Spoken Language Derived Measures for Detecting Mild Cognitive Impairment

Area of Science:

• Neuropsychology
• Computational Linguistics
• Gerontology

Background:

• Spoken responses in neuropsychological exams offer diagnostic insights beyond performance metrics.
• Linguistic characteristics of speech can distinguish between different subject groups, including healthy and cognitively impaired individuals.

Purpose of the Study:

• To evaluate the effectiveness of spoken language markers in differentiating healthy elderly subjects from those with mild cognitive impairment (MCI).
• To assess the utility of automatically derived speech and linguistic features for MCI detection.

Main Methods:

• Collected audio and transcripts from a spoken narrative recall task.
• Derived speech features (e.g., pause frequency, duration) and linguistic complexity measures.
• Compared measures from manual annotations with those from automatic (forced) alignments and parses.

Main Results:

• Identified statistically significant differences in several spoken language measures between healthy elderly subjects and MCI subjects.
• Demonstrated that these differences are largely maintained when using automated methods.
• Achieved statistically significant improvement in the area under the ROC curve (AUC) for MCI detection by incorporating automatic spoken language features with existing neuropsychological test scores.

Conclusions:

• Spoken language analysis provides valuable, complementary markers for neuropsychological assessment.
• Automated derivation of speech and linguistic features is effective for identifying differences between clinical groups.
• Integrating automatically derived spoken language markers can enhance the accuracy of mild cognitive impairment detection systems.