番茄社区

An artificial intelligence (AI) model using a short language test was modestly accurate in spotting whether healthy people would have a future diagnosis of Alzheimer's disease, researchers reported.

The model predicted whether cognitively normal older adults would develop Alzheimer's symptoms before age 85 with an area under the receiver operating curve (AUC) of 0.74 and an accuracy of 70%, wrote Elif Eyigoz, PhD, of IBM Research in Yorktown Heights, New York, and colleagues in .

Currently, neuropsychological tests and biomarkers like cerebrospinal fluid and imaging are used to monitor Alzheimer's progression. Potential blood-based biomarkers are being studied to predict Alzheimer's disease in cognitively normal older adults, but none are available for clinical use.

Language competence is a sensitive indicator of mental dysfunction, Eyigoz said. "Aging-related cognitive decline manifests itself in almost all aspects of language comprehension and production, because even seemingly mundane linguistic abilities -- such as object naming -- engage extensive brain networks," she told 番茄社区.

The IBM model used short written samples from participants in the (FHS) who had regular cognitive assessments using the and completed a neuropsychological test battery in successive visits.

The model derived linguistic variables from written responses to the Cookie Theft Task, part of the Boston Aphasia Diagnostic Examination included in the FHS test battery. In the task, participants were asked to describe a drawing of three people: a boy on a toppling stool reaching into a cookie jar, a girl standing next to him, and a woman wiping dishes at an overflowing sink.

The study included 703 samples from 270 FHS participants. A dataset consisting of a single sample from 80 participants was held for testing; half of these participants developed Alzheimer's symptoms on or before age 85 (cases); the other half did not (controls). Alzheimer's diagnoses were based on criteria.

Cases and controls were matched on approximate age, sex, and education. The test data set included only one sample per participant, and only samples collected before cognitive impairment onset were used.

Among the test cases, mean time to diagnosis of mild Alzheimer's from cognitive normalcy was 7.59 years. Alzheimer's diagnoses were harder to predict among people with a college degree (AUC 0.70) than those without (AUC 0.76). They also were predicted more confidently among women (AUC 0.83) than men (AUC 0.64).

Future Alzheimer's onset was associated with telegraphic speech, repetitiveness, and misspellings. Telegraphic speech is common in non-fluent aphasia, the researchers noted: it's simple in grammatical structure and marked by a lack of determiners (like "the" or "a"), auxiliary verbs (like "is" or "are"), and entire subjects.

Using generic terms like "boy," "girl," or "woman" instead of more specific words like "son," "brother," "sister," "daughter," or "mother" to describe the people in the drawing was tied to a higher risk of Alzheimer's disease. Mentioning details like the dishcloth or dishes was linked with lower risk of Alzheimer's.

Linguistic variables from a single administration of the Cookie Theft Task performed better than predictive models that incorporated APOE genotype, demographic variables, and other neuropsychological test results, the researchers noted.

A limitation of the model is that it relied on the written version of the Cookie Theft Task, they added: the spoken version may show different aspects of linguistic dysfunction.

"We designed our study with primarily clinical trial enrichment in mind," Eyigoz said. "Beyond that, we think that our method, mutatis mutandis, can help augment the information and data that clinicians have access to in their practice."

The model may offer a low-burden approach to monitoring treatment and disease progression if patients consent, Eyigoz added. "For example, a practical application of the methods presented in this study could be utilization of its results in a larger system, which would additionally include other automated assessments of cognitive decline -- such as analysis of motor impairments, olfactory impairments, acoustic variations, etc. -- as sophisticated decision-making in artificial intelligence usually involves combining results obtained from multiple methods," she said.

Comment