AI in Healthcare: Facial Screening, FDA Oversight Gaps, and the Demand for Disclosure

A week of contrasts in health AI: striking demonstrations of what the technology can detect, and a sobering audit of how little we know about whether it actually helps patients. Plus new moves on regulation, safety, and public trust.

A 5-Second Selfie Video Could Screen for Hypertension and Diabetes

Researchers from the University of Tokyo presented findings at ESC Congress 2026 showing that an AI model can analyze subtle signals in a brief facial video — blood flow patterns, micro-movements, skin color variations — and identify hypertension with more than 90% accuracy and diabetes with more than 80% accuracy.

The system requires only a smartphone camera and five seconds of video, making it a potential low-cost screening tool for settings where lab tests and specialist visits are scarce. Primary care clinics across the Mediterranean and the developing world are among the obvious beneficiaries: hypertension often goes undetected for years simply because testing is inconvenient or inaccessible.

The results are preliminary and presented at a conference, meaning full peer review is still in progress. But the directional finding — that AI can extract clinically meaningful signals from ordinary video — points to a category of non-invasive, low-friction diagnostics that could eventually integrate into telehealth platforms and community screening programs.

FDA-Cleared AI Devices: Approved, But Not Proven to Help Patients

A study published in PLOS Digital Health on August 19 examined 1,357 medical AI devices cleared by the FDA through the 510(k) pathway — and found a stark gap between regulatory approval and evidence of patient benefit.

Only three of those devices, roughly 0.2%, were linked to studies evaluating patient-centered outcomes such as mortality, hospital readmission rates, or quality of life. Only 2.5% had any registered prospective clinical trial on record. The rest were cleared based on technical performance metrics — accuracy against a reference standard, for example — without evidence that they change what happens to patients.

This is not a niche concern. The 510(k) pathway was designed to speed clearance for devices deemed "substantially equivalent" to already-approved products. It was not designed to demonstrate that a device improves care. As AI devices multiply, many are being deployed in hospitals and clinics with no trial evidence that they benefit — or even that they don't harm — the people they are used on.

FDA Moves Toward a Formal Framework for Generative AI in Medicine

Separate from the clearance-gap study, the FDA published a discussion paper on August 18 signaling a formal policy shift for generative AI medical devices — large language models, multimodal systems, and similar tools used in clinical decision support.

The agency is moving away from case-by-case review toward a systematic framework covering three areas: premarket evaluation standards, risk classification, and postmarket surveillance. Public comments on the discussion paper are due October 19, 2026.

The timing matters. GenAI tools are already being used in hospitals for documentation, radiology interpretation, and clinical decision support — largely under existing guidance that was written long before large language models existed. The FDA is acknowledging the mismatch and beginning to build a framework fit for purpose. Clinicians, developers, and health systems have until mid-October to shape what that framework looks like.

ECRI Builds an Error-Reporting System Specifically for AI

The patient safety nonprofit ECRI has expanded its Problem Reporting Network to specifically capture AI-related errors and near-misses in clinical settings. The move follows a survey of 124 healthcare quality and safety leaders: 31% reported encountering an incorrect or misleading AI output in the past year, and 9% said an AI error had actually reached a patient.

Healthcare has well-established pipelines for tracking drug errors, device failures, and surgical complications. AI-related incidents have largely fallen through the cracks — an AI tool that quietly generates a wrong dosage suggestion or misclassifies an imaging finding may never appear in any adverse event database.

ECRI's expansion creates a structured channel for clinicians to report AI errors the same way they report other patient safety events. The data collected will inform safety guidance and, over time, provide a clearer empirical picture of where AI tools fail in practice — the kind of evidence base regulators and hospitals will need to make better deployment decisions.

Public Expects Transparency: 8 in 10 Americans Want AI Disclosed

A Pew Research Center survey conducted June 22–28, 2026 found that 80% of Americans want their healthcare provider to tell them when AI is being used in their care. Yet only 16% say their doctor has disclosed such use, and 46% are not sure whether AI has been involved in their treatment at all.

Privacy concerns are prominent, particularly around AI health chatbots that may not fall under HIPAA protections — patients using consumer tools to research symptoms may not realize their inputs could feed commercial model training or be shared with third parties.

The survey lands at a relevant moment. The EU AI Act includes transparency requirements for high-risk AI systems, and EU member states are implementing those requirements now. Patients have a legal right to know when consequential AI is influencing decisions about their care, and the Pew data makes clear they want that right exercised even where no mandate yet applies.

PISA: A New Tool Catches What Genomic AI Is Really Learning

Finally, a methodological advance from the Stowers Institute and Kundaje Lab at Stanford, published in Nature Communications in August 2026. The tool, called PISA (pairwise influence by sequence attribution), maps what a genomic deep-learning model actually attends to when making predictions — base by base, across the full genome.

Researchers built PISA to understand their models, rather than trust them as black boxes. What they found was more alarming than expected: PISA exposed a hidden experimental bias that had been silently corrupting results. The model was learning artifacts in the data — technical noise from the sequencing protocol — rather than genuine biological signal. On standard benchmarks it looked fine; PISA revealed the problem.

The finding has two implications. First, it is a reminder that AI models trained on biological data can and do learn spurious patterns that appear valid on held-out test sets but reflect measurement artifacts rather than biology. Second, PISA is itself a practical audit tool — now publicly available for other genomics labs to examine whether their own models are learning what they think they are.