AI for neglected tropical diseases is advancing faster than the evidence behind it
Artificial intelligence is moving into neglected tropical disease research with striking speed, but the clinical foundation remains thin. A systematic review of prediction-model studies in leprosy, leishmaniasis, and schistosomiasis found a field dominated by diagnostic tools, a sparse focus on prognosis, and a level of methodological fragility that would be hard to ignore in any other area of medicine.
Among 69 studies published between 2015 and 2025, 89.9% appeared after 2020. The publication surge reflects how quickly machine learning has entered the conversation around endemic infections, yet the review makes clear that volume has not translated into robustness. In practical terms, the models are proliferating faster than the safeguards needed to know whether they would work outside the datasets that trained them.
Diagnostic use cases dominate while prognostic questions remain underused
The clearest pattern in the review was the imbalance between diagnosis and prognosis. Diagnostic support accounted for 81.2% of the studies, while only 18.8% examined prognosis or adverse outcomes. That split matters because early detection is only one part of care in these diseases; clinicians also need tools that anticipate deterioration, complications, treatment failure, or other downstream harms.
The current emphasis on diagnosis is understandable. Image-based recognition, pattern detection in clinical records, and classification tasks are natural entry points for AI, especially when access to specialists is limited in endemic settings. But prognosis is where a model begins to intersect with resource allocation, triage, and follow-up intensity. Without credible prognostic models, the technology remains better at labeling a case than helping decide what happens next.
High risk of bias was the rule, not the exception
Methodological quality was the central weakness. Of the 69 studies, 89.9% of models were judged to have a high risk of bias using PROBAST+AI. The dominant reasons were straightforward and familiar: 82.6% lacked external validation, and 66.7% failed to report or appropriately handle missing data.
External validation is the difference between an interesting model and a clinically believable one. A tool trained in one hospital, one region, or one patient mix can look impressive in development and collapse when exposed to different epidemiology, different recording practices, or different disease prevalence. In neglected tropical diseases, where geography, access to care, coinfections, and socioeconomic context shape presentation, that weakness is especially damaging. A model that never leaves its original dataset may be measuring local quirks rather than disease biology.
Missing data handling was another major fault line. In real-world clinical records, missing values are not a rare nuisance; they are part of the operating environment. Poor handling of missingness can distort model performance, inflate apparent accuracy, and conceal the uncertainty clinicians need to see before trusting an automated output. When the data pipeline is fragile, the final prediction can look more precise than it deserves to be.
One sentence captures the problem: most of these models were never tested hard enough.
Transparency lagged behind model development
The review also identified a reproducibility problem. Source code was unavailable in 76.8% of cases, creating a barrier to independent verification and reuse. That absence is more than an academic inconvenience. Without code, outside investigators cannot inspect preprocessing steps, compare model architectures, reproduce the analysis, or determine whether performance depends on choices hidden from the published report.
For a field that aims to serve resource-limited health systems, this is a serious contradiction. Open science is not a luxury add-on to AI research; it is one of the few ways to separate a promising algorithm from a result that cannot be trusted, scaled, or audited. Code availability also matters for adaptation. Endemic regions differ in lab infrastructure, imaging capacity, digital records, and case definitions, and a model that cannot be examined cannot be realistically adjusted for those differences.
Why the burden of proof is higher in endemic settings
The diseases in this review are not rare curiosities. They remain major public health problems in communities facing socioeconomic vulnerability, and the article places them in the context of more than one billion people affected globally by neglected tropical diseases overall. That burden creates genuine pressure to find tools that can extend clinical reach where specialist access is limited.
AI can, in principle, help with that challenge because it handles high-dimensional data and can detect nonlinear relationships that ordinary methods may miss. In infectious diseases shaped by clinical, demographic, environmental, and social variables, such models may eventually support screening, triage, and surveillance. But the same complexity also increases the chance of false confidence. If the underlying datasets are narrow, the outcome labels inconsistent, or the validation weak, the model may encode local bias rather than transferable biology.
A second sentence stands out for the bedside reality it implies: a model that fails external validation is not ready for patient care.
Open science and external validation should define the next phase
The review points toward a simple research agenda. Future studies need external validation, transparent handling of missing data, and shared code that allows other teams to reproduce the work. Those are not cosmetic improvements. They are the minimum requirements for moving from proof-of-concept to something that can influence diagnosis or prognosis in the settings where these diseases actually occur.
The article’s own data availability statement points to a GitHub repository containing the data and code for the systematic review, but the broader literature it examined has not yet adopted that standard consistently. Until that changes, the field will continue to generate many models and too little confidence. The larger problem is not a shortage of algorithms. It is a shortage of evidence that those algorithms will survive contact with real clinics, real populations, and the variability that defines neglected tropical disease care.