A-79: Implementation of a Speech Analysis Pipeline to Distinguish Between Dementia and Healthy Controls
Abstract
Objective: Existing research suggests that alterations in natural speech patterns can be indicative of cognitive decline. This study aims to determine if analysis of lexical and temporal features of natural speech during a picture description task can distinguish between groups with and without dementia. Method/Data Selection: Transcribed speech data from the Cookie Theft Task was derived from DementiaBank, an open-access database. A speech analysis pipeline was adapted using feature methodology implemented in prior research (Skirrow et al., 2024). High-level models extracted both lexical and temporal features of speech during image description. Logistic regressions examined accuracy of these features in distinguishing healthy controls vs. dementia. Accuracy, F1, and ROC-AUC were examined. Results/Data Synthesis: Logistic regressions revealed that examining combined features of speech together (Lexical + Temporal) was the strongest in all the evaluation outputs for distinguishing between participants with and without dementia. The lexical model provided a meaningful baseline but lower distinction than the combined set. The temporal model was informative however weaker than the combined and lexical models. Conclusions: The current study replicated and extended a language processing pipeline that produced stable benchmark metrics for distinguishing dementia participants based solely on speech features. When temporal and lexical speech features were combined, global class separation of dementia vs. control participants was better than either feature examined separately. Future research could extend to conversational data and incorporate raw audio signals to refine timing-based features.