AI Helps Predict Dementia Using Speech Patterns

Voice recordings helped predict which patients with mild cognitive impairment developed Alzheimer’s dementia in 6 years.

Combined with basic demographic information, speech patterns recorded
in neuropsychological exams achieved an accuracy of 78.5% and a
sensitivity of 81.1% in predicting progression from mild cognitive
impairment to dementia in a 6-year window, reported Ioannis Paschalidis,
PhD, of Boston University, and colleagues in Alzheimer’s & Dementia.

“However, the specificity of predicting whether an individual with
mild cognitive impairment will progress to Alzheimer’s disease within 6
years was moderate, at 75%,” Paschalidis and co-authors wrote. “To reduce the costs associated with recruiting subjects for clinical trials, it is important to improve the specificity.”

The
study leveraged AI methods for speech recognition and processed the
resulting text using language models. The researchers used the content
of the interview — words spoken and how they were structured — not
acoustic features like enunciation or talking speed.

The approach could be developed into a remote screening tool for
predicting progression to Alzheimer’s dementia, the researchers noted.
“If you can predict what will happen, you have more of an opportunity
and time window to intervene with drugs, and at least try to maintain
the stability of the condition and prevent the transition to more severe
forms of dementia,” Paschalidis said in a statement.

In previous work, Paschalidis and colleagues reported that a model
using natural language processing (NLP) discerned normal cognition from
mild cognitive impairment and dementia based on voice recordings.
Other researchers have found that speech patterns in phone
conversations could spot people with early-to-moderate Alzheimer’s
dementia.

The current study evaluated neuropsychological test interviews of 166 Framingham Heart Study
participants, including 90 people who had progressed from mild
cognitive impairment to dementia within 6 years, and 76 people who had
stable mild cognitive impairment in that period. The median age was 81,
and nearly two-thirds of participants were women.

Neuropsychological test interviews were digitally recorded in the
Framingham Heart Study. These hour-long interviews include cognitive
tests like the Boston Naming Test, the Hooper Visual Organization Test,
and the Wechsler Memory Scale.

“The neuropsychological test, triggered by patient history and in
conjunction with a clinical examination, provides a comprehensive
evaluation of cognitive function, including attention, memory, language,
and visuospatial abilities,” Paschalidis and co-authors observed.

“Researchers have explored computer-based approaches to predict the
progression from mild cognitive impairment to dementia using
neuropsychological tests, primarily relying on hand-crafted features and
cognitive scores extracted from the neuropsychological test by
clinicians,” they pointed out. “However, these approaches have not yet
achieved full automation, limiting their potential for more precise and
efficient cognitive evaluations.”

Paschalidis
and colleagues used recorded neuropsychological test interviews to
predict the likelihood of participants transitioning to Alzheimer’s,
training a model to spot connections among speech, demographics,
diagnosis, and disease progression. The analysis used text automatically
transcribed from the recordings.

The model’s accuracy and sensitivity outperformed other measures at
predicting progression to dementia in 6 years. Standard
neuropsychological tests had an accuracy of 74.7% and sensitivity of
77.2%, for example. The Mini-Mental State Examination (MMSE) had an
accuracy of predicting progression to dementia over 6 years of 62.9% and
a sensitivity of 66.7%.

The study demonstrates the potential of automatic speech recognition
and NLP techniques to develop a prediction tool to identify which
patients with mild cognitive impairment are at risk of dementia, the
researchers said.

“Our method achieved high accuracy and outperformed other
non-invasive approaches,” Paschalidis and co-authors wrote. “However,
further prospective studies with larger populations are necessary to
validate the generalizability of our models.”

The definition of mild cognitive impairment needs to be standardized
to better compare results, they noted. “With continued development and
refinement, our approach may contribute to early intervention and
selection in clinical trials for novel Alzheimer’s disease treatments,
ultimately improving patient outcomes,” they wrote.