Screening for early Alzheimer's disease: enhancing diagnosis with linguistic features and biomarkers

Abstract

Introduction: Research has shown that speech analysis demonstrates sensitivity in detecting early Alzheimer's disease (AD), but the relation between linguistic features and cognitive tests or biomarkers remains unclear. This study aimed to investigate how linguistic features help identify cognitive impairments in patients in the early stages of AD.

Method: This study analyzed connected speech from 80 participants and categorized the participants into early-AD and normal control (NC) groups. The participants underwent amyloid-β positron emission tomography scans, brain magnetic resonance imaging, and comprehensive neuropsychological testing. Participants' speech data from a picture description task were examined. A total of 15 linguistic features were analyzed to classify groups and predict cognitive performance.

Results: We found notable linguistic differences between the early-AD and NC groups in lexical diversity, syntactic complexity, and language disfluency. Using machine learning classifiers (SVM, KNN, and RF), we achieved up to 88% accuracy in distinguishing early-AD patients from normal controls, with mean length of utterance (MLU) and long pauses ratio (LPR) serving as core linguistic indicators. Moreover, the integration of linguistic indicators with biomarkers significantly improved predictive accuracy for AD. Regression analysis also highlighted crucial linguistic features, such as MLU, LPR, Type-to-Token ratio (TTR), and passive construction ratio (PCR), which were sensitive to changes in cognitive function.

Conclusion: Findings support the efficacy of linguistic analysis as a screening tool for the early detection of AD and the assessment of subtle cognitive decline. Integrating linguistic features with biomarkers significantly improved diagnostic accuracy.

Keywords: Alzheimer’s disease; amyloid-β; cognitive impairment; hippocampal volume; linguistic features; speech analysis.

Figure 1

A set of three pictures on Taiwanese culture as visual stimuli to elicit descriptive responses from study participants for speech analysis.

Figure 2

Correlation heatmap of linguistic features and biomarkers variables in early-AD detection. The heatmap displays Pearson correlation coefficients, with color intensity and hue indicating the strength and direction of correlations, respectively. Dark blue represents strong positive correlations (+1), while dark red indicates strong negative correlations (−1).

Figure 3

The curve of the test set for comparing the performance of three different classifier models: support vector machine (SVM), k-nearest neighbors (KNN), and random forest (RF). The left plot (A) shows that when only linguistic features are used, KNN shows the highest AUC. The middle plot (B) shows that KNN shows the highest AUC when using biomarkers alone. In the right plot (C), including biomarker data alongside linguistic features substantially increased model performance for all three classifiers.