Towards the Speech Features of Mild Cognitive Impairment: Universal Evidence from Structured and Unstructured Connected Speech of Chinese

Tianqi Wang, Chongyuan Lian, Jingshen Pan, Quanlei Yan, Feiqi Zhu, Manwa L. Ng, Lan Wang, Nan YanInterspeech 2019

TL;DR

This study investigates how speech patterns change in individuals with Mild Cognitive Impairment (MCI) by analyzing connected speech from both structured (picture description) and spontaneous (self-introduction) tasks in Mandarin Chinese speakers. They found that while lexical features showed little change, acoustic, semantic, and speech fluency features, along with syntactic complexity, were significantly altered in MCI patients, suggesting their potential as early biomarkers for cognitive decline.

What Problem Does This Paper Address?

The early detection of cognitive decline, particularly Mild Cognitive Impairment (MCI), is crucial for timely intervention and management of neurodegenerative diseases like Alzheimer's. While traditional neuropsychological tests have limitations in identifying very early stages of decline, analyzing connected speech has emerged as a promising avenue. However, previous research has often relied on single speech tasks (either structured or spontaneous) and has been limited in its scope, particularly for non-English languages like Chinese. This study aims to bridge this gap by examining a broader range of speech features across two different speech elicitation tasks in a Chinese-speaking population, seeking universal evidence of MCI-related speech alterations.

Methods

The researchers employed a sophisticated statistical approach using linear mixed-effect regression models. This method is particularly well-suited for analyzing complex linguistic data because it can account for multiple sources of variation simultaneously. Specifically, it allowed them to model:

  • Fixed effects: The primary factors of interest, namely the group (MCI vs. healthy control) and age.
  • Random effects: Variations introduced by individual participants and the specific speech task (picture description vs. self-introduction).

This approach enables a more robust analysis of connected speech, moving beyond the limitations of single-task studies and allowing for the identification of features that are consistently altered across different speaking contexts. They extracted a comprehensive set of features, categorized into:

  • Lexical: Total words, part-of-speech (POS) rates, content density, type-token ratio, Brunét's index, Honoré's index, word/vocabulary frequency.
  • Semantic: Semantic units, idea density, idea efficiency.
  • Syntactic: Utterance length, dependent elements linked to the noun, syntactic complexity.
  • Speech fluency: Filled pause count/rate, disfluency count/rate.
  • Acoustic: Silence duration, speech duration, voice-silence ratio, verbal rate, standard phonation rate.

Datasets

The study utilized speech samples from 75 elderly native Mandarin Chinese speakers. This cohort comprised:

  • 19 individuals diagnosed with Mild Cognitive Impairment (MCI).
  • 56 healthy elderly controls.

The speech data was collected through two distinct tasks:

  1. Structured connected speech: Elicited via the "Cookie Theft" picture description task from the Boston Diagnostic Aphasia Examination (BDAE).
  2. Spontaneous connected speech: Elicited through predefined open-ended questions designed to prompt self-introduction and discussion on topics like family, career, and hobbies.

All speech samples were constrained to one minute in duration.

Key Insights and Ideas

  • Lexical features are less sensitive in Chinese MCI: Unlike findings in Indo-European languages, lexical features like POS rates and lexical richness measures did not show significant differences between MCI and control groups in this Chinese sample. This suggests that the linguistic structure of Chinese might influence which features are most affected by early cognitive decline.
  • Acoustic and fluency markers are strong indicators: Features related to the temporal aspects of speech (silence duration, speech duration, voice-silence ratio, verbal rate) and speech fluency (disfluencies, filled pauses) were significantly altered in MCI patients. This highlights the importance of the how of speech production, not just the what.
  • Semantic content suffers early: A reduction in semantic units and lower idea efficiency were observed in the MCI group, indicating that the ability to convey meaningful content is compromised early in the cognitive decline process.
  • Syntactic complexity shows subtle changes: While overall syntactic complexity didn't differ significantly, reduced utterance length and less complex noun phrase relations were noted, suggesting a more nuanced impact on sentence structure.
  • Task-agnostic evidence is crucial: By using both structured and spontaneous speech tasks, the study provides evidence for features that are consistently affected by MCI, regardless of the speaking context. This strengthens their potential as universal biomarkers.

Key Results

The study found significant differences between the MCI and control groups across several feature categories:

  • Lexical:
    • Total words/characters: Significantly lower in MCI participants in both tasks.
  • Vocabulary frequency: Significantly higher in MCI participants in the picture description task, which is an unexpected finding and warrants further investigation.
  • Brunét's index and Honoré's index: Showed contradictory trends, suggesting complexity in interpreting lexical richness.
  • Semantic:
    • Semantic units: Significantly reduced in MCI participants across both tasks.
  • Idea efficiency: Significantly lower in MCI participants across both tasks.
  • Syntactic:
    • Utterance length: Significantly shorter in MCI participants in both tasks.
  • Dependent element linked to the noun: Significantly lower in MCI participants in both tasks.
  • Speech fluency:
    • Disfluency: Significantly higher in MCI participants in both tasks.
  • Disfluency rate: Significantly higher in MCI participants in both tasks.
  • Acoustic:
    • Silence duration: Significantly longer in MCI participants in both tasks.
  • Speech duration: Significantly shorter in MCI participants in both tasks.
  • Voice-silence ratio: Significantly lower in MCI participants in both tasks.
  • Verbal rate: Significantly slower in MCI participants in both tasks.

Limitations and Future Work

  • Lexical features in Chinese: The study notes that the lack of significant findings in well-established lexical features like content density might be due to the intact cognitive functions of lexical retrieval in their selected MCI participants. This suggests that even salient features need careful interpretation, and further neurophysiological studies might be needed to confirm their nature.
  • Generalizability: While the study provides evidence for Chinese speakers, further research is needed to confirm the universal applicability of these findings across diverse linguistic backgrounds and cultural contexts.
  • Future directions: The authors suggest that the identified features (acoustic, semantic, speech fluency, and syntactic complexity) hold significant promise for the screening and diagnosis of suspected MCI. Future work could focus on developing automated systems that leverage these features for early detection.

Summary updated

BibTeX
@inproceedings{wang2019towards,
  author = {Tianqi Wang and Chongyuan Lian and Jingshen Pan and Quanlei Yan and Feiqi Zhu and Manwa L. Ng and Lan Wang and Nan Yan},
  title = {{Towards the Speech Features of Mild Cognitive Impairment: Universal Evidence from Structured and Unstructured Connected Speech of Chinese}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure