Study: Artificial intelligence program identifies linguistic markers that predict, with 70% accuracy, who gets Alzheimer’s Disease years later

by

Fig. 3. Cook­ie-theft pic­ture descrip­tion task (CTT) exam­ples from the Fram­ing­ham Heart Study, includ­ing an unim­paired sam­ple (a), an impaired sam­ple (b), and an even more impaired sam­ple show­ing sig­nif­i­cant mis­spellings and min­i­mal gram­mat­ic com­plex­i­ty ©. Cred­it: Eyigoz et al (2020)

Alzheimer’s Pre­dic­tion May Be Found in Writ­ing Tests (The New York Times):

… the researchers looked at a group of 80 men and women in their 80s — half had Alzheimer’s and the oth­ers did not. But, sev­en and a half years ear­li­er, all had been cog­ni­tive­ly normal.

The men and women were par­tic­i­pants in the Fram­ing­ham Heart Study, a long-run­ning fed­er­al research effort that requires reg­u­lar phys­i­cal and cog­ni­tive tests. As part of it, they took a writ­ing test before any of them had devel­oped Alzheimer’s that asks sub­jects to describe a draw­ing of a boy stand­ing on an unsteady stool and reach­ing for a cook­ie jar on a high shelf while a woman, her back to him, is obliv­i­ous to an over­flow­ing sink.

The researchers exam­ined the sub­jects’ word usage with an arti­fi­cial intel­li­gence pro­gram that looked for sub­tle dif­fer­ences in lan­guage. It iden­ti­fied one group of sub­jects who were more repet­i­tive in their word usage at that ear­li­er time when all of them were cog­ni­tive­ly nor­mal. These sub­jects also made errors, such as spelling words wrong­ly or inap­pro­pri­ate­ly cap­i­tal­iz­ing them, and they used tele­graph­ic lan­guage, mean­ing lan­guage that has a sim­ple gram­mat­i­cal struc­ture and is miss­ing sub­jects and words like “the,” “is” and “are.”

The mem­bers of that group turned out to be the peo­ple who devel­oped Alzheimer’s disease.

The A.I. pro­gram pre­dict­ed, with 75 per­cent accu­ra­cy, who would get Alzheimer’s dis­ease, accord­ing to results pub­lished recent­ly in The Lancet jour­nal EClinicalMedicine.

The Study:

Lin­guis­tic mark­ers pre­dict onset of Alzheimer’s dis­ease (EClin­i­calMed­i­cine). From the Abstract:

  • Back­ground: The aim of this study is to use clas­si­fi­ca­tion meth­ods to pre­dict future onset of Alzheimer’s dis­ease in cog­ni­tive­ly nor­mal sub­jects through auto­mat­ed lin­guis­tic analysis.
  • Meth­ods: To study lin­guis­tic per­for­mance as an ear­ly bio­mark­er of AD, we per­formed pre­dic­tive mod­el­ing of future diag­no­sis of AD from a cog­ni­tive­ly nor­mal base­line of Fram­ing­ham Heart Study par­tic­i­pants. The lin­guis­tic vari­ables were derived from writ­ten respons­es to the cook­ie-theft pic­ture-descrip­tion task. We com­pared the pre­dic­tive per­for­mance of lin­guis­tic vari­ables with clin­i­cal and neu­ropsy­cho­log­i­cal vari­ables. The study includ­ed 703 sam­ples from 270 par­tic­i­pants out of which a dataset con­sist­ing of a sin­gle sam­ple from 80 par­tic­i­pants was held out for test­ing. Half of the par­tic­i­pants in the test set devel­oped AD symp­toms before 85 years old, while the oth­er half did not. All sam­ples in the test set were col­lect­ed dur­ing the cog­ni­tive­ly nor­mal peri­od (before MCI). The mean time to diag­no­sis of mild AD was 7.59 years.
  • Find­ings: Sig­nif­i­cant pre­dic­tive pow­er was obtained, with AUC of 0.74 and accu­ra­cy of 0.70 when using lin­guis­tic vari­ables. The lin­guis­tic vari­ables most rel­e­vant for pre­dict­ing onset of AD have been iden­ti­fied in the lit­er­a­ture as asso­ci­at­ed with cog­ni­tive decline in dementia.
  • Inter­pre­ta­tion: The results sug­gest that lan­guage per­for­mance in nat­u­ral­is­tic probes expose sub­tle ear­ly signs of pro­gres­sion to AD in advance of clin­i­cal diag­no­sis of impairment.

Study in Context: