Abstract
BACKGROUND: Speech and language impairments, long recognized as early symptoms of Alzheimer's disease (AD), can now be quantified with unprecedented precision due to recent advances in natural language processing (NLP) and artificial intelligence (AI). Despite growing interest in AI-enabled speech biomarkers, few studies have linked spontaneous speech to biologically verified AD, and most have focused on English-language data or acoustic features with limited linguistic interpretability. Here, we present the first end-to-end machine-learning framework for automatic AD detection from German speech, using clinical-biological criteria validated by cerebrospinal fluid (CSF) biomarkers. METHODS: 44 participants were included: 22 biomarker-defined AD cases from a prospective observational study (German Clinical Trials Register, DRKS00030633) and 22 socio-demographically matched cognitively healthy controls (CHC). Connected speech was elicited using the standardized Cookie Theft picture description task. Recordings were transcribed with a state-of-the-art automatic speech recognition (ASR) system. From these transcripts, 32 theory-driven linguistic biomarkers were computed with an advanced NLP tool, falling into three categories: information-theoretic, lexical richness, and syntactic. AD-versus-CHC classification used five supervised models (logistic regression, support vector machine with a radial basis function kernel, random forest, gradient boosting, XGBoost) under stratified five-fold cross-validation, with stability-based recursive feature elimination performed within training folds. Model interpretability was assessed using SHapley Additive exPlanations (SHAP). RESULTS: Recursive feature elimination retained seven of 32 candidate speech biomarkers as consistently informative across folds. Trained on this subset, all classifiers showed strong discrimination between biomarker-defined AD and CHC. Logistic regression, SVM, random forest, and gradient boosting achieved ~91% mean accuracy with F1 ≈ 0.90 and sensitivity ≈ 0.90, while XGBoost was slightly lower (~89% accuracy). SHAP analyses indicated that model decisions were primarily driven by information-theoretic and structural markers: lower compressibility, reduced lexical density, shorter clauses and sentences, and weaker predictive sequencing indexed by higher-order n-gram statistics. CONCLUSION: Clinically meaningful linguistic biomarkers can be robustly derived from spontaneous speech, even in small, well-characterized clinical samples. Theory-driven features and stability-focused modeling show that information-theoretic and structural properties of connected speech capture core Alzheimer-related impairments with robust classification performance. These findings support AI-enabled speech analysis as a non-invasive, scalable complement to established biological biomarkers of Alzheimer's disease.