AI models looked accurate for early dengue diagnosis—but the evidence rarely left the lab

The pooled numbers are encouraging—and easy to overread

AI models reported strong results for identifying mosquito-borne viral infections near a patient's first presentation, but a new peer-reviewed synthesis finds that the evidence is not ready to justify routine clinical use in hospitals and public-health services worldwide today. The reviewers searched three databases to 11 August 2025, screened 4,531 records and included 43 studies. Their random-effects meta-analysis estimated sensitivity at 86.1% and specificity at 90.2%; pooled area under the receiver-operating curve was 0.954.

Those summaries combine very different datasets, predictors, reference tests and validation methods. Statistical heterogeneity for the univariate outcomes ranged from 95.8% to 97.7%, meaning most variation was not sampling noise alone. The authors therefore treat the pooled figures as a description of published studies, not a performance promise for a clinic choosing a model today.[1]

This is mainly dengue evidence from Asian hospitals

Forty-one of the 43 studies assessed dengue. Thirty-five involved Asian settings, including 12 studies from India, while evidence for Zika and chikungunya was sparse. The search found no eligible model study for West Nile virus, yellow fever or Rift Valley fever. Results should therefore not be stretched across pathogens whose symptoms, prevalence, laboratory pathways and patient populations differ.

Twenty studies used retrospective diagnostic-accuracy designs and ten modelled existing datasets. Only five were clearly prospective: three diagnostic-accuracy studies, one cohort and one modelling study. Most drew on clinical and laboratory variables available around assessment, and 37 used laboratory confirmation in at least part of the reference standard. Yet only a few studies used genuinely separate temporal or cross-dataset validation.[1]

Bias, calibration and workflow evidence are the missing pieces

Using the PROBAST prediction-model tool, the reviewers rated 30 studies—69.8%—at high overall risk of bias, seven at low risk and six unclear. Missing-data and class-imbalance handling were often poorly reported. Only four studies presented a formal or semi-formal calibration check. Calibration asks whether a stated risk behaves like that risk in practice; good discrimination alone cannot tell a clinician whether a threshold will produce an acceptable number of false alarms or misses.

Forty studies remained research prototypes and three were pilots or proofs of concept; none reported routine deployment. The review also found no consistent evidence that one model family outperformed another. A useful next study would freeze a model and threshold, test it prospectively across several countries, report calibration and subgroup performance, and compare ordinary care with an AI-supported pathway. It should count delayed or missed diagnoses, unnecessary testing, clinician workload and patient outcomes—not only retrospective accuracy.

The review was registered prospectively, used duplicate screening and extraction, and its authors reported no specific funding or competing interests. Its own limits include English-language, peer-reviewed publications and the possibility of publication bias. The careful conclusion is not that AI has failed, but that early diagnostic promise has outrun the evidence needed for dependable use across outbreaks and health systems.[1]