From big data to big everything.

Why this page exists.

Episode 11 of Road to CEPAS 2026 is deliberately selective. The Substack post does not survey omics in neonatal sepsis; it asks which biological layer can help with which of the three decisions from Episode 08, and on what clock. Omics is also the title of the session in Lyon.

This page holds the studies behind each layer in one place, with population, size and result as reported, and sets each layer against the gate it could serve. It adds one thing the post leaves implicit: how far each study population sits from a preterm infant on a NICU at the moment of suspicion.

A good test that arrives late answers the wrong question.

The three gates from Episode 08: at gate 1 the alarm fires and we decide whether to examine the infant; at gate 2, having examined, whether to take a culture and start antibiotics; at gate 3, 36 to 48 hours later, whether to stop. Each layer of biology runs on its own clock, and the clock decides which gate it can reach.

Who is at risk this week, not who is becoming ill tonight.

Two kinds of evidence. The first shows that the organism in the blood often lived in the gut first. The second shows that the gut community itself changes before sepsis begins.

The volatiles are the exception in this table: one of the few body-derived signals with any lead time at all, though on small numbers, with modest performance and a daily laboratory step.

What is the body fighting?

At gate 2 a blood sample already exists. The question is whether there is more in it to read than a CRP. The host-RNA evidence is strongest where it is furthest from our population, and thins out as it approaches a preterm infant on a NICU.

Infected or not, rather than bacterial or viral.

On a NICU the question at gate 2 is rarely bacterial versus viral. It is whether a deteriorating infant is infected at all. Viruses do turn up: in a two-centre US study, a respiratory virus was found in 8 of 135 late-onset sepsis evaluations, 6% (Ronchi 2014). That is why the metabolomics study, which compares sepsis with non-infectious inflammation, sits closest to our question despite a lower AUC than the transcript scores.

Different signals for different decisions, in practice.

Rotterdam implemented heart-rate-variability monitoring to trigger an evaluation, followed by inflammatory markers to decide what happens next. Across 1,135 infants under 32 weeks before and after the change, sepsis severity (nSOFA course) was lower afterwards, with more biomarker blood tests but no significant increase in blood cultures or antibiotic use (Kurul 2022). A before-and-after design cannot establish cause. It is also a sequential protocol, physiology then conventional biomarkers, not a combined prediction model with a molecular layer.

Fast enough for gate 3, and finding a lot.

Instead of the infant's response, look for the bacterium itself: PCR, broad-range 16S sequencing, microbial cell-free DNA, metagenomics. Results within hours to a day put these tests at gate 3, the decision to stop.

• Oeser 2020, UK. 208 newborns evaluated for early-onset sepsis. Culture positive in five, all also found by PCR. PCR detected bacteria in 91 of the 203 culture-negative samples (45%) (Oeser 2020).
• Sauberan 2026, San Diego. Plasma microbial cell-free DNA detected an organism in 10 of 20 well-appearing hospitalised preterm infants (Sauberan 2026).
• CHAMPIONS, Netherlands. Multicentre prospective study of Molecular Culture, a rapid broad-range PCR, against blood culture in infants under 90 days, explicitly to stop or avoid antibiotics sooner. Protocol published; results not yet found (Groen 2024).

Two readings fit the positives equally well. Blood culture misses genuine infections, plausible given the volumes we can draw. Or bacterial DNA passes through the circulation of preterm infants without causing disease, plausible given the gut evidence above. The available studies do not separate the two.

The largest item in the inventory, and mostly a different question.

• Diagnostic yield. A meta-analysis of 23 studies (1,567 critically ill infants) pooled a diagnostic rate of 0.37 for rapid whole-genome sequencing, 0.42 across all rapid genomic sequencing (Xiao 2022).
• Clock. In the first NHS-commissioned trio rapid-genome service, in Wales, mean time to report was nine days (Sloper 2024).
• Susceptibility. In the NICHD Neonatal Research Network, concordance of LOS did not differ between same-sex and unlike-sex VLBW twin pairs: no evidence that susceptibility is strongly genetically determined (Boghossian 2013).

More measurements per infant, not more episodes.

Episode 10 concluded that the limit for sepsis AI is the number of labelled episodes, not bytes. Omics makes the ratio more extreme.

• The arithmetic. The Valencia study measured about 28,000 transcripts in 36 infants: 28,000 ÷ 36 ≈ 780 measurements per infant. With ratios like that, finding a combination that separates two small groups is not difficult.
• The familiar drop. A four-gene signature measured at birth in term infants in The Gambia predicted early-onset sepsis with AUC 0.94 in training and 0.72 in an external cohort (An 2024). A very different population from ours, the same lesson as Episode 03.
• What sparse models can and cannot do. They can reduce thousands of candidate features to a handful. They do not create more sepsis episodes.

Support each decision with what is available at that moment.

The post's conclusion is not one large sepsis score built from every layer. It is decision support per gate, carrying information from the slower layers into the faster ones.

The gap, at search strength.

The post looked for studies combining continuous physiology with a molecular or microbial layer to predict late-onset sepsis and did not find one. Kurul 2022 is the nearest neighbour, and it pairs physiology with conventional inflammatory markers in sequence rather than in a model. This was a literature search, not a systematic review, so absence from it is not evidence of absence. If the gap is real, it is an interesting one.

What this episode settled, and what it still owes.

• Settled: the thread from Episode 10. The infant's own chemistry, set up at the end of Episode 10, is followed layer by layer and tied to the three gates.
• Search-strength only: two absences. Gut information used to tune a physiological alarm, and a model combining continuous physiology with a molecular layer for LOS. Neither was found; neither search was systematic.
• From abstracts: the numbers. All figures were taken from abstracts, not checked against every full text.
• Pending: CHAMPIONS. The trial closest to gate 3 has a published protocol and no results found yet.
• Still open from Episode 10: notes as the gate-1 timestamp. A proposal, not yet done.
• Closed without being discharged: the money. The funding instruments promised in Episode 08, drafted and cut in Episode 09, will not be covered in this series. Recorded here so the promise is closed in the open rather than left hanging.
• Next: the presentation. This is the last essay. Episode 12 is the talk for Lyon taking shape.
• Standing: the audit is unpublished. The gate-1 audit and the alarm-versus-evaluation comparison still appear in general terms only.