A working notebook from a neonatologist at the Wilhelmina Children's Hospital, UMC Utrecht. Read what I'm building, why I'm skeptical of half of it, and where the field is actually going.
In late October 2026 I'll stand on a stage in Lyon for twenty minutes and talk about big data, AI, and sepsis prediction in the NICU. I've been quietly uneasy about it ever since. Not about the speaking part, about what I'm supposed to say.
Read on Substack →A live PubMed search for late-onset sepsis prediction models, with Claude as the research partner. Seven academic groups, two systematic reviews, every query archived. The full evidence base for the series begins here.
Read on Substack →Four late-onset sepsis models, read side by side. The headline AUCs land in a tight band, but each one answers a different question. Similar, but not the same. The full eight-dimension comparison lives on the companion page.
Read on Substack →The same two decisions, turned onto my own paper. The detection fraction went in the abstract; the 4% precision stayed in Table 3. A self-critique of Berg 2023, read against its own supplement, the working file is on the companion page.
Read on Substack →A footnote we cited and walked past turns out to be the argument: a performance claim has to be good enough to catch the hazard it creates. Read against Episode 04's false-alarm rate, our own precision number stops being a metric and becomes a risk, the MDR classification, Rule 11, and the Epic Sepsis Model as cautionary tale are on the companion page.
Read on Substack →Does acting on the alarm at three in the morning leave the baby better off? That is a trial question, and it has been asked properly once, in a study that missed its own primary endpoint and is remembered for a secondary one. What a real clinical-impact trial of our model would cost is on the companion page.
Read on Substack →The device was already cleared before the 2011 trial began, so the trial was never commercially required. That removes the comfortable explanation for the missing second trial and leaves a harder one: the thing that can be sold does not need a trial, and the thing that needs a trial cannot be sold. The incentive structure, the enrolment arithmetic, and six obstacles that multiply are on the companion page.
Read on Substack →My first attempt at specifying a response protocol turned into thresholds, refractory periods and escalation rules, which is the model again in different clothes. The protocol is not something to design. It is a cascade that already runs several times a week, opened as readily by a nurse's concern as by an algorithm, and it separates into three decisions with different costs and different evidence behind each. The evidence map, the design contrast, and the gate where a model actually fires are on the companion page.
Read on Substack →Put a kitchen timer beside every incubator, set it for two hours, and examine the infant every time it rings. Routine care already runs that timer, and whatever a model adds, it has to add on top of it. Every paper in this series anchors onset to something recorded; our own audit puts documented evaluation closer to once a day than three. Whether a model replaces a look or adds one is on the companion page.
Read on Substack →I'm Daniel Vijlbrief, a neonatologist and clinical researcher at the Wilhelmina Children's Hospital, UMC Utrecht. I work on AI implementation in neonatal care, clinical decision support, sepsis prediction, and the unglamorous question of whether any of this actually changes what we do at the bedside.
Lately I've also been working with large language models in clinical workflow, including an AI-generated discharge letter now in routine use on our department, and the experiment you're reading now.
This site is where I think out loud.