The study behind our sepsis model was called Big Data for Small Babies. This page is the working file behind taking that phrase literally: how much data one very preterm infant produces over an admission, how much of it is kept, which datasets hold the infants a sepsis model needs, and what AI can do with each layer.
Episode 10 of Road to CEPAS 2026 is an inventory rather than an argument. The Substack post starts with a thought experiment, one very preterm infant staying about two months, and counts what that stay generates, layer by layer. It then sets the per-infant numbers against the datasets that actually exist, and asks what each kind of data can do for late-onset sepsis.
This page holds the arithmetic and its assumptions, the dataset landscape in one table, the four uses of the data the post walks through, and the sourcing. It also connects the notes to the comparator problem of Episode 09, because that is where they matter most.
Each layer of a two-month admission, with the assumption behind its size and whether it usually survives. The order follows the post: monitor, waveforms, charting, laboratory, notes, imaging, video, research data.
| Assumption | Size over the stay | Usually kept? | |
|---|---|---|---|
| Monitor numerics, 1 Hz | HR, SpO2, RR, BP, temperature; 86,400 values per signal per day | ~5.2 million per signal; ~26 million for five signals | Rarely at this resolution |
| Monitor numerics, per minute | 1,440 values per signal per day; the resolution our 2023 model used | ~86,000 per signal | Yes, in the record |
| ECG waveform | One lead at 500 samples per second | ~2.6 billion samples, ~5 GB | Rarely |
| All waveforms | ECG plus plethysmography and respiration at lower rates | ~5–10 GB | Rarely |
| Structured charting | Hourly nursing observations, medication, fluid balance, ventilator settings | Tens of thousands of entries; little storage | Yes |
| Laboratory | Blood gases, counts, CRP, electrolytes, renal function, triglycerides; the blood cultures | ~1,000–2,000 values | Yes |
| Clinical notes | Five a day, ~300 words each: medical, nursing, consultations | ~90,000 words: a novel about one baby | Yes |
| Imaging | Serial cranial ultrasound, chest radiographs, perhaps a term-equivalent MRI | ~1 GB | Yes |
| Video | Continuous camera at modest quality, ~22 GB a day | > 1 TB: more than everything above combined | Few NICUs record |
| Research layers | Stool samples analysed by eNose; a whole genome | Tens of GB; ~100 GB | Research only |
Most of the data disappears: the waveform, the camera image and the continuous physiology exist, but what survives is minute-level numerics, charting, laboratory results and notes. And there is an enormous amount of data per infant with very few events in it.
What teaches a sepsis model is the labelled episode: the blood culture, the organism and the timestamp. The scale of that label is set by the largest cohort the post cites. In 47,187 infants of 22–28 weeks from 243 centres in the Pediatrix network, 2009–2018 (Speier et al. 2023):
For this kind of machine learning the number of infants matters more than the number of bytes. A billion heartbeats from one infant do not compensate for the absence of a second infant.
The datasets the post names, set side by side. They trade resolution per infant against number of infants, and that trade decides which sepsis question each can answer.
| What it holds | Scale | What it can answer for sepsis | |
|---|---|---|---|
| WKZ cohort · van den Berg 2023 | Minute-by-minute HR and SpO2 from the EHR; ≤ 32 weeks, 2008–2019; single centre | 292 LOS, 1,497 matched controls | Early warning from physiology, in one unit |
| UK NNRD · Chehrazi 2026 | Daily clinical data from every NHS neonatal unit in England and Wales | 26,132 extremely and 55,789 very preterm admissions, 2013–2023 | What happens next, at national scale; no minute-level physiology |
| Pediatrix data warehouse · Speier 2023 | Multicentre clinical data, US | 47,187 infants, 243 centres | How often the cascade opens, and how often it finds an organism |
| MIMIC-III · Johnson 2016 | Single Boston hospital; vital signs, labs, notes; waveforms for a subset | 7,870 neonates, 2001–2008 | Openly accessible method development; one centre, older practice |
| PreMo · Mesner 2026 (preprint) | Continuous HR and SpO2 with clinical events and outcomes; four US NICUs | 3,829 VLBW infants | External validation of physiology models without first building a consortium |
| Artemis · Blount 2010 | Platform for real-time analysis of monitor streams, IBM and SickKids | Piloted in one NICU | Keeping the disappearing waveforms by analysing them as they arrive |
| NeonatalBERT · Xie 2025 | Language model pre-trained on neonatal clinical notes, Stanford; external testing at a second centre | 32,321 newborns (primary); 7,061 (external) | Risk of 19 morbidities, sepsis among them, from text |
A sophisticated model built on beautifully sampled physiology from one centre usually has no easy route to finding out whether it works elsewhere. A multicentre physiology dataset does not solve that, but it changes what an ordinary research group can attempt.
| Question | The catch | |
|---|---|---|
| Physiology · early warning | Does heart rate and saturation change before the infant looks septic? | A model from one NICU learns one population, one set of monitors, one record and one way of practising. Some of the signal may be local |
| Registries · the cascade at scale | What does the existing cascade look like, before a model claims to shrink it? | No timing of deterioration. In England and Wales, 96% of very preterm infants who survived to discharge received at least one antibiotic (Saunders 2024) |
| Notes · when suspicion began | When did someone first become worried? | Used as predictors, notes let a model reconstruct suspicion that already exists. Models trained on clinician-initiated data alone came close to full-record models across 42.9 million admissions (Beaulieu-Jones 2021) |
| Large datasets · target trial emulation | Which questions are worth taking into a real trial? | Cannot evaluate an alarm and response protocol that has never existed. Applied in the UK to probiotic strategies with late-onset sepsis as a primary outcome (Chehrazi 2026, protocol) |
If an infant changes at 03:00, the first trace may not be a laboratory value or a code. It may be a nurse writing that the infant is less responsive, has more apnoeas, looks mottled or “is just not himself”, with the culture taken an hour later. The post's proposal is to use the notes to reconstruct the moment suspicion first arose, rather than feeding them to the prediction model.
That is the comparator Episode 09 found missing. Every model paper in this series anchors onset to something recorded, usually the culture, and the internal audit showed that most documented evaluations never lead to one. The notes are where gate 1 leaves a trace.
An alarm that precedes the blood culture by six hours sounds impressive. An alarm that precedes a nurse writing “baby looks septic” by twelve minutes is a very different proposition. Only the notes can tell the two apart retrospectively.
Stripped of the phrase, big data in the NICU is concrete: a monitor samples physiology, the record stores what we did, the laboratory what we measured, and the notes what we noticed and thought. Most of the highest-frequency data is discarded, and the one thing a sepsis model needs most, the labelled event, is rare. The opportunity the post names is connecting these layers across centres, so that rare events add up to something a model can learn from.
Every layer observes the infant from the outside: electrical activity and light absorption, treatments and decisions, the occasional blood sample, and the infant as seen by doctors and nurses. None of them sees the infant's own chemistry continuously.
The episode delivers the landscape it promised, and adds one proposal that ties it back to the comparator. The money remains the longest-running open item in the series.
Citations 1–4 and 6–10 were resolved against PubMed; the cohort sizes, centre counts and percentages above were checked against the abstracts. Citation 5 is a preprint outside PubMed and was checked by the author. The per-infant inventory in Table 1 is arithmetic on the assumptions stated with it, not a measurement. Figures for our own model are as reported in the 2023 paper. The internal audit referred to from Episode 09 is unpublished and appears only in general terms. The reading of the notes as a gate-1 timestamp is the author's own, offered as a proposal rather than a finding of any cited study.