JNNP: A fully automated cognitive screening tool based on speech and language assessment

Memory clinics assess patients with a variety of cognitive impairments, including those associated with Alzheimer's disease (AD), mild cognitive impairment (MCI), and functional cognitive impairment (FCD).
early referral to appropriate care pathways is conducive to the health of patients and the effective allocation of medical resources.
between 2009 and 2014, memory outpatient referrals increased by 682 per cent, indicating the need for accurate preclinical screening tools to guide patients in choosing the most appropriate service.
AD is associated with subtle language impairments that can be decades before the loss of situational memory.
previous studies have shown that qualitative analysis of session characteristics, inspired by session analysis methods (CA), can distinguish between patients with FCD and patients with neurodegenerative diseases (ND).
, however, relies on trained experts and is not easy to expand.
that while the use of automated speech analysis to recognize cognitive impairment has been studied, most studies do not describe fully automated solutions.
they rely on data collected from human interactions for automated analysis or using manually generated transcriptions.
this article created a fully automated layered tool.
"CognoSpeak" consists of a virtual clinician, a talking head displayed on a computer screen that asks questions and records the patient's verbal response.
the system uses automatic speech recognition (ASR) and journals (which divide recordings into contributions from different speaker) to extract acoustic and linguistic metrics, and machine learning classifiers to select the most likely diagnostic categories.
60 participants; each of the four different diagnostic groups had 15: AD, MCI, FCD, and Health Control Group (HC).
patient participants came from a specialist memory clinic in Sheffield.
patients can record themselves or in the absence of an escort.
the nervous system is based on standard diagnostic criteria and has been reviewed by a multidisciplinary team.
significant emotional disorders (identified by the Clinical History and Patient Health Questionnaire (PHQ-9>15)) and significant cerebrovascular diseases led to removal.
subjects were evaluated using Addenbrooke's Cognitive Examination Revision (ACE-R) tool or a detailed neuropsychological assessment.
brain imaging (including CT, MRI, and TC99m-HMPAO single photon emission CT) according to clinical needs.
in the health control group did an MRI (VPH-DARE@IT) in a previous study, all participants were native English speakers, mostly white British.
, although a small number of participants are of South Asian descent, they are native or bilingual.
use the ASR system to transply audio into a string of words.
automatically tweed text and recorded audio are used to extract a series of characteristics that machine learning-based classifiers (logistic regressions) use to assign participants to diagnostic classes.
72 features were extracted from the voices of patients and entourage, including 17 CA-inspired features, 24 pure acoustic features, 24 lexical features, and 7 word vector features.
11 minutes and 24 seconds for all participant groups and people with cognitive impairment.
HCs took an average of 9 minutes and 58 seconds; FCD patients 11 minutes 2 seconds; MCI patients 9 minutes 34 seconds; and AD patients 15 minutes 2 seconds.
difference between the MCI group and the AD group is significant (U-59, p-0.026).
In the bidirectional distinction between patients who may have neurodegenerative and non-aggressive diseases, the Cognospiak system was 87 percent accurate in identifying MCI or AD, while the participants were correctly assigned to HC or FD with 77 percent accuracy.
this equates to neurodegenerative memory impairment sensitivity of 86.7 percent and specificity of 76.7 percent.
AD, Alzheimer's disease; FCD, functional cognitive impairment; HC, health control.
65 per cent of cases are classified correctly.
80% accuracy in identifying AD patients.
two participants were incorrectly assigned to MCI and the other to FCD and HC groups.
MCI has an accuracy of 80%.
one participant was incorrectly assigned to AD and two participants were incorrectly assigned to FCD and HC.
50% accuracy of the identification of patients with HC or FD.
8 participants were incorrectly assigned to MCI and 7 were incorrectly assigned to AD.
this equates to 80.0 per cent sensitivity and 77.8 per cent specificity.
in four groups, 60 percent of the subjects had higher accuracy of identification throughout the task.
accuracy of identifying AD participants was 80%, the accuracy of identifying MCI participants was 60%, the accuracy of identifying FCD participants was 47%, and the accuracy of identifying HC participants was 53%.
the sensitivity of the bidirectional classification system was 86.7% compared to the most commonly used screening tools.
was slightly less specific, at 76.7 per cent, but this was acceptable given the focus of early identification and the fact that "false positives" would be investigated by professional memory services.
using a three-way classification method, CognoSpeak is able to identify MCI patients with sensitivity of 80.0% and specificity of 77.8%, exceeding the sensitivity (66.34%) and specificity (72.94%) of MMSE (mini-mentalstate examination, MMSE) in distinguishing MCI from HCs.
accurate identification of MCI patients is helpful for early intervention and for participation in research.
't know of any other screening tools that include FCD patients in their studies.
in the validation study of screening layer tools, inclusion in this patient group is essential, as FCD patients account for 24% of specialist memory service referrals.
tool in this article can distinguish between FCD and MCI/AD, not just HC and MCI/AD groups, which increases its effectiveness.
future research will also include groups of patients with non-AD dementia and Alzheimer's disease.
O'Malley RPD, Mirheidari B, Harkness K, et al Fully automated cognitive tool screening on assessment of speech and language Journal of Neurology, Neurosurgery and Psython Published Online First: 20 November 2020. doi: 10.1136/jnnp-2019-322517MedSci Original Source: MedSci Original Copyright Notice: All text, images and audio and video materials on this website that indicate "Source: Met Medical" or "Source: MedSci Original" are owned by Mets Medicine and are not authorized to be reproduced by any media, website or individual Source: Mays Medicine.
all reprinted articles on this website are for the purpose of transmitting more information and clearly indicate the source and author, and media or individuals who do not wish to be reproduced may contact us and we will delete them immediately.
at the same time reproduced content does not represent the position of this site.
leave a message here