Where AI Is Already Changing Medicine — and Where the Results Remain Uncertain

Where AI Is Already Changing Medicine — and Where the Results Remain Uncertain

In a single week in September 2026, a peer-reviewed validation of AI-assisted stroke diagnosis landed, a study of an AI-developed drug reported exploratory "aging clock" findings, and a company announced a brain implant that gave a patient back her voice — a useful moment for sorting which claims about AI in healthcare are well supported, and which are not.

The Week That Produced Three Benchmarks

The past few days have delivered three kinds of news about AI in clinical medicine. A research team from Korea University College of Medicine published a multinational validation of AI-assisted stroke diagnosis in the Journal of NeuroInterventional Surgery (Medical Xpress). Insilico Medicine announced a study in Nature Biotechnology on its AI-developed drug rentosertib, with exploratory findings on reduced biological age (Pulse2). And Paradromics announced that the first patient in its clinical trial of a wireless brain implant can converse in real time (New York Times).

The three cases represent different levels of evidentiary strength: peer-reviewed validation, early exploratory findings, and a company's own statement without independent verification. Distinguishing them from one another is a useful exercise for anyone who has to evaluate future headlines about AI in medicine.

Stroke Diagnosis: The Strongest Evidence

The best-documented development comes from imaging diagnostics. The Korea University team validated an algorithm that detects large vessel occlusions on non-contrast CT — a scan that is fast and widely available, but harder to interpret than contrast CT. In a validation across 963 patients, the algorithm's AUC was 0.963 in the Korean cohort and 0.899 in the US cohort (Medical Xpress).

More relevant for practice: when eight clinicians read CT scans without AI assistance, their diagnostic accuracy was 0.718 AUC. With AI support it rose to 0.852, sensitivity increased from 46.6 to 63.7 percent, and specificity from 91.9 to 94.9 percent. The researchers also report finding no clear signs of automation bias — that is, clinicians uncritically accepting AI answers (Medical Xpress). That last point is a claim from the research group, not an independent finding, but it points to a real problem in such systems.

This is not merely a promise about the future. In England, the NHS has deployed AI decision support such as Brainomix e-Stroke on brain scans, which has shortened the time from a patient's hospital arrival to the start of treatment by roughly an hour, according to the country's health department, as reported by USA Today (USA Today). Another example from the same review: a study published in JAMA Neurology in 2025 found that the AI system MELD Graph detected 64 percent of epilepsy lesions that radiologists had previously missed.

Psychiatry: Promising, but Early

A more cautious signal comes from psychiatry. According to an article in Knowable Magazine, republished by Scientific American, a Dutch research team trained AI on 88 speech features to distinguish people with schizophrenia from healthy controls — with 86.2 percent accuracy on audio recordings from patients the model had never heard before (Scientific American).

The interest stems from a real problem: symptom scoring in psychiatry is difficult to standardize, and clinicians' assessments of the same patient can diverge by 30 to 50 percent. An objective, language-based measure could in principle reduce that variation. But it is worth noting that this evidence comes from a single magazine article; the underlying publication details are not included in the available coverage. The result is interesting, not established.

Brain Implant: The Company's Own Statement

Paradromics announced this week that the first patient in its clinical trial — a 68-year-old woman from Michigan who has lost most of her ability to speak — can hold spontaneous conversations via a computer-synthesized voice, including on the phone with her grandchildren (New York Times). She received the wireless Connexus implant, with more than 400 electrodes, in surgery in June at University of Michigan Health.

All of the evidence here comes from the company's own announcement, relayed through the New York Times. There is no peer-reviewed publication or independent clinical verification in the available material. The company's chief clinical officer, Dr. William Marks, says the trial is planned to include ten patients nationally with conditions that cause speech loss and paralysis, such as stroke and ALS. The result may be significant — but for now it is a promise, not a documented finding.

Drug Design: Possible Signals, Many Caveats

The most heavily caveated case is rentosertib, developed end-to-end with AI, from target identification of the protein TNIK to generative molecular design. According to Insilico's announcement, relayed by Pulse2, a study in Nature Biotechnology shows the drug reduced calculated biological age as measured by six independent proteome-based "aging clocks" in a phase IIa analysis (Pulse2).

The figures vary by dose and measurement method, and there is no single headline number. In the group receiving 30 mg twice daily, a reduction of roughly three to four years was measured at week 4 on some measures, and up to six years on one clock. In the group receiving 60 mg once daily, the clocks designed to estimate chronological age showed reductions of 2.71 to 3.46 years, according to ScienceAlert's coverage (ScienceAlert). The signal plateaued after week four.

The caveats are substantial. The findings are explicitly exploratory and do not establish that rentosertib extends lifespan or reverses aging. All participants had idiopathic pulmonary fibrosis (IPF), and if the drug improved the lung disease, that improvement alone could have made the blood look biologically younger — the study did not fully separate the two explanations. Dr. Dung Trinh of MemorialCare Medical Group, who was not involved in the study, noted in comments to Medical News Today that the study was small, lasted only 12 weeks, and involved IPF patients, so the biomarker changes may reflect improvement in the disease rather than reversed aging (Medical News Today).

There is also an unresolved detail: Pulse2 states the analysis included 42 patients, while ScienceAlert writes that the original study recruited 71 people and that the new analysis examined samples from 42 participants. The discrepancy must be clarified against the primary paper before conclusions are drawn about the study population.

Rentosertib's further course — it is now said to be in phase III in China for IPF — and Insilico's reported revenues of around $106 million in the first half of 2026, up 287 percent from the same period a year earlier, come from the company's own statements and should be read in that light. The FDA told USA Today that AI use across all phases of drug development has increased significantly in recent years.

What the Evidence Gradient Means

The four cases span a spectrum: measured accuracy improvements in peer-reviewed studies (stroke CT), realized clinical deployment with reported time savings (NHS), early results based on sparse documentation (speech and schizophrenia), exploratory biomarker findings with known confounders (rentosertib), and pure company statements without peer review (Paradromics). For the reader, the practical rule is simple: ask about the study, its size, who is reporting the result — and whether anyone independent has confirmed it. The closer a claim stands to a company's own announcement, the more cautiously it should be read.