While a primary care clinician meets with a patient for their annual check-up, an AI model analyzes the patients’ speech patterns for signs of cognitive impairment. Such a screening could be possible within the next few years, according to a recent feasibility study published in JAMA Neurology.

Researchers tested how well machine learning algorithms analyze conversational speech patterns to identify red flags for potential dementia or Alzheimer’s cases — one model accurately identified about 30% of patients who had cognitive impairment.

“While the clinical note is being generated, all of the information that we use in our algorithm is also captured during that process, just listening to the conversation,” said Joseph Colonel, PhD, a postdoctoral research fellow at the Icahn School of Medicine at Mount Sinai in New York City, who led the study. “All the pieces are there so that at the end of the doctor’s visit, there could be a risk score that goes up [in the patient’s chart] and says, you might want to consider that this person should be screened for cognitive impairment.”

Cognitive impairment, including Alzheimer’s disease, affects nearly 20% of adults aged 65 years or older, yet cases often go undiagnosed. But primary care clinicians are uniquely poised to screen for these issues, potentially with AI tools, which Colonel and colleagues sought to test.

For the feasibility study, Colonel and his team recruited nearly 1000 patients aged 55 years or older (mean age, 67.2 years; 55% female, 35% Black) who had no history of cognitive impairment between 2020 and 2021 from five primary care clinics. New York-based patients (n = 787) were recruited to train the AI model, while Chicago-based patients (n = 179) were used to validate the model’s effectiveness.

Patients and clinicians wore a lavalier microphone during primary care visits, amounting to nearly 400 hours of recorded audio used to train the machine learning model by detecting rate of speech, number of pauses, tone, and how quickly the patient responded.

After the visit, patients were screened with a cognitive assessment for dementia, which revealed 21% of participants in each group had undiagnosed cognitive impairment.

The AI model positively predicted cognitive impairment in 30.4% (95% CI, 28.7%-32.1%) of these cases when noise-reduced audio was used, achieving sensitivity of 68.2% (95% CI, 61.8%-74.6%) and specificity of 63.6% (95% CI, 59.8%-67.4%).

“If matured, this approach could meaningfully change how and when cognitive decline is detected,” Gabriela Meade, PhD, an assistant professor of speech pathology, and Hugo Botha, MB ChB, a neurologist, both with the Mayo Clinic in Rochester, Minnesota, wrote in an accompanying editorial.

Ellie Fishbein, MD, director of Geriatric Primary Care at WashU Medicine in St. Louis, said the proof-of-concept results are promising.

“We already use ambient technology as scribes in a lot of primary care clinics, so I’m interested in optimizing a tool we already use so we can pull as much data from it as possible,” she said.

Colonel said a path to clinical implementation remains undefined due to rapid, unpredictable technological shifts. Deploying such a tool prematurely would come with risks, as false positive results can trigger invasive follow-up procedures and severe patient anxiety.

“It only just opens the door to a workup that can take considerable time and be emotionally charged for patients and their loved ones,” Fishbein said. “I’m excited to see where the research goes next, so we can shape this into something useful for people, something that ultimately adds to their quality of life.”

The study was funded by the National Institute on Aging. Colonel and Fishbein reported having no relevant financial relationships.

Kelsey Mesmer, PhD, is a freelance journalist and journalism professor at Saint Louis University in St. Louis.