TD‐P‐014: COGID: A SPEECH RECOGNITION TOOL FOR EARLY DETECTION OF ALZHEIMER'S DISEASE
Brianna M Broderick, Si Long Jenny Tou, Emily Mower Provost · Alzheimer s & Dementia · 2018
Due to the variety of symptoms and progression, dementia and Alzheimer's disease (AD) can often go undiagnosed. The diagnostic process is complicated, expensive, requires repeated testing in an unfamiliar environment and can cause added stress and confusion. Language can often be early signs of the disease. A tool that analyzes acoustic and semantic properties of speech and recognizes characteristics of AD could be a diagnostic aid to make the process more accessible. We combined findings from literature in attempt to create a robust AD detection tool. The tool was built using audio from the Dementia Bank's Pitt Corpus which includes audio and transcriptions from 142 subjects with diagnosed AD and 96 healthy elderly controls. A portion of subjects have repeat visits over time, totaling 223 control interviews and 234 from individuals with AD. We used the Cookie Task which allows us to ensure a controlled context which should elicit the same type of vocabulary and speech. We extracted features based on the linguistic abilities affected by AD. Transcription based features captured a decline of vocabulary and semantic processing and included features of lexical richness, utterance length, frequency of filler words, pronouns, verbs, adjectives and proper nouns. Acoustic features focused on word finding errors, fluidity and rhythm of speech and included pause frequency and duration, speech rate, and articulation rate. Principal Component Analysis (PCA) was used for feature reduction. Classification was done using SVM and KNN, and we used leave-one-out cross validation to evaluate the models. The best model used data from all visits for each subject, PCA components = 3, SVM with RBF kernel and achieves an F1 of 0.73. This model has a recall of 0.83.