AI in Medicine: From Stroke Scans to Aging Clocks

AI in Medicine: From Stroke Scans to Aging Clocks

Within a single week in September 2026, results arrived showing what artificial intelligence is already doing in health care: improved stroke diagnosis in a multinational study, speech analysis that distinguishes schizophrenia from healthy controls, a wireless brain implant conversation, and an AI-designed drug that in an early analysis shifted biological aging markers — though that last signal is exploratory and cannot be separated from improved lung disease.

From Overview to Concrete Results

USA Today's September 14 overview describes how AI is already transforming diagnosis, drug design, and surgery — from research to treatment (6d9a4956). The same week brought a cluster of dated results that put substance behind that frame. They sit, however, on very different levels of evidence, and the gap between them is the real story.

Stroke Diagnosis With a Measured Effect

A multinational validation, published in the Journal of NeuroInterventional Surgery and reported by Medical Xpress, found that AI support improves clinicians' ability to detect large vessel occlusions (LVO) on non-contrast CT. Without AI, diagnostic accuracy was 0.718 AUC; with AI it rose to 0.852 AUC in a reader study of 963 patients (f88fa4c7). The research team from Korea University College of Medicine also found no clear automation bias — that is, no tendency for clinicians to accept AI results uncritically (f88fa4c7).

This is the most mature part of the week's cluster: a peer-reviewed effect in routine diagnostics, with measurement of both the gain and a known risk of AI support.

Speech as a Diagnostic Tool

A Dutch research team trained AI on speech features and tested the model on audio recordings from patients it had never heard before. It distinguished schizophrenia patients from healthy controls with 86.2 percent accuracy, and could also differentiate between subtypes of the disorder, according to Scientific American/Knowable Magazine (8c93660c). This points toward speech as a possible, objective supplement in early psychiatric diagnosis — with the caveat that it remains a research result, not a clinical tool.

Brain Implant and Real-Time Conversation

Paradromics announced on September 14 that a 68-year-old patient with a paralyzing neurological condition has held conversations in real time using the company's wireless Connexus system — including via imagined speech, that is, simply by thinking about what she wanted to say (8c94bea1). The patient, the first in the study, received the system with more than 400 electrodes in June at University of Michigan Health; the study is intended to include 10 patients (8c94bea1).

Note: this is the company's own announcement, conveyed through The New York Times. It is not peer-reviewed, and it concerns one patient so far.

The Drug That Moved Aging Clocks

Rentosertib, whose molecular structure was generated with AI assistance, is Insilico Medicine's candidate against idiopathic pulmonary fibrosis. A new analysis published in Nature Biotechnology on September 7 shows, according to the company, that the drug reduced biological aging markers measured by six different proteomic "aging clocks" — AI technology designed to predict illness and death (e5177584).

In the analysis, blood samples from 42 participants were collected over 12 weeks, and more than 2,800 proteins were measured against the six clocks (f563b2b8). The strongest signal came in week 4 among participants who received 30 mg twice daily, with an estimated reduction in predicted biological age of around three to four years, and up to six years on one clock (612bc717).

Why the Signal Cannot Be Taken as Proof That "Aging Is Reversed"

There is an important catch: all participants had serious lung disease. If rentosertib reduced fibrosis or otherwise improved the lung disease, that improvement alone could have made the blood look biologically younger, ScienceAlert notes in its coverage (a6ca5f11). In other words: the aging signal cannot, for now, be separated from improvement in the disease itself.

In addition, the drug candidate is, according to The New York Times, still years away from regulatory approval, even for sick patients, and Insilico has not tested the anti-aging effect in healthy people (e5177584). The company states that rentosertib has reached phase III development in China against pulmonary fibrosis (612bc717). The aging results must be attributed to Insilico Medicine's own clock measurements and lack independent replication.

What Is Verified — and What Is Not

The week's results sit at different levels of evidence:

• Peer-reviewed study: AI support improved LVO detection from 0.718 to 0.852 AUC in 963 patients, with no clear automation bias (f88fa4c7).
• Research result: speech analysis with 86.2 percent accuracy for schizophrenia (8c93660c).
• Company's own analyses: rentosertib's aging-clock signals, published in Nature Biotechnology, but exploratory and confounded by disease improvement (e5177584; a6ca5f11).
• Company announcement without peer review: Paradromics' real-time BCI conversation, one of ten planned patients (8c94bea1).

What is needed for confirmation is the same at both ends: independent replication. For stroke AI, that means broader implementation with continued measurement of both the gain and the bias risk. For rentosertib, it means studies that can separate an effect on aging from an effect on the lung disease — ideally including healthy participants. Until then, the right view of the week's news is curious, but calibrated.