Skip to main content
medicine

A Record Dengue Year: What 14 Million Cases Demand of Medical AI

22 Jun 2026 · 8 min read · ASI Research Lab

In 2024 the World Health Organization received reports of 14,434,584 dengue cases, including 7,718,585 laboratory-confirmed cases, 52,738 severe cases, and 11,201 deaths, across all six WHO regions and more than 100 countries. It was the highest global dengue burden ever recorded (WHO, Dengue – Global situation, 2024). The Region of the Americas accounted for more than 90 percent of cases, with Brazil alone reporting over 10 million. Those 2024 figures are the last complete year of global surveillance, and in 2026 they still define the scale of the problem.

Those are reported cases. The underlying burden is far larger: the widely cited modelling estimate puts true infections at around 390 million per year, of which roughly 96 million are clinically apparent (Bhatt et al., Nature, 2013). The gap between 14 million reported and 390 million estimated is itself the first thing to understand about dengue, and the first place data systems fail.

That record year is a useful stress test for a claim we take seriously: that AI belongs in this fight. It does — but only in specific places, and only under conditions that are easy to state and hard to meet.

Where AI genuinely helps

Severity triage. Most dengue infections are self-limiting. The clinical danger is the minority that progress to plasma leakage, severe bleeding, or shock, often around the time the fever breaks. The operational problem in an outbreak is not diagnosis; it is deciding which patients need a hospital bed before they deteriorate. Routine markers collected anyway — platelet count trajectories, haematocrit, and the timing of warning signs — carry predictive signal, and models trained on them can support triage when wards are overwhelmed. This is a ranking-and-prioritization task, which is exactly what supervised models are good at, and it sits alongside a clinician rather than replacing one.

Outbreak forecasting. Dengue transmission is strongly climate-sensitive: temperature, rainfall, and humidity drive the mosquito population and the viral replication rate, and the record surge tracked that sensitivity closely (World Economic Forum, 2024). Because the climate signal leads cases by weeks, models that combine meteorological data with historical case counts can give public-health units lead time to pre-position resources. The value here is days and weeks of warning, not perfect prediction.

Where the honest limits are

The limits are not subtle, and pretending otherwise is how medical AI loses trust.

Local data or nothing. A severity model trained on a Brazilian cohort does not transfer cleanly to South Asia. Serotype distribution, prior-infection history, comorbidities, and care pathways all differ, and dengue severity is notoriously dependent on prior exposure. A model is only as good as the population it learned from, which is why the work has to be done with local clinical data and local partners, not imported wholesale.

Prospective validation, not retrospective flattery. Almost any model looks good on the data it was tuned on. The number that matters is performance on patients it has never seen, ideally in a different year and a different hospital. Retrospective accuracy is a hypothesis; prospective accuracy is evidence.

Surveillance is the bottleneck. The 14-million-versus-390-million gap means the case data feeding any forecasting model is incomplete and biased toward places with testing capacity. AI does not fix under-reporting; it inherits it. Improving surveillance is often a higher-value intervention than improving the model.

What this means for our work

Dengue is one of the diseases our Advanced Medicine laboratory works on precisely because it is underserved relative to its burden — concentrated in tropical and subtropical regions, including South Asia, that receive a small share of global medical-AI investment. The discipline we hold ourselves to follows directly from the limits above: build on local data under ethics-board approval, validate prospectively before any clinical claim, and treat the model as a decision-support tool for clinicians, not a substitute for them.

A record year is not an argument for moving faster than the evidence. It is an argument for doing the unglamorous parts — data partnerships, validation, surveillance — properly, because that is where the lives are.

Sources