Utilising speech-derived biomarkers to detect Alzheimer's disease with BERT-based language models: a machine learning study.

Background: Early diagnosis is critical for effective management of Alzheimer's disease (AD). While prior studies have shown that speech features can be indicative of AD, most existing work consolidates multiple biomarkers, making it difficult to isolate the contribution of individual features. Objective: This study systematically isolates individual speech biomarkers to quantify their distinct contributions to AD classification performance of language models (LMs) and determine whether targeted biomarker selection improves over consolidated feature sets. Methods: We processed speech transcriptions from DementiaBank to surface discriminatory speech biomarkers-verbal pauses, disfluencies, and unintelligible words. We then fine-tuned and evaluated lightweight LMs (BERT, AlBERT, and DistilBERT) on these biomarker-conditioned transcripts for automatic AD classification. Results: Pauses emerged as the most discriminatory speech biomarker (F1 = 0.8326), significantly outperforming the No-biomarker baseline and other biomarkers. Combining all biomarkers degraded performance relative to pauses alone on average across models, though the effect was model-dependent: BERT's All-biomarkers condition exceeded its own No-biomarker baseline, suggesting that feature combination benefits higher-capacity models. BERT yielded the best performance (F1 = 0.8154) across conditions. Conclusions: Selective use of speech biomarkers such as pauses can meaningfully improve AD detection with lightweight LMs, suggesting that targeted biomarker selection may offer a more interpretable and clinically actionable path than broad feature consolidation.