension comparison, read across Berg 2023 (WKZ Utrecht), Kausch 2023 (UVA and two partner NICUs), Yang 2024 (Eindhoven/Máxima) and Meeus 2024 (Antwerp/Innocens). Three supporting-cast papers from the Episode 02 map — Honoré, Leon, Rio — sit at the foot. Full citations and licensing are in the references.

The four papers at a glance.

Cohort design.

• Berg — Retrospective, single-centre, 2008–2019. Every consecutively admitted infant GA ≤32 wk regardless of admission reason: the broadest, least selected population of the four. Controls drawn from negative-culture and out-of-window patients, matched on GA and time-since-birth.
• Kausch — Retrospective, three NICUs, 2012–2021. Restricted to VLBW (<1500 g) — a narrower, sicker slice. Strict definition: positive culture after day 3, ≥5 days antibiotics, ≥2 antibiotic-free days before.
• Yang — Retrospective, single-centre, 2016–2018. The smallest and most selected cohort: 51 LOS + 68 controls. Clean cases, clean controls — ideal for modelling, least representative of a real ward.
• Meeus — Retrospective, single-centre, 2012–2020. Outcome is LOS and NEC jointly — 206 combined episodes. Not a LOS-only paper.

Signal stack.

Really three decisions in one: which signals, sampled how fast, from which device.

• Berg — Two signals, 1/min: HR and SpO2 only. Deliberately excluded CRP and blood pressure (clinician-initiated, leakage risk) and temperature and RR (artefact-prone).
• Kausch — Two signals plus their cross-correlation, 0.5 Hz. Key finding: performance is essentially unchanged whether HR comes from ECG or pulse oximetry — the model can run on a standalone oximeter.
• Yang — Three signals, 1 Hz, 102 features. The sampling-rate ablation: raw waveform 0.886 → 1 Hz 0.875 → 1/min 0.825 → 1/h 0.687. Minute-by-minute is the deployable floor; hourly destroys the signal.
• Meeus — Seven signals, 1/min — the widest stack. But breadth does not produce a cleanly higher comparable AUC.

ML methodology.

• Berg — LR vs GAM vs XGBoost; all within 0.01 AUC, no significant difference by DeLong's test. Chose LR for interpretability.
• Kausch — LR (cubic splines) vs neural net vs XGBoost vs random forest; similar across all. Chose LR — "equal performance, better explainability."
• Yang — Seven methods including two deep-learning architectures. XGBoost won; deep learning underperformed.
• Meeus — XGBoost only, no bake-off. SHAP for explainability.

Validation rigour.

• Berg — 75/25 patient-grouped split, bootstrap CIs, longitudinal alarm simulation. Single-centre; no external validation.
• Kausch — The field's strongest: trained on one NICU, externally validated on two others. TRIPOD-reported, calibration assessed per site. External drop ≈0.03 AUC.
• Yang — 7-fold cross-validation, nested for hyperparameters. Single-centre; no external validation — named as the main limitation.
• Meeus — Cross-validation plus a temporal hold-out (year-2020 patients). Still single-centre; generalisability stated as the main limitation.

Performance.

Each number, with its conditions attached — that is the point.

• Berg 0.79 — cross-sectional AUC at t=0 (0.73 train / 0.79 test), whole-NICU, culture-positive LOS. Longitudinal: ≈58–67% recall under the multi-threshold policy; ≈47–60% of patients detected pre-clinically.
• Kausch ≈0.79 — prediction within 24 h, VLBW-only, external. Pulse-rate ≈ ECG. Risk rises significantly 23–24 h before blood culture.
• Yang 0.875 — measured 6 h before CRASH, on 119 hand-cleaned infants. Patient-wise sensitivity 96.1%, but specificity 19.1%.
• Meeus 0.93 — for LOS+NEC jointly, on cross-validated time windows (a metric Meeus flag as optimistic on imbalanced data). Sliding-window: 69% all / 81% severe; median time gain ≈10 h.

Alarm policy.

What converts an AUC into something a clinician actually experiences — chosen four different ways, rarely justified.

• Berg — Hourly scores; thresholds set to fixed unit-wide rates. 8-hour refractory period (a nursing shift); a muted alarm re-fires only if a higher threshold is crossed. Burden reported as alarms per patient-day.
• Kausch — An alert switches on at a threshold and stays on until 24 h with no further crossing. No refractory period, no escalation.
• Yang — Adopts Berg's framework wholesale — same 8-hour silencing, same multi-threshold escalation — and cites it.
• Meeus — Hourly probabilities aggregated into 24-hour buckets; reports alarm-days per week, echoing Berg's burden accounting without the refractory machinery.

Honest limitations.

• Berg — Imputed blood-culture timestamps; culture-positive-only labelling; blood-culture time as a suspicion proxy may overestimate lead time; may fire on non-sepsis deterioration.
• Kausch — Missing-data exclusions; culture-proven-only definition; co-morbidities alter features; a prospective multicentre study still needed.
• Yang — Selected, non-representative cohort; maturation difference between groups; small dataset; no external validation; deep learning underperformed.
• Meeus — LOS labelling is error-prone; retrospective-labelling risk; no generalisability data; AUROC optimistic on imbalanced data; metrics are backward-looking.

Unique contributions.

• Berg — Largest unselected single-centre cohort; the most extensive retrospective clinical-impact simulation in the LOS literature; the alarm-fatigue framework that became a field convention.
• Kausch — The only true multi-NICU external validation; the pulse-rate-equals-ECG finding; HR–SpO2 cross-correlation as a feature; a direct benchmark against the legacy HRC index.
• Yang — The sampling-rate ablation; the seven-method bake-off showing deep learning does not help; concrete evidence of Berg's alarm framework propagating.
• Meeus — Joint LOS+NEC prediction; the widest signal stack; the only commercial spinoff (Innocens BV); a pointed critique of AUROC on imbalanced data.

Supporting cast.

Three further Branch A groups mapped in Episode 02, included in the close read as context rather than as anchors.

• Honoré 2023 — Karolinska / KTH. 325 infants, only 20 LOS cases; Naïve Bayes, chosen to avoid overfitting a tiny positive set; AUROC 0.82 up to 24 h before suspicion.
• Leon 2021 — Rennes. 49 infants; HRV from ECG only, visibility-graph features; AUROC 0.877 in the 6 h before antibiotics. The methodologically distinctive single-signal end of the spectrum.
• Rio 2022 — Lausanne. Not a model paper: an independent validation of the commercial HeRO/HRC index. Strongly gestational-age-dependent (sensitivity 76% at <28 wk vs 25% at >32 wk); optimal cutoff 2.76, not the FDA-cleared 2.0.