Predicting MCI Status From Multimodal Language Data Using Cascaded Classifiers.

Area of Science:

• Computational linguistics
• Neuroscience
• Machine learning

Background:

• Automated language analysis shows promise for detecting mild cognitive impairment (MCI).
• Existing studies often focus on single language tasks, limiting comprehensive analysis.

Purpose of the Study:

• To investigate a cascaded approach combining multiple language tasks for improved MCI detection.
• To compare the performance of this multimodal approach against single-task classifiers and traditional neuropsychological tests.

Main Methods:

• Participants (26 MCI, 29 controls) completed picture description, silent reading, and reading aloud tasks.
• Multimodal data (audio, text, eye-tracking, comprehension) were collected and features extracted.
• A cascaded classifier architecture was employed, integrating predictions at feature, mode, task, and session levels.

Main Results:

• The cascaded approach combining data at the task level achieved an AUC of 0.88 and accuracy of 0.83.
• This outperformed single-task classifiers, neuropsychological tests (AUC=0.75), and early fusion multimodal classification (AUC=0.79).
• Further improvement to AUC=0.90 and accuracy=0.84 was achieved by integrating language and neuropsychological classifier predictions.

Conclusions:

• A cascaded, multimodal language analysis approach offers superior performance for MCI detection.
• This modular system enhances interpretability and can be extended to include diverse data types for broader applications.