Peer-reviewed progress and cautious signals: AI in medicine this September

Peer-reviewed progress and cautious signals: AI in medicine this September

Mid-September 2026 delivered a series of concrete results for artificial intelligence in healthcare: a multinational validation of JLK's stroke-diagnosis AI published in the Journal of NeuroInterventional Surgery on September 14, an overview feature in USA TODAY the same day, and a new Nature Biotechnology study of Insilico Medicine's AI-developed drug candidate rentosertib. Together, they paint a picture of a field where diagnostics are showing measurable, peer-reviewed gains, while the drug and aging results remain early and partly exploratory.

Stroke diagnosis: the strongest evidence

The most solidly documented news this month is the validation of JLK's AI tool for detecting large vessel occlusions (LVO) on non-contrast CT scans — crucial, because such findings determine referral for thrombectomy.

The study, led by Dr. Chi Kyung Kim's team at Korea University Guro Hospital and collaborators, covered 963 patients in South Korea and the United States. The AI's diagnostic performance, measured as area under the ROC curve (AUC), was 0.963 in the Korean cohort and 0.899 in the American one. When clinicians read the images with AI assistance, their own accuracy rose from 0.718 to 0.852 AUC. Sensitivity — the ability to identify patients with a vessel occlusion — increased from 46.6% to 63.7%, while specificity rose from 91.9% to 94.9%. The researchers found no clear signs of automation bias, meaning clinicians blindly following the AI's suggestions (Medical Xpress, DOI: 10.1136/jnis-2026-025339).

Stroke treatment in practice: claims from UK authorities

The USA TODAY overview points to how such tools are already in use: England's NHS has deployed AI decision support for brain scans through the Brainomix e-Stroke system. According to the country's Department of Health and Social Care, as relayed by USA TODAY, the time from a patient's arrival at hospital to treatment has been cut by roughly an hour, and the number of stroke patients recovering with little or no disability has tripled since the rollout (USA TODAY).

These figures come from government reporting and have not been independently verified by AIMag; they should be read as reported outcomes from a real implementation program, not as the results of a controlled study.

Epilepsy: lesions radiologists missed

The USA TODAY article also looks beyond stroke: a 2025 study published in JAMA Neurology found that the AI system MELD Graph detected 64% of the epilepsy-causing lesions that radiologists had previously missed (USA TODAY). For patients with drug-resistant epilepsy, such a finding can mean the difference between continued suffering and possible surgical treatment — but the figure is known here through a secondary source, not the original study.

Drug development: slow and fallible

USA TODAY recalls the baseline: developing a new drug often takes a decade or more, and roughly 9 in 10 candidates fail in clinical trials. According to the FDA, as cited in the article, the use of AI has increased significantly across all phases of drug development in recent years.

The most concrete example this month is rentosertib, a candidate for idiopathic pulmonary fibrosis that was discovered and developed by Insilico Medicine using AI. A new study published in Nature Biotechnology, announced by the company and relayed by Pulse2, reports that the drug reduced estimated biological age as measured by six independently developed proteomic "aging clocks" (Pulse2). Note that this point rests on the company's own announcement, as reported by Pulse2 — not on neutral coverage.

The aging findings: exploratory, not proven aging reversal

It is tempting to read the results as "AI drug makes patients younger." The evidence does not support that conclusion. In the proteomic analysis, the clearest findings came at week four, with reductions of 2.71–3.46 years in the group receiving 60 mg once daily, measured on clocks based on chronological age (ScienceAlert). The central observation was the direction of change across all six clocks — not any single headline number. It is worth noting that the available secondary sources are not entirely in agreement about which dose group showed the strongest signal; Pulse2 reports the strongest effect in the 30 mg twice-daily group, while ScienceAlert highlights the 60 mg figures. AIMag cannot resolve this point of disagreement from the available evidence.

And the study cannot separate the two explanations: all participants had severe lung disease, and if rentosertib reduced the fibrosis or otherwise improved the lungs, that improvement alone could make the blood look biologically younger (ScienceAlert). Reductions in biological age are thus not documentation of halted aging or extended lifespan.

What the evidence actually supports

Taken together, this month's results point in one direction: AI-based image analysis is accumulating genuine, peer-reviewed clinical documentation — the JLK study shows measurably better clinician performance with no signs of blind trust in the machine. The British stroke figures and the MELD Graph finding point the same way, but are known through government and secondary sources respectively.

On the drug side, the picture is more open. That an AI-discovered candidate has reached clinical testing and is now showing consistent proteomic signals is real progress against a baseline where 9 in 10 candidates fail. But the aging findings in the rentosertib study are explicitly exploratory, the study cannot rule out that disease improvement explains the younger-looking blood, and the company's own announcements still require independent follow-up. The clear dividing line this September, then, is between diagnostic documentation that is beginning to measure up, and drug and aging claims that still await stricter validation.