Three AI advances in medicine in one week — with very different levels of solidity

Three AI advances in medicine in one week — with very different levels of solidity

The week of September 7–14, 2026 brought three dated milestones: a peer-reviewed study of AI-assisted stroke diagnosis, a published phase 2a analysis of an AI-designed drug, and a company-declared breakthrough for a brain–computer interface. Together, they offer a concrete but differentiated picture of what AI is actually delivering in clinical medicine right now — and where the line runs between documented effect and corporate claims.

The most solid first: AI support for stroke diagnosis

The most robust of the three developments is a multinational study published in the Journal of NeuroInterventional Surgery, reported on September 14 via Medical Xpress based on a press release from Korea University College of Medicine (Medical Xpress). The study included 963 patients — 723 from Korea and 240 from the US — and tested JLK's AI system on ordinary, non-contrast CT. On its own, the system achieved an AUC of 0.963 in the Korean dataset and 0.899 in the American one for detecting large vessel occlusions, the condition requiring acute thrombectomy.

Most relevant for clinical practice: clinicians' own reading performance measurably improved with AI support. Without AI, diagnostic accuracy was 0.718 AUC; with AI it rose to 0.852. According to the study, there were no clear signs of automation bias — clinicians blindly following the machine's assessment — a well-known concern with decision-support systems.

As context, not as direct evidence for this particular system: England's Department of Health (via USA Today) reports that the number of stroke patients recovering with no or minimal disability has tripled since the NHS introduced the Brainomix e-Stroke system. The figure is attributed to the department and cannot be credited to any single technology, but it points in the same direction: AI-assisted image diagnostics is today the best-documented clinical contribution from AI in medicine.

Image diagnostics is not the only diagnostic track, incidentally: researchers in the Netherlands have trained AI on 88 speech features to distinguish people with schizophrenia from healthy controls, with 86.2 percent accuracy on audio recordings of new patients the model had never heard before (Scientific American/Knowable, September 12). This is early research, but it illustrates that AI diagnostics is moving beyond images.

Promising, but confounded: rentosertib and biological age

The same week, on September 7, Nature Biotechnology published phase 2a results for rentosertib, a drug developed by Insilico Medicine (reported by Medical News Today). The analysis included 42 participants from an earlier trial in idiopathic pulmonary fibrosis (originally 71 people) and measured nearly 3,000 proteins in the blood. In the group receiving 60 mg once daily, six different proteomic "aging clocks" showed a reduction in estimated biological age of 2.71 to 3.46 years after four weeks (ScienceAlert).

It is tempting to read this as aging reversal. It should not be read that way. The study itself points to a central confounding factor: if rentosertib reduced fibrosis or otherwise improved the lung disease, that improvement alone could have made the blood appear biologically younger. The analysis failed to distinguish between the two explanations. The result is therefore exploratory and says nothing about whether the drug reverses aging or extends lifespan.

That the drug has come this far at all is nonetheless remarkable. Insilico states (via Pulse2) that the target protein TNIK was identified using the company's AI platform and the molecule designed with Chemistry42, and that rentosertib is now in phase III for pulmonary fibrosis in China. These are the company's own claims, but phase III status, if accurate, is a concrete indication that AI-driven drug design has moved from concept to late-stage clinical development.

Worth noting: coverage is not fully consistent on the dose. ScienceAlert gives 60 mg once daily as the group with the clearest signal, Pulse2 mentions 30 mg twice daily, and Medical News Today refers to roughly "3–4 years." None of the available sources resolves this against the original paper, so the figures should be read with caution.

Declared, not independently verified: Paradromics' speech interface

On September 14, Paradromics announced that the first participant in its Connect-One study — conducted under an FDA-approved Investigational Device Exemption — has used the Connexus brain–computer interface for real-time speech and text, including during a live phone call, following implantation in June at University of Michigan Health (press release via Business Wire/AOL).

The New York Times (Pam Belluck, September 14) covered the case the same day, including that the wireless system could produce speech when the woman thought about what she wanted to say, without moving her speech muscles, and that the study is planned to include ten patients. However, there is no independent clinical verification of the result in the available sources — both the press release and the NYT report rely on the company's own information. The news is therefore significant as an indicator of where the field stands, but should not be presented as documented clinical effect.

What the week actually shows

Taken together, the three stories point to a clear pattern: diagnostic AI today has the strongest evidence base — peer-reviewed, multinational, with a measurable lift in clinicians' own accuracy. AI-driven drug design is real and progressing, but concrete health claims (such as "younger blood") run ahead of what the studies can actually document. And neuro-interfaces remain in small, company-reported studies.

For readers following AI in health care, the practical conclusion is simple: weigh the type of source. Peer review with patient data carries more weight than press releases, and exploratory results — especially with known confounding factors — are not the same as established clinical effect.