Design of an artificial intelligence model to screen spontaneous speech to detect Alzheimer's Disease

There is a shortage of physicians trained in the specialized care of Alzheimer's disease (AD). One possible solution is to use machine learning (ML)/artificial intelligence (AI) techniques to screen for the effects of AD on speech patterns, a technique that has the potential for automation. Our objective was to evaluate the ability of various language processing/artificial intelligence (AI) techniques to classify subjects into AD and controls. Our study used the Pitt dataset (n = 549), from DementiaBank, a shared dataset of audio files of spontaneous speech using the standardized Cookie Theft Picture Description task. The dataset was divided into training (n = 337), development (n = 115) and test (n = 97) datasets. A series of language processing techniques were tested on their ability to classify subjects into AD and controls, including classical (Tfidf, Term Frequency-Inverse Document Frequency), hybrid (Tfidf + text embeddings), neural classification and large language models. The hybrid model (Tfidf + text embeddings) was the best performing one, with accuracy and recall of 0.92 and 0.93 respectively on the test dataset. However, when tested on an independently collected validation dataset it only achieved 0.70 accuracy and 0.44 recall, performing well in identifying controls but missing some true AD cases. The model also demonstrated similar performance in identifying controls in a dataset that included a high proportion of diverse cognitive conditions, similar to referrals to a specialized memory clinic. AI techniques have demonstrated some utility for the classification of spontaneous speech audio files into those with AD and controls although these techniques could benefit from having enlarged datasets.