Automated Analysis fo Spoken Story Recall Tests

Abstract

Clinical research into Alzheimer's disease (AD) and the mild cognitive impairment (MCI) that precedes its full onset is increasingly focused on early diagnosis. Currently, the reliable clinical diagnosis of MCI requires expensive and time-consuming evaluations by skilled clinicians utilizing neuropsychological tests administered in person by trained psychometricians. Automation of these evaluations would allow frequent, large-scale testing for MCI. Numerous neurolinguistic and neuropsychological studies have demonstrated that there are objective, measurable differences in the spoken and written language produced by patients with MCI and AD, and healthy age-matched adults. Our long-term goal is to provide technologies that enable minimally-obtrusive collection of speech data in a subject’s home, and analysis of these data to signal potential impairment. Our main hypothesis for the proposed work is that automatic speech processing techniques will yield statistically significant discrimination between healthy and MCI subjects on both a word-list recall and a story-recall task. We will address the technological feasibility within a constrained and controlled setting, namely automatic analysis of spoken neuropsychological test batteries administered in subjects’ homes. This study has two specific aims. First, to evaluate whether high-accuracy word transcriptions of story retellings can be produced by automatic speech recognition (ASR), given the task requirements (e.g. home environment, moderately large vocabulary) and task constraints (e.g. single speaker, known topic). Second, to determine whether speech and language features extracted from the story recall test can be used to effectively discriminate between healthy and MCI subjects. A number of techniques will be applied, including (i) training an ASR system on speech from elderly speakers, (ii) adaptation of the system to individual speakers, (iii) language modeling based on statistics from appropriate corpora, (iv) automatic parsing to determine sentence structure, (v) automatic analysis of sentence structure and cross entropy. Results from this project will allow a preliminary evaluation of the feasibility of effective continuous speech ASR in home environments, and the utility of features derived from the spoken language analysis for discrimination between healthy and MCI subjects.