Fully-automated analysis of voice recordings–from neuropsychological tests–found to help differentiate normal cognition from dementia and mild cognitive impairment

Cred­it: James Byrne

Voice Record­ings Spot Cog­ni­tive Impair­ment (Med­Page Today):

A machine-learn­ing mod­el iden­ti­fied mild cog­ni­tive impair­ment and demen­tia from dig­i­tal voice record­ings of neu­ropsy­cho­log­i­cal tests, an ear­ly study showed.

Among 1,084 peo­ple in the Fram­ing­ham Heart Study whose tests were record­ed, the aver­age area under the curve (AUC) reached 92.6% for dif­fer­en­ti­at­ing nor­mal cog­ni­tion from demen­tia, 88.0% for dis­cern­ing nor­mal cog­ni­tion or mild cog­ni­tive impair­ment from demen­tia, and 74.4% for dis­tin­guish­ing nor­mal cog­ni­tion from mild cog­ni­tive impairment.

The mod­el used voice recog­ni­tion to tran­scribe record­ings to text and lever­aged nat­ur­al lan­guage pro­cess­ing meth­ods for analy­sis, report­ed Ioan­nis Pascha­lidis, PhD, of Boston Uni­ver­si­ty, and co-authors in Alzheimer’s & Demen­tia … “It sur­prised us that speech flow or oth­er audio fea­tures are not that crit­i­cal; you can auto­mat­i­cal­ly tran­scribe inter­views rea­son­ably well and rely on text analy­sis through AI to assess cog­ni­tive impair­ment,” Pascha­lidis said.

The Study:

Auto­mat­ed detec­tion of mild cog­ni­tive impair­ment and demen­tia from voice record­ings: A nat­ur­al lan­guage pro­cess­ing approach (Alzheimer’s & Demen­tia). From the Abstract:

• Intro­duc­tion: Auto­mat­ed com­pu­ta­tion­al assess­ment of neu­ropsy­cho­log­i­cal tests would enable wide­spread, cost-effec­tive screen­ing for dementia.
• Meth­ods: A nov­el nat­ur­al lan­guage pro­cess­ing approach is devel­oped and val­i­dat­ed to iden­ti­fy dif­fer­ent stages of demen­tia based on auto­mat­ed tran­scrip­tion of dig­i­tal voice record­ings of sub­jects’ neu­ropsy­cho­log­i­cal tests con­duct­ed by the Fram­ing­ham Heart Study (n = 1084). Tran­scribed sen­tences from the test were encod­ed into quan­ti­ta­tive data and sev­er­al mod­els were trained and test­ed using these data and the par­tic­i­pants’ demo­graph­ic characteristics.
• Results: Aver­age area under the curve (AUC) on the held-out test data reached 92.6%, 88.0%, and 74.4% for dif­fer­en­ti­at­ing Nor­mal cog­ni­tion from Demen­tia, Nor­mal or Mild Cog­ni­tive Impair­ment (MCI) from Demen­tia, and Nor­mal from MCI, respectively.
• Dis­cus­sion: The pro­posed approach offers a ful­ly auto­mat­ed iden­ti­fi­ca­tion of MCI and demen­tia based on a record­ed neu­ropsy­cho­log­i­cal test, pro­vid­ing an oppor­tu­ni­ty to devel­op a remote screen­ing tool that could be adapt­ed eas­i­ly to any language

The Study in Context:

• Start­up Aur­al Ana­lyt­ics rais­es $4.3M to mon­i­tor neu­ro­log­i­cal health by ana­lyz­ing speech patterns
• Mind­strong Health iden­ti­fies dig­i­tal bio­mark­ers of cog­ni­tive func­tion using smart­phone data
• Neu­rotech­nol­o­gy can improve our lives…if we first address these Pri­va­cy and Informed Con­sent issues