From the stage.

The last episode of Road to CEPAS 2026. On a Saturday morning in Lyon, the Omics in Sepsis session put three talks side by side: thirty years of neonatal biomarkers, big data and AI on the NICU, and reading the immune system. They started in different places and ended at the same one. This page is the working file behind the post, with the slides, the animation there was no time for, and the papers behind the other two talks.

Why this page exists.

Episode 12 of Road to CEPAS 2026 is the finale. It merges the two episodes that were planned around the talk into one Substack post: my own talk in short, the two other speakers, the discussion, and a wrap-up of the series. There is no Episode 13.

The session, Omics in Sepsis at CEPAS 2026, was chaired by Hans Jürgen Dornbusch (Austria) and Chiara Minotti (Switzerland). The speakers were Jan Janota (Prague, Czechia), me, and Ruud Nijman (Imperial College London, UK). This page keeps three kinds of material apart: what was on my slides, what is in the published papers the other speakers drew on, and what was shown as conference data that has not yet been published.

A note on sources. The other two talks are paraphrased from a written account of the session, never quoted; audience questioners are not named. Every figure attributed to a paper was checked against its abstract as indexed in PubMed, not against the full text. Figures labelled conference data were presented in Lyon and could not be found on PubMed; they are reported as presented and carry no reference. My own talk is summarised from the slides.

The talk, as given

Big data, AI and sepsis prediction: opportunities in the NICU.

Fifteen minutes, thirteen slides and an animation, one argument carried through the series since Episode 01: the field can predict late-onset sepsis in preterm infants reasonably well, and has not yet built the bridges from an alarm to a better outcome. The talk opened with a show of hands on who had used AI to prepare a poster or talk for the congress, and closed on a call to share models and data.

Slides · PDF · 14 pages, 0.3 MB

The deck included an animation on the care round that was skipped for time. It follows one preterm infant from the monitor stream to an alarm, then through the three gates. It is here instead. It runs 47 seconds and has a music track.

Late-onset sepsis, from signal to decision. One preterm infant's heart rate and saturation; tens of millions of values over two months, most never kept; a risk score crossing a threshold before clinical recognition; the alarm entering gate 1 alongside a nurse's concern, a parent's concern, more apnoeas and an incidental CRP; most alarms ending in a look, not antibiotics. Proportions in the gate sequence are illustrative.

Table 1. The figures on the slides and where each comes from. Every one was built in an earlier episode; this table only points back to where.
Slide	Figure	Source	Built in
4 · One infant, two months	~26 million monitor values; ~5 GB ECG at 500 Hz; ~90,000 words of notes; >1 TB of video in the few NICUs that keep it	Order-of-magnitude estimates for a 60-day stay	Ep 10
4 · Events, not bytes	47,187 infants of 22–28 weeks; 67% had at least one LOS evaluation; 10% of evaluations were culture-positive	Speier 2023 [1]	Ep 10
5 · Heart rate and saturation	Half to two-thirds of late-onset episodes flagged before clinical recognition; AUC 0.78–0.88 across groups, in different cohorts, outcomes and horizons	Seven late-onset sepsis groups	Ep 02, 03
6 · Our own model	At least 47% of infants with sepsis flagged before the blood culture; ~4% of alarms matched a true episode. Retrospective, one centre, 2008–2019	van den Berg 2023 [2]: 47% in the abstract, ~4% from Table 3	Ep 04
7 · Four bridges	Bytes to labelled episodes; one unit to every unit; alarm to the right decision; decision to proven benefit	The series' own frame	Ep 08–11
8 · Three gates	Look, treat, stop. Evidence at gate 1: none found (literature search, not a systematic review)	Sola 2020; Stocker 2017; van Leeuwen 2024; Feng 2025	Ep 08
9 · The care round	Replace or add; every model paper anchors onset to a documented action, none counts the look before it	Masino 2019; Cabrera-Quiros 2021; Mithal 2018; Mani 2014; van den Berg 2023	Ep 09
10–11 · Omics	Each layer on its own clock; ~780 transcripts per infant studied; AUC 0.94 → 0.72 on external validation; 0.79 for five metabolites plus IL-6	Cernada 2014; An 2024; Thangavelu 2025	Ep 11
12 · To the bedside	One randomised trial in 15 years: 3,003 VLBW infants, primary endpoint missed, mortality 10.2% to 8.1% as a secondary outcome. Likely MDR Class IIb (a reasoned reading). An adult sepsis model at AUC 0.63 externally	Moorman 2011 [3]; Kausch 2023; Peng 2025; Wong 2021; Koppens 2023	Ep 03, 05, 06

Takeaway

Every number in the talk was already on this site before Lyon. The slides added nothing new; they put eleven episodes into fifteen minutes and ended on the same gap: the bridges, not the data or the models.

Jan Janota · biomarkers

Thirty years, and no single marker.

Jan Janota spoke mainly about early-onset sepsis, where a reliable marker would decide who gets antibiotics in the first days of life. Publications on neonatal sepsis biomarkers are still rising year on year. His group placed more than 250 biomarkers tested in neonates over three decades onto a map of the inflammatory cascade. None meets all the classic criteria for an ideal marker.

His conclusion, in paraphrase: stop looking for one ideal marker, keep looking for markers that add something to a combined system of markers, clinical data and AI. He added that cost, speed and complexity will test AI-driven multi-omic systems as much as they tested single markers.

Published: NeoPInS, a machine-learning analysis of 1,685 neonates of more than 34 weeks treated for suspected early-onset sepsis. Biomarkers, mainly CRP and white cell count, beat risk factors and clinical signs for culture-proven sepsis. The full model reached an AUROC of 83% and an AUPRC of 28%; risk factors alone performed close to random [4]. Janota is a co-author.
Conference data: the map of more than 250 neonatal sepsis biomarkers on the inflammatory cascade. Not found on PubMed.

Takeaway

The gap between an AUROC of 83% and an AUPRC of 28% is the same gap this series found in physiology models: good discrimination, and most positive calls still wrong when the condition is rare.

Ruud Nijman · host response

Reading the immune system instead of hunting the bug.

Ruud Nijman described a strategy that reads the host response rather than searching for the pathogen. His group works on signatures for bacterial versus viral infection and for several specific diseases, from samples ranging from blood to breath, urine and stool. Most of his talk was about endotypes: subgroups of sepsis defined by clinical data, biomarkers or gene expression, and whether they agree.

Table 2. Sepsis subtypes in adults and children, as reported in the abstracts. Adult schemes first.
	Built from	Population and size	Finding
Seymour 2019 [5]	29 clinical variables in the health record	Adults, 20,189 in derivation; 43,086 in validation; three RCTs	Four phenotypes, α to δ; 28-day mortality from 5% (α) to 40% (δ)
van Amstel 2023 [6]	Four existing schemes: clinical, biomarker, two transcriptomic	522 critically ill adults	No clear relationship between any of the schemes (Cramér's V 0.086–0.456)
Scicluna 2025 [7]	Blood transcriptome, three classification methods aligned	Adults, MARS (678) and GAinS (444 at admission); tested in VANISH (176) and Uganda (128)	Three consensus subtypes, CTS1–3; a harm signal with corticosteroids in CTS2
Wong 2016 [8]	100-gene signature plus a protein-biomarker risk model	288 children with septic shock	Endotypes A and B; see Table 3
Schlapbach 2024 [9]	Whole-blood RNA sequencing; two ten-gene signatures	907 children evaluated for sepsis (RAPIDS); EUCLIDS external set, 362	Internal AUC 94.1% for bacterial vs viral, 82.2% for organ dysfunction within 24 h; about 70% in EUCLIDS
Dunican 2026 [10]	RNA velocity of whole blood (VeloCD)	Several cohorts, including febrile children	From one sample, predicted the direction of illness, including which febrile children needed most care

Two reading notes. The paediatric validation of the consensus subtypes was presented in Lyon but is not in the Scicluna 2025 abstract. And the EUCLIDS figure of about 70% is for the two signatures used together, predicting organ dysfunction at the time of sampling.

Table 3. Corticosteroids by subtype. All three are post hoc analyses; adult and paediatric schemes are different classifications and cannot be mapped onto each other.
	Population	Subtype	Association
Antcliffe 2019 [11]	176 adults in septic shock (VANISH)	SRS2, relatively immunocompetent	Hydrocortisone with higher mortality: OR 7.9 (95% CI 1.6–39.9)
Scicluna 2025 [7]	Pseudo-randomised cohort and VANISH reanalysis	CTS2	A harmful signal with corticosteroids
Wong 2016 [8]	288 children with septic shock	Endotype B at intermediate to high baseline risk	Corticosteroids with a lower risk of a complicated course: RR 0.09 (0.01–0.54). Complicated course = two or more organ failures at day 7, or death by day 28; not mortality alone

Also shown, as conference data not found on PubMed:

a four-gene severity signature for children in the emergency department, about 900 children, AUC above 0.9 as presented;
point-of-care formats for host-response signatures: a handheld lateral-flow device intended for a trial in The Gambia, an RNA lab-on-a-chip with a result in about 15 minutes, and a commercial bacterial-versus-viral signature in about 40 minutes;
a planned trial at Imperial of real-time endotyping to guide steroid treatment;
a proof of principle in 22 children combining several omic layers.

Takeaway

Four ways of subtyping the same adults found four different populations. Where subtypes do look useful, it is because a treatment decision changes with them, which is the series' argument from the other side: the signal matters through the decision it changes.

The discussion

The clinician is still the comparator.

Paraphrased; audience questioners are not named.

The best tool. Janota held that an experienced clinician who sees the infant repeatedly is probably still the best instrument there is. He is working on a unified definition of neonatal sepsis.
Calculators travel badly. On the early-onset sepsis calculator: it reduces overtreatment where it was built, but could increase treatment in countries where incidence is lower.
Signatures overlap. Nijman noted that critical-illness signatures overlap with burns, trauma and ARDS, and that neonatal signatures still struggle to separate infection from other causes, especially after hypoxic-ischaemic encephalopathy. My point: at the bedside the question is what to do at this moment.
Context shapes the response. Maternal infection, milk and mode of delivery all change the newborn's inflammatory response; Janota's group has seen differences after vaginal birth compared with caesarean section.
Low-resource settings. I suggested prevention may matter more there than models, and that a mobile phone can carry a tool. Nijman described work in Zimbabwe, AI on phones and AI-guided ultrasound for pneumonia, and noted that global health funding is shrinking.

What the three talks share

Different starting points, one destination.

Table 4. Six threads that ran through all three talks. The synthesis is the post's, not a view stated by the other speakers.
	Where it showed up
No single number	No ideal marker after more than 250 candidates; a model as one input to gate 1 alongside a nurse's and a parent's concern; subtyping schemes that disagree. All three talks ended at combined systems.
The definition underneath	Prediction models anchor onset to the blood culture; endotypes were derived while sepsis definitions changed; a unified neonatal definition is still being worked on.
Time	Omic layers run on different clocks; RNA velocity adds direction to a single sample; cost and speed test any multi-omic system.
The decision	Antibiotics in the first days of life; look, treat and stop; corticosteroids by subtype, with a trial of real-time endotyping planned.
The comparator	The experienced clinician who sees the infant repeatedly; the care round that puts hands on every infant every few hours.
Scale	Performance falls in a new unit for physiology models, and from 82–94% internally to about 70% externally for gene signatures; a calculator that reduces treatment where it was built may increase it elsewhere.

Takeaway

External validation lowers performance for physiology and for gene signatures alike. More layers per infant do not replace more infants, more centres and more time.

Twelve episodes

The series, end to end.

Table 5. Each episode, the question it asked and where its working file lives. Episode 01 has no companion page.
Episode	Question	Working file
01 The Honest Problem	Why does prediction not become utility?
02 Mapping the Field	Who builds these models?	Search log
03 A Field Full of ROC Curves	Do the headline AUCs mean the same thing?	Comparison grid
04 Our Own Work	What does our own paper actually show?	Methodology log
05 The Model Leaves the Lab	What does regulation ask before the bedside?	Methodology log
06 From Alarm to Better Outcomes	What would a clinical-impact trial take?	Methodology log
07 Who Would Run It	Who pays for the trial nobody can sell?	Methodology log
08 Three Gates	What does an alarm actually open?	Methodology log
09 The Comparator	What does a model compete with?	Methodology log
10 Big Data in the NICU	What data exist, and what is labelled?	Methodology log
11 From Big Data to Big Everything	Which biological layer serves which gate?	Methodology log
12 From the Stage	Where did three talks end up?	This page

The ledger

What the series settled, and what it leaves open.

Settled: prediction is not utility. Data, cases and models exist. What is missing are the four bridges: from bytes to labelled episodes, from one unit to many, from an alarm to the right decision at each of the three gates, and from that decision to proven benefit.
Settled: the care round is the comparator. A model either chooses better than the nurse who looks anyway, or adds looks that come early in the right infants (Episode 09).
Settled: omics layers run on their own clocks. Each can serve a different gate (Episode 11), and external validation lowers performance for physiology and gene signatures alike.
Open: a trial of the response protocol. Testing what happens after the alarm, not the model itself.
Open: multicentre external and prospective validation of late-onset sepsis models.
Open: CHAMPIONS. The trial closest to gate 3 has a published protocol and no results found yet.
Open: the internal gate-1 work. The audit and the alarm-versus-evaluation comparison remain unpublished and appear in general terms only.
Open: a unified neonatal sepsis definition, the work Janota described.
Search-strength only. The absence of evidence at gate 1 and of a model combining continuous physiology with a molecular layer come from literature searches, not systematic reviews.
Closed without being discharged: the money. The funding instruments raised in Episode 08 were closed in Episode 11 without being covered.

