Artificial Intelligence Tools to Evaluate Language and Speech Patterns in Alzheimer's Disease.

Background: Speech and language problems are one of the earliest signs of neurodegenerative diseases including Alzheimer's disease (AD), but they are often difficult to detect especially in the early stages. Utility of assessing speech and language problems using artificial intelligence (AI) has been examined in the past. However, the accuracy of different methods has been variable, preventing incorporation of these techniques in clinical use. In this study, we present different speech‐based machine learning techniques. We focused on the automatic extraction of a wide array of interpretable features from the speech signals, then compared these features across different neurological disorders. Method: Spoken responses to four different tasks (i.e., Cookie Theft Picture (CTP) and Stroop Task) were recorded from 80 participants, 5 with AD, 28 with Parkinson's disease (PD), 18 normal controls (NCs), and the remaining with other neurological diseases (OTRs). We automatically extracted several acoustic and linguistic features using speech and language technologies. We conducted pairwise Kruskal–Wallis tests to assess how speech and language abilities differed between different neurodegenerative disease vs. control groups. We used SciPy Python library. Result: The total amount of speech time and pauses as well as the average duration and the variance of the number pauses significantly differed between AD and NC and between PD and NC in almost every task (p <0.05). Prosodic features based on F0 contour and F0 on the first and last voiced segment resulted in significant differences across tasks between AD and NC, AD and PD, AD and OTR (p <0.05). The phonological posterior for the phonological classes of consonantal, voiced, labial, nasal, velar, trill, and close significantly differed in almost every task between AD and NC, AD and PD and AD and OTR (p <0.05). Moreover, from the transcriptions of the CTP task, the number of sentences and noun phrases differed significantly between AD and NC, whereas the number of contrastive connectives and the moving average type‐token ratio were significant between NC and OTR (p <0.05). Conclusion: Automated analysis of speech employing machine learning techniques can provide objective assessment of patients with AD that may be useful in distinguishing AD from other neurological disorders. [ABSTRACT FROM AUTHOR]