摘要
This study aims to investigate whether mild cognitive impairment (MCI) can be detected by analyzing the speech produced by elderly participants when they attempt to describe commonly used cartoon pictures in the field of dementia study. We collected and transcribed a mixed Mandarin and Southern Min speech dataset from elderly subjects in Taiwan. In addition, an English dataset was retrieved from DementiaBank for comparison purposes. Then, prosodic and linguistic features in six categories were computed - including acoustic, speaking rate, filled pause, part-of-speech, syntactic, and lexical richness features. Linear discriminant analysis (LDA) was applied to each of the six feature spaces to examine the data separability between MCI and the healthy control (HC) group. Subsequently, we selected the most discriminative feature categories for both datasets and trained support vector machine (SVM) classifiers on LDA-projected features. Performance of the classifiers is evaluated in terms of F1 score and accuracy via 3-fold cross validation. We discovered that the discriminative power of acoustic and linguistic features generally differ across the two datasets. This creates a challenge for early detection of MCI, and we suggest that the best classifier should be designed in language-specific manners.