Proof Must Reach the Patient

August put a hard question beside the promise of medical AI: can a tool change an outcome that matters to the person in the clinic?

Retrospective, authored 16 October 2026. This article focuses on signals recorded in August; it does not claim these observations were published then.

The question in the waiting room

A clinician opens a scan on a crowded afternoon. An algorithm marks a patch of grey and gives it a number. The patient is not waiting for a score. She is waiting to learn whether someone has noticed a problem in time, whether the referral will happen, and whether the next step is affordable.

Medical AI is often evaluated on a clean technical question: did a system classify the image correctly? August's evidence raised the harder one: what changed for the people whose care passed through it?

On 19 August, researchers in PLOS Digital Health published an analysis of 1,357 FDA-cleared or approved AI medical devices. They found that 34 had links to registered prospective trials, 12 had results posted, and three evaluated patient-centred outcomes such as mortality, morbidity or readmission. The finding is not that every device fails, or that the authors proved no benefit exists. It is that device authorisation and evidence of improved outcomes are different things. The study and its methods are available here.

What the signal meant

The denominator matters. The study examined devices cleared through 5 December 2025, using the FDA device database and searches of trial registries and published papers. Its authors found a thin chain from clearance to prospective evaluation to reported patient outcomes. Because this was an evidence review, its conclusions depend on its search and linkage methods; unlinked evidence may exist. The paper itself also notes that evidence limitations are not unique to AI devices.

At the same time, the regulator was confronting a new class of product. On 18 August the US Food and Drug Administration issued a discussion paper on generative-AI-enabled medical devices. It asked for public input on risk assessment, premarket evaluation, postmarket monitoring, foundation models and agentic systems. The proposed approach included competency assessment, non-clinical benchmarks and clinical confirmation. This was a discussion paper, not a final rule. Its importance lay in making explicit that a model with variable outputs may need a different kind of scrutiny from a fixed calculator. The FDA's announcement and the paper itself set out the questions. The FDA's AI-enabled device catalogue provides a separate view of the authorised-device landscape; authorisation is a regulatory status, not a substitute for evidence of benefit in a local care pathway.

These two signals fit together without being identical. The PLOS analysis looks backward at authorised devices and the evidence connected to them. The FDA paper looks forward at how generative systems might be assessed. One counts a gap; the other opens a consultation about what an adequate process could become.

The math: the evidence funnel

Take the study's three devices with patient-centred outcome evaluation and divide by the 1,357 devices in its analysis: 3/1,357 = about 0.22 per cent. That is the share identified by the authors, rounded to about 0.2 per cent. It is not an estimate that only 0.2 per cent of devices work. It measures the presence of a particular kind of published or registered evidence under the study's method.

Now view the preceding steps. Thirty-four trial links out of 1,357 is 2.5 per cent; twelve with results or peer-reviewed publications is 0.9 per cent. Each stage narrows the visible evidence base. The useful question for a clinic is not simply, “Is this tool cleared?” It is: what population was evaluated, what outcome was measured, what happens when the model is wrong, and who receives the alert?

The local test

In Kenya, a validation study cannot be a US dataset with a local flag added afterwards. Patients may speak Kiswahili, Sheng, Somali, Kikuyu, Luo or several languages in one encounter. Clinics vary in staffing, connectivity, equipment and referral options. A tool that works in a tertiary hospital can fail in a dispensary because the workflow around it is different.

That is not an argument against using international evidence. It is a reason to ask what travels. A safe evaluation should define a task before deployment, include the people who will use and be affected by it, measure consequences that matter to patients, and record where the system abstains or sends a case to a human. If a model suggests a referral, did the patient reach the referral facility? If it helps summarise a consultation, did the clinician save time without losing a relevant detail? These are concrete outcomes, and they belong in the evaluation plan.

The counterargument is real: prospective clinical studies take time and money. A rare outcome may require a large sample; a model may change before a trial ends. Strict evidence demands can preserve the advantage of large institutions and slow tools that could help underserved patients. The answer is proportionate evidence: evaluate risk, let lower-risk administrative tasks use lighter review, and demand stronger clinical confirmation where a wrong answer can delay urgent care. The FDA discussion paper itself framed risk and competency as questions for feedback rather than a settled solution.

Pawanax builds clinical AI for doctors and a WhatsApp health agent for families, so this is not an abstract standard. The product question is whether a family gets a clearer next step and a clinician a safer, more complete handoff. Any benefit claim should be tied to a measurable change in that journey, not a benchmark alone.

The next 90 days

The FDA's comment period is scheduled to close on 19 October. That date is a process milestone, not evidence that a final framework has been adopted. The registered forecasts most relevant to Pawanax are F6 (80%), whether WHO's Bundibugyo Ebola PHEIC remains in force on 31 January 2027, and F7 (20%), whether Kenya confirms a locally acquired case by that date. They are subjective, uncalibrated probabilities published on 16 October, and separate epidemiological questions rather than predictions about clinical AI. Their outcomes will be scored at the stated date with (p − outcome)².

The lesson for the next quarter is narrower and more useful: count the evidence chain from model to workflow to patient outcome. Keep a record of the population, version and failure path. Let the patient, not the press release, be the unit at the end of the chain.

Read the World Ledger

All 30 articles. Previous: The Machine Has a Power Bill. Next: A Rain Forecast Is Also a Clinic Forecast.

Mathematical methods: 45 parameters, assumptions and limits. Download the research and executable model.

AI-assisted research and writing using parallel Luna agents; editorial synthesis under the APEX lens. Historical scenes are illustrative. Sources support the stated observations; forward scenarios and inherited probabilities remain conditional, uncalibrated judgments. Kenya Vision 2060 is a founder-authored thesis, not an adopted government plan.

Cover image: Doctor using ultrasound machine in medical office.jpg — Shixart1985 · CC BY 2.0. Photographed 20 September 2026; location not stated on Commons. Illustrative clinical image, not a record of a particular 2026 case. Resized for web display.