AI beats traditional tools in early dengue diagnosis

Early diagnosis of arboviral infections remains a persistent clinical bottleneck, particularly in regions where healthcare systems must make rapid decisions with limited resources. A new study finds that artificial intelligence (AI) can outperform traditional diagnostic methods in early detection, though concerns over reliability and validation continue to limit its real-world use.

The study, titled Comparative Diagnostic Performance of Artificial Intelligence Versus Conventional Approaches for Early Detection of Mosquito-Borne Viral Infections: A Systematic Review and Meta-Analysis, with Evidence Predominantly from Dengue Studies and published in Machine Learning and Knowledge Extraction, found that AI and machine learning models improved sensitivity, specificity, accuracy, and negative predictive value over conventional approaches under random-effects analysis, while the evidence for area under the curve remained inconsistent and positive predictive value results were unstable.

The apparent gains are accompanied by extreme heterogeneity, widespread reliance on internal validation, and frequent methodological weaknesses in the analysis domain, all of which limit confidence in how well these tools would work once deployed across new health systems, new outbreaks, and new patient populations.

AI gains ground in early triage where conventional diagnosis remains weak

AI appears to improve several threshold-dependent diagnostic metrics that matter in early triage. Across pooled random-effects models, AI-based systems showed significant gains in sensitivity, specificity, accuracy, and negative predictive value compared with conventional approaches. Sensitivity improved with an effect size of 2.64, specificity with 5.55, accuracy with 3.19, and negative predictive value with 13.84, all with statistical significance except that gains in area under the curve did not hold consistently across the more conservative model.

That pattern is important because it points to where AI may actually be useful in practice. The authors argue that these models seem better suited to operational tasks such as early triage, case exclusion, and prioritizing who should receive confirmatory testing, rather than serving as definitive stand-alone diagnostic systems. In settings where large numbers of patients present with acute febrile illness and testing capacity is limited, a tool that safely rules out low-risk cases could ease pressure on clinics and laboratories.

The review found that the included evidence was dominated by dengue studies. Ten of the 13 studies focused primarily on dengue, while only one centered on chikungunya, one on Zika, and one assessed dengue, chikungunya, and Zika together. Most were hospital-based, though a few were community-based or multicenter, and nearly all relied on laboratory confirmation through RT-PCR or ELISA, sometimes combined with clinical criteria.

The models themselves were highly varied. Random forests, gradient boosting machines, neural networks, support vector machines, decision trees, naive Bayes systems, and deeper architectures all appeared in the evidence base. Clinical and laboratory variables were the most common predictor domains, each used in 10 of the 13 studies, followed by demographic or socio-economic variables in nine studies. By contrast, genomic, virological, entomological, and imaging data were absent from the included literature, underscoring how dependent this evidence base remains on routine point-of-care inputs rather than richer multimodal datasets.

The authors interpret the findings as evidence that AI models are not necessarily transforming global discrimination across all thresholds, but are changing performance at the decision points that matter clinically. That helps explain a central tension in the results: why several threshold-based measures improved while overall area under the curve did not show a robust random-effects advantage. In short, AI may be better at making useful calls at specific operating thresholds even when it does not radically reorder overall case-versus-noncase discrimination.

Stronger performance comes with major weaknesses in validation and bias

Although AI often outperformed conventional approaches on practical performance measures, the evidence base is fragile in ways that matter for real-world adoption. Substantial heterogeneity was present across outcomes, with I2 values reaching 100 percent in the pooled analyses, indicating that study-to-study differences were extreme rather than marginal.

That heterogeneity reflects more than statistical noise. The included studies varied widely in geography, study design, sample size, disease target, case definitions, feature sets, handling of missing data, threshold choices, and intended use. Most studies were conducted in Asia, especially Bangladesh and Thailand, with limited representation from Africa and the Americas. Only one study came from South America alone, and two were multi-country efforts.

Validation emerged as another major weakness. Every included study relied on internal validation only, such as train-test splits or cross-validation. None reported external validation in a way that would establish performance across independent settings. That makes it hard to know whether the reported gains would survive distributional shifts in prevalence, clinical presentation, laboratory practices, or data quality. The review also notes that calibration was rarely reported, with only two studies describing formal calibration procedures or goodness-of-fit assessments.

Risk-of-bias assessment reinforced those concerns. Using PROBAST, the authors found that participants, predictors, and outcome domains were usually low risk, but the analysis domain was the biggest source of trouble. More than half of studies were judged high risk in analysis, and overall most studies were classified as high risk. The paper links this to problems such as overfitting, handling of missing data, threshold optimization, and validation procedures.

Reporting quality also remained inconsistent. Interpretability or explainability was explicitly described in only a minority of studies, even though explainability is particularly important for clinical adoption. Most models were still effectively research tools or proof-of-concept systems, with no study reporting deployment in routine clinical practice.

Publication bias further complicates the picture. The review found no evidence of publication bias for some outcomes such as AUC, accuracy, specificity, and NPV, but signs of bias were detected for sensitivity and PPV. That matters because positive AI findings are more likely to be published, especially in a field where model development studies can generate attractive headline metrics from relatively small or highly controlled datasets.

AI looks most useful as a decision-support layer, not a replacement for clinical judgment

The authors' do not argue that AI should replace conventional diagnostics for mosquito-borne viral infections. Instead, they state that AI-based models may serve best as adjunctive decision-support tools, especially in the earliest phase of patient evaluation, when clinicians need help triaging febrile patients and deciding who can be deprioritized or who should move quickly to confirmatory testing.

This interpretation is consistent with how the pooled results behave. Gains in negative predictive value and sensitivity support a role for AI in ruling cases out and flagging potentially important ones early. But the instability in positive predictive value, especially under random-effects analysis, means these systems are far less suited to acting alone as rule-in tools. That is why the review stresses integration with high-specificity laboratory tests such as NS1 assays or RT-PCR inside sequential diagnostic pathways rather than letting AI outputs drive definitive conclusions by themselves.

The paper also notes that AI frequently outperformed conventional statistical comparators in sensitivity analyses restricted to logistic-regression-type models, particularly for accuracy, specificity, and NPV. Still, the magnitude and consistency of these advantages shifted depending on whether fixed-effects or random-effects assumptions were used. That divergence itself is revealing: it suggests the benefits of AI are highly context-dependent and may shrink once between-study variability is taken seriously.

For clinicians and health systems, the message is practical rather than revolutionary. AI may help in overcrowded or testing-constrained settings by sorting patients faster and improving early exclusion. For public health, it may support surveillance and outbreak response by strengthening first-line case identification. But the evidence is not yet strong enough to support broad, uncritical deployment, particularly in communities and outbreak settings that differ from the training environments used in model development.