Automatic remote screening of speech features associated with Alzheimer's disease

点击翻译按钮获取中文摘要

Assessing Alzheimer's disease (AD) is an expensive and laborious process that isunsustainable given our rapidly aging population, especially if one accounts for indirect costssuch as lost time in travel, wait times, and hours spent in assessment, which is a process solabourious that it is often repeated only every few years. Moreover, the high variability ofsymptoms in AD cannot be ascertained accurately from a single assessment, whichminimizes the effectiveness of current practice. Rapid, repeatable, and remote assessment isessential.Many individuals with AD have diminished language abilities. Linguistic measures, like theratio of pronouns to nouns, relate strongly to cognitive decline. By measuring such featuresand using advanced machine learning software, we have developed computational methodsthat are over 85% accurate in identifying Alzheimer's disease from short speech samples. Aswe continue to improve this technology, we will make it available over the public phonenetwork accessible by a 1-800 number. This system will automatically receive calls from (andmake calls to) outpatients, and completely automate speech-based tests including questionanswering,narratives, and random item generation. This will be encapsulated in an interfacethat healthcare workers can use to assign and monitor assessments and exercises.Currently, to extract language measures, speech must be transcribed by professionals, whichis impossible on large scales. Unfortunately, no automatic speech recognition (ASR) systemexists for people with AD. We will therefore provide the first complete ASR system for peoplewith AD.Our objectives are: 1) to improve the accuracy of ASR for people with AD; 2) to refine andvalidate our software in remote monitoring/assessment; 3) to embed our software into a userfriendlyphone-based interface.