Skip to main content
medicine

Reading the Platelet Curve: How AI Sharpens Dengue Triage

26 Jun 2026 · 7 min read · ASI Research Lab

Most dengue is survivable without much intervention. The danger is the minority of patients who slide into plasma leakage, severe bleeding, or shock, frequently around the time the fever breaks. In an outbreak that fills wards past capacity, the operational question is not diagnosis. It is triage: which patient, looking stable now, will deteriorate, and therefore needs the bed.

The clinical answer for two decades has been a checklist of warning signs. It is practical, and it is not very precise.

What the numbers say

Platelet count is the marker everyone watches, and the data behind that instinct is real. In one clinical series, patients with no warning signs had a median platelet count around 114,000 per microlitre on admission; those with warning signs, around 35,500; and those with severe dengue, around 25,000. Patients below 25,000 had roughly 7.5 times the odds of progressing to severe disease (clinical profile and platelet study, NCBI). The trajectory of that count over hours, not a single reading, is what carries the signal.

But warning signs as a binary rule perform poorly on their own. A systematic assessment found their sensitivity and specificity for predicting severe dengue are limited enough that relying on them alone produces both false alarms and missed cases (warning-signs accuracy, NCBI).

This is the gap machine learning has been shown to close. A validated 8-gene model predicted progression to severe dengue at 86.4% sensitivity and 79.7% specificity in an independent cohort, against 77.3% sensitivity and 39.7% specificity for warning signs at presentation — and cut the number of patients needed to predict one severe case by about 80% (8-gene model, NCBI). Using only routine inputs, a larger study in Puerto Rico reported gradient-boosting models reaching an area under the ROC curve around 97% on a full feature set (Puerto Rico study, NCBI). Low platelet count, abdominal pain, and vomiting recur as the top-ranked predictors across these studies.

Why this is the right kind of AI problem

Severity triage suits supervised learning for a precise reason: it is a ranking-and-prioritization task with an outcome that is recorded anyway. The model does not diagnose; it orders patients by risk so a clinician can allocate scarce attention. The output is a probability, not a verdict, and it sits beside a doctor rather than in front of one.

The limits that decide whether it is safe

The performance numbers above are encouraging and incomplete, and a research institution has to say so plainly.

Severity is population-specific. Dengue severity depends heavily on prior infection with a different serotype — the immunological reason a second dengue infection is often worse than the first. A model trained where one serotype dominated will not transfer cleanly to a population with a different exposure history. The South Asian setting our Advanced Medicine work targets is not interchangeable with the Latin American cohorts that produced much of the published evidence.

Validation has to be prospective. An AUC measured on the data a model was tuned on is a hypothesis. The number that licenses clinical use is performance on patients from a different hospital and a different season. Almost everything looks good in-sample.

The inputs must be available in the field. An 8-gene assay is powerful in a study and useless in a rural clinic without the lab to run it. A model built on a complete blood count and basic chemistry is less accurate and far more deployable. The right model is the one the clinic can actually feed.

Triage is the part of dengue care where the evidence for AI is strongest. It is also the part where a model that skips local validation does the most harm. Both things are true, and the second is the one that governs how we work.

Sources