GPT-3 can spot early Alzheimer's with 80% accuracy

While millions of people are using ChatGPT to write emails and debug code, researchers are discovering that the models behind it might have a far more consequential calling: catching dementia before doctors do.

Alzheimer’s disease is notoriously stealthy in its earliest stages. By the time patients undergo standard neurological evaluations and memory screenings, significant brain damage has often already occurred.

Now, a study from Drexel University’s School of Biomedical Engineering, Science and Health Systems suggests that large language models could offer a fast, low-cost shortcut. Published in PLOS Digital Health, the research reveals that OpenAI’s GPT-3 can predict early-stage Alzheimer’s from spontaneous speech transcripts with 80% accuracy.

Listening between the lines

Language impairment affects between 60% and 80% of dementia patients. Long before severe memory lapses appear, people often begin hesitating, confusing word meanings, or making subtle grammatical errors.

Historically, digital detection tools focused on acoustics, measuring things like speech pauses, voice strength, and articulation. But Drexel researchers Felix Agbavor and Dr. Hualou Liang decided to look at semantic structure instead.

Using voice transcripts from an NIH-supported speech dataset, the team tapped GPT-3 to analyze word choice and sentence construction. The model created dense text embeddings that captured the linguistic fingerprint of cognitive decline, essentially turning the program into a dementia screening engine.

When pitted against two top conventional natural language processing systems, GPT-3 outperformed both in identifying Alzheimer's cases with fewer false negatives. In a secondary evaluation predicting patients' scores on the Mini-Mental State Exam, GPT-3’s text analysis proved roughly 20% more accurate than traditional acoustic assessments.

The team is now working on converting the concept into a simple web application for home and community clinic use.

Why it matters

Early detection doesn’t cure Alzheimer’s, but it gives patients more treatment options and time to plan before severe cognitive decline sets in. If a patient can simply speak into a browser for five minutes instead of sitting through hours of stressful clinical evaluations, early testing becomes accessible to a far wider population.

For the AI sector, this demonstrates the unexpected utility of frontier foundation models in medicine. Instead of hand-crafting narrow diagnostic algorithms from scratch, researchers were able to take an off-the-shelf general language model and extract clinical signals that specialized acoustic systems missed.

That said, an 80% accuracy rate is suitable for pre-screening, not a definitive diagnosis. The real test will be whether Drexel can maintain this level of sensitivity across noisy, real-world conversations and diverse demographic groups without generating an overwhelming number of false alarms.