This study by PLOS Digital Health utilizes acoustic features, GPT-3 embeddings, and a combination of the two approaches in order to classify Alzheimer's Disease (AD) vs. Non-AD. Furthermore, this study uses the same methods to predict the Mini-Mental State Exam (MMSE) scores of a set of participants. The results show that GPT-3, when combined with acoustic features, yielded the highest accuracy for the AD vs. Non-AD classification, as well as for MMSE score prediction. This study demonstrates the potential of large language models for predicting dementia from spontaneous speech, and provides a method for accurate and reliable early diagnosis of dementia.
