Automatic classification of AD versus FTLD pathology using speech analysis in a biologically confirmed cohort

Abstract

Background: Logopenic variant primary progressive aphasia (lvPPA) is a non‐amnestic presentation of Alzheimer’s disease (AD). Many studies have proposed automatic classification systems comparing AD to healthy controls (HC), showing an accuracy range of 70%–88%, but have not included lvPPA, nor distinguished patients with AD neuropathologic change (ADNC) from patients with other neurodegenerative changes. We propose automatic classification systems that utilize language features to detect ADNC. Method: Using automatic language processing tools, we extracted 34 language features from digitized 1‐minute picture descriptions produced by 69 patients with likely ADNC (49 amnestic AD (62.6±7.6y), 28 non‐amnestic lvPPA (63.2±7.1y); confirmed by autopsy or CSF analytes consistent with ADNC), 20 PPA patients with likely FTLD‐tau pathology (67.8±7y), and 35 HC (64.6±7y). Extracted features include part‐of‐speech categories per 100 words, lexical characteristics of content words (e.g. frequency, concreteness), and acoustic properties (e.g. pause rate, articulation rate). Multiple random forest (RF) and support vector machine (SVM) classifiers were built for three binary (ADNC vs. HC; ADNC vs. FTLD‐tau; lvPPA vs. FTLD‐tau) and one three‐way classifications (ADNC vs. FTLD‐tau vs. HC). We performed leave‐one‐out cross‐validation and feature selection with Pearson correlations (r <0.5 to 0.9) and t‐tests/ANOVA (p=0.01 or 0.05) in all models, and reported the best performance after hyperparameter tuning. Result: The RF classifier of ADNC versus HC (features selected at r<0.7, p=0.01) had 90% accuracy (sensitivity=0.95, specificity=0.8, AUC=0.94) with prepositions, number of phonemes, and pause rate selected as the most important features. The SVM classifier of ADNC versus FTLD‐tau showed 93% accuracy (sensitivity=0.99, specificity=0.7, AUC=0.9), selecting features such as nouns, pronouns, and articulation rate (p=0.05). The RF classifier of lvPPA versus FTLD‐tau showed 85% accuracy (sensitivity=0.93, specificity=0.75, AUC=0.92), selecting features such as nouns, articulation rate, and total words (r<0.6, p<0.05). The RF multiclass classifier had 80% accuracy (sensitivity=0.92, macro‐average recall=0.7, macro‐average AUC=0.88) with word frequency, articulation rate, and number of prepositions selected as the most important features (r<0.6, p=0.01). Conclusion: Automatic classifiers trained with language features extracted from one‐minute picture descriptions using our automatic pipelines showed high accuracy in identifying the pathological grouping of patients, which can potentially be used in the screening stage.