2024 American Academy of Neurology Abstract Website

Objective:

To leverage trustworthy, interpretable machine learning techniques by evaluating a Decision Tree classification of spontaneous conversational speech samples of participants with Alzheimer's Disease (AD), Amnestic Mild Cognitive Impairment (aMCI), and healthy controls.

Background:

There is an increasing need for early and accurate screening for Alzheimer's Disease (AD) and Amnestic Mild Cognitive Impairment (aMCI). Machine learning applied to spontaneous speech is emerging as a promising, cost-effective, and rapid screening method. However, machine learning methods such as artificial neural networks are often challenging for clinicians to interpret.

Design/Methods:

A cohort of 153 participants was analyzed: 59 AD, 30 aMCI, and 64 healthy controls (HC), all frequency-matched for age and gender. Conversational samples, averaging two minutes, represented typical clinical interactions. Speech features were distilled and the most salient identified using an ANOVA F-test. A Decision Tree Classifier was employed, focusing on 'coherence'—the logical flow and consistency of conversation, operationalized as the cosine distance between word embedding representations of text. Coherence was quantified using GloVe, capturing semantic relationships between words. Different comparison approaches (e.g., between successive windows of text vs within single windows of text), window lengths (i.e., n-grams), and statistics (e.g., average, minimum, maximum, and standard deviation) were considered to identify patterns.

Results:

The classifier registered an accuracy of approximately 95.42%, with a sensitivity of 96.63% for aMCI/AD and a specificity of 93.75% for HC.

Conclusions:

AD and aMCI groups demonstrated heightened conversational coherence compared to HC. Such deviations could indicate shifts in cognitive strategies or neural pathways. The Decision Tree Classifier provides both transparency and commendable performance. Validation in independent samples remains pivotal for ensuring its broader applicability and robustness.