The 16 Out of 237

A scoping review published in Tropical Medicine & International Health in June 2026 mapped 237 studies of AI-enhanced point-of-care diagnostics for infectious diseases in resource-limited settings, published between 2015 and early 2026. Only 16 of 237 studies reported prospective field validation, and 62% did not report a validation level at all. The rest live in the literature as demonstrations of what works in a dataset. The clinician who is supposed to use these tools sees both numbers and knows which one matters.

The gap is not academic. The 2025-2026 period accounts for 44% of all included studies, which means the publication rate is exponential and the field-validation rate is not. Every month brings another imaging model for malaria microscopy, another convolutional net trained on COVID chest films, another lateral-flow interpreter that claims laboratory-grade sensitivity when the technician holds the phone at the right angle. None of that changes the operational question: who puts their name on the report.

Rigorous validation under CLIA guidance, including analytical accuracy, clinical performance in intended-use settings, and usability testing, remains central to ensuring reliability in decentralized environments. That sentence describes the standard. The scoping review describes the corpus. The distance between them is the distance between a conference poster and a test you are willing to run on a febrile patient two hours from the nearest lab when the result determines whether they get antibiotics tonight or a referral tomorrow.

The studies that did reach prospective field validation share a pattern. Malaria microscopy and COVID-19 molecular diagnostics are the most densely studied combinations, which is another way of saying the validation effort followed the funding and the funding followed the headlines. A rapid test for malaria in a clinic without a microscope has an immediate referent: the technician who used to squint at a slide for twenty minutes can now photograph it, and the model produces a result the supervising clinician can compare to their own read or to a PCR send-out. The workflow is legible, the error modes are known, and the performance metric has a ground truth.

Most of the 221 other studies do not have that. They have sensitivity and specificity measured against a test set that may or may not reflect the population the device will see, and a cross-sectional review of 691 FDA-cleared AI medical devices found that only 1.6% cited data from a randomized clinical trial and fewer than 1% reported actual patient health outcomes. Regulatory clearance happened, but clearance is a narrow gate. It asks whether the thing does what the submission said it would do in the setting the submission described. It does not ask whether a nurse practitioner in a rural health post will use it the same way a research coordinator did during the pivotal study, and it does not ask what happens when the patient demographic shifts or the phone is two Android versions behind or the lighting is worse.

Image-based diagnostics remain foundational to women's health point-of-care testing, yet traditional microscopy is subjective and frequently inaccurate, contributing to high rates of inappropriate treatment. As care increasingly shifts beyond specialty practices into decentralized settings, improving diagnostic accuracy at the moment of care is imperative. AI-assisted digital microscopy can responsibly augment clinician judgment, supported by real-world deployment insights and a practical regulatory pathway from LDT toward CLIA-waiver. That is the promise. The scoping review is the scorecard.

The structural problem is that validation in the field is expensive, slow, and unrewarding from a publication standpoint. A model trained on 18,000 annotated images and tested on a held-out set of 4,000 produces a paper. A six-month deployment in three clinics with 200 patients, continuous quality monitoring, and a comparison arm produces operational data and regulatory evidence, but it does not produce a methods section that fits the format most journals want. The journal wants the architecture diagram and the loss curve. The regulator and the clinician want to know whether the thing drifted when the rainy season started and half the samples came in with condensation on the cassette.

Only 26% of the 237 studies report any cost information. Cost is the other half of the decision. A test that requires a $1,200 smartphone, stable connectivity, and a monthly SaaS fee is a different procurement conversation than a $40 handheld with a one-time capital cost and no recurring license. The studies that omit cost are not lying, they are optimizing for a different performance function. They want to demonstrate technical feasibility. The clinician who signs the purchase order needs to know the five-year total and who pays for the calibration consumables.

The FDA has now cleared more than 1,000 AI medical devices, and by early 2026, about 95% to 97% went through 510(k) rather than De Novo or PMA. That pathway asks for substantial equivalence to an existing device, not a new standard of evidence. It is faster and it is appropriate for incremental improvements, but it also means the evidentiary bar for most of these tools is lower than the bar a hospital applies to a new chemotherapy protocol. The clinician at the point of care inherits that gap.

What the clinician needs is not another paper showing 96.2% sensitivity in a dataset. It is a deployment guide that names the failure modes, a validation protocol they can run locally with 50 samples, and a support path that does not require them to debug a Python dependency on their own time. Models require local tuning to avoid bias or performance drops across scanners and populations, and continuous monitoring and governance frameworks are essential. That is the operational reality. The 221 studies without field validation do not address it.

The review authors are careful not to overstate the finding. They describe a field producing studies at an exponential pace and ask why so few convert into prospective field-validated tools. The answer is not mysterious. Field validation requires a partner site, ethical approval in the deployment country, a comparator arm, outcome tracking, and someone who will still answer the phone eight months in when the model starts flagging every sample as indeterminate because the lot number on the reagent strips changed. None of those things appear on a conference leaderboard.

The gap will close, but it will close on the supply side, not the demand side. The studies will keep coming because the models are cheap to train and the datasets are growing and the publication incentive has not changed. The validation bottleneck is elsewhere. It is in the regulatory capacity of the agencies that approve diagnostics for low-resource settings, in the capital budgets of the clinics that would deploy them, and in the liability framework that governs what happens when an AI-flagged negative turns out to be a false negative and the patient was sent home.

Until that bottleneck clears, the clinician at the point of care will keep seeing two numbers. 237 tools in the literature. 16 with the evidence that would let them use it without holding their breath. The second number is the one that determines what goes into the protocol, and it is growing slower than anyone publishing the first number wants to admit.

Tarry Singh is the founder and CEO of Real AI (realai.eu), an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan (earthscan.io) for Energy AI, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.