Abstract: Mild Cognitive Impairment (MCI) is a condition where an individual’s ability to perform regular activities starts deteriorating. It is often an early indicator of Dementia. By identifying individuals with MCI as soon as possible, affected individuals can receive care to help alleviate their symptoms as well as slow down the progression of MCI towards Dementia. In this paper, we explore the adoption of speech-based features as well as embeddings from pre-trained models to train machine learning models for detecting MCI. Our results show that using an XGBoost tree algorithm in combination with TRILLsson features as the input can successfully predict MCI with an accuracy of 81%. We compare the answers of patients to two different questions to detect MCI, the first being a “How are you?” question and the second asking the patient to describe a particular picture. Our results show that approaches based on the answers to the picture description question typically perform better than those using the “How are you?” answers.
Reference: Vinod Subramanian, Namhee Kwon, Raymond Brueckner, Nate Blaylock, Henry O’Connell, “Detecting Mild Cognitive Impairment using Vocal Biomarkers from Spontaneous Speech,” in Proceedings of the Interspeech Workshop on Speech, Music and Mind 2024 (SMM), September 2024.
