Modeling Human Behavior in Strategic Settings
My dissertation research was motivated by this question: Can we design a computer program to detect whether someone has dementia simply by listening to them talk? It is estimated that 35.6 million people worldwide were living with dementia in 2010, and over half will never be diagnosed. The Alzheimer Society of Canada reports that 747,000 Canadians suffered from cognitive impairment or dementia in 2011, and suggest that these numbers will double in the next 20 years.
Many researchers and clinicians agree that the analysis of narrative or conversational speech is important for assessing the extent of an individual's linguistic and cognitive impairment. However, there is simply not enough time to perform an in-depth analysis of speech in a clinical setting. Therefore, in my research I developed a software application to automatically analyze speech for signs of dementia.
The program works by going through the following steps: First, the audio file is analyzed for relevant features such as number and length of pauses, changes in pitch and loudness, etc. Then a transcript is generated using automatic speech recognition. Next, natural language processing algorithms are used to extract information about syntactic complexity, vocabulary richness, word specificity, and the amount of content contained in the narrative. This results in a high-dimensional vector that precisely characterizes an individual's speech and language function. We can then train a machine learning classifier to distinguish between the patterns associated with dementia versus those patterns associated with healthy aging.
Ultimately, I was able to classify certain language-based types of dementia with up to 100% accuracy, and detect Alzheimer's disease with 82% accuracy. Beyond classification accuracy, and in contrast to
earlier work in the field, my dissertation focused on questions of interpretability and practical utility. In particular, rather than assuming perfect output at each stage in the analysis pipeline, I considered the types of errors introduced by automatic speech recognition and sentence segmentation, and their effect on the machine learning classifiers. Potential future applications include screening for dementia, monitoring the effects of therapy or pharmaceutical interventions, and predicting trajectories of decline in people experiencing mild cognitive impairment.