Takeaway

No unit and no discipline can build these bridges alone. The model is available, and every search, paper and comparison behind the talk is on /sepsis.

How this page was built

The last working file.

The inputs were a written account of the session and the slide deck. Claude asked for the names of chairs and speakers rather than guessing them, matched the papers mentioned in the other talks to PubMed records, and checked every figure attributed to a paper against its abstract. Where a figure could not be found in a publication, it is labelled conference data. I decided that this would be the final episode, that the other speakers would be paraphrased, which papers to include and how to frame the synthesis. Claude drafted; I edited and approved.

References & sourcing.

Speier RL, Cotten CM, Benjamin DK, et al. Late-onset sepsis evaluation and empiric therapy in extremely low gestational age newborns. J Pediatric Infect Dis Soc 2023;12(Suppl 2):S37–S43. doi:10.1093/jpids/piad068 PMC · PMID 38146858
van den Berg MAM, Medina OOAG, Loohuis IIP, et al. Development and clinical impact assessment of a machine-learning model for early prediction of late-onset sepsis. Comput Biol Med 2023;163:107156. doi:10.1016/j.compbiomed.2023.107156 PubMed · PMID 37369173
Moorman JR, Carlo WA, Kattwinkel J, et al. Mortality reduction by heart rate characteristic monitoring in very low birth weight neonates: a randomized trial. J Pediatr 2011;159(6):900–906.e1. doi:10.1016/j.jpeds.2011.06.044 PMC · PMID 21864846
Stocker M, Daunhawer I, van Herk W, et al. Machine learning used to compare the diagnostic accuracy of risk factors, clinical signs and biomarkers and to develop a new prediction model for neonatal early-onset sepsis. Pediatr Infect Dis J 2022;41(3):248–254. doi:10.1097/INF.0000000000003344 PubMed · PMID 34508027
Seymour CW, Kennedy JN, Wang S, et al. Derivation, validation, and potential treatment implications of novel clinical phenotypes for sepsis. JAMA 2019;321(20):2003–2017. doi:10.1001/jama.2019.5791 PMC · PMID 31104070
van Amstel RBE, Kennedy JN, Scicluna BP, et al. Uncovering heterogeneity in sepsis: a comparative analysis of subphenotypes. Intensive Care Med 2023;49(11):1360–1369. doi:10.1007/s00134-023-07239-w PMC · PMID 37851064
Scicluna BP, Cano-Gamez K, Burnham KL, et al. A consensus blood transcriptomic framework for sepsis. Nat Med 2025;31(12):4119–4130. doi:10.1038/s41591-025-03964-5 PMC · PMID 41028542
Wong HR, Atkinson SJ, Cvijanovich NZ, et al. Combining prognostic and predictive enrichment strategies to identify children with septic shock responsive to corticosteroids. Crit Care Med 2016;44(10):e1000–e1003. doi:10.1097/CCM.0000000000001833 PMC · PMID 27270179
Schlapbach LJ, Ganesamoorthy D, Wilson C, et al. Host gene expression signatures to identify infection type and organ dysfunction in children evaluated for sepsis: a multicentre cohort study. Lancet Child Adolesc Health 2024;8(5):325–338. doi:10.1016/S2352-4642(24)00017-8 PubMed · PMID 38513681
Dunican C, Wilson C, Habgood-Coote D, et al. Predicting trajectories of illness using RNA velocity of whole blood. Nat Commun 2026;17(1). doi:10.1038/s41467-026-71685-5 PMC · PMID 42091589
Antcliffe DB, Burnham KL, Al-Beidh F, et al. Transcriptomic signatures in sepsis and a differential response to steroids: from the VANISH randomized trial. Am J Respir Crit Care Med 2019;199(8):980–986. doi:10.1164/rccm.201807-1419OC PMC · PMID 30365341

All 11 citations were resolved against PubMed; sample sizes, proportions, AUCs, odds ratios and relative risks above were checked against the abstracts, not the full texts. “PMC” marks a free full text in PubMed Central; “PubMed” marks an indexed abstract only. Papers named on the slides without a number here (Table 1) are referenced on the episode pages linked beside them. Figures labelled conference data were presented in Lyon and are not publications. The synthesis in Table 4 and the ledger is the post's reading, not a finding of any cited study or a view stated by the other speakers.