• Speech samples were obtained from a publicly available, de-identified ADReSS dataset from DementiaBank5, containing recordings from  individuals diagnosed with AD and HC speakers. 

• Samples included Cookie Theftpicture description transcripts from 48 participants: 24 with Alzheimer's disease (AD) and 24 healthy controls (HC), matched by age (AD: 66.1 ± 7.4 years; HC: 66.1 ± 7.1 years) and gender. The transcripts were obtained from DementiaBank as CHA files.

• All analyses, including LLM use, followed DementiaBank data-use policies.5

• Formulaic language was identified and counted using a structured, prompt-guided GPT-5 procedure. 

• Formulaic language ratios were calculated for each participant sample. 

• Mini-Mental State Examination (MMSE) scores from the dataset were used to examine the relationship between formulaic language use and cognitive status to evaluate its predictive value (AD: 19.5 ± 5.4; HC: 28.8 ± 1.5). 

• Agreement between the researcher and GPT-5 was assessed using Intra Class Coefficient (ICC) on 50% of randomly selected samples. 

• Inter-rater reliability analysis revealed excellent agreement between the GPT-5 and the human expert for formulaic language ratios (ICC = .91)

• Formulaic language ratio achieved an AUC of .71, indicating fair and statistically significant classification of AD vs HC (p = .005).

• Across all participants, formulaic language ratio was negatively associated with MMSE scores (r = −.421, p = .003) and significantly predicted MMSE, accounting for 17.8% of the variance in cognitive status (R² = .178, p = .003).

• Among individuals with AD, formulaic utterance ratio was not a significant predictor of MMSE scores (R² = .123, p = .093).

Discussion

• AD participants produced higher overall formulaic language ratios than HC, consistent with previous studies.

• Formulaic language showed potential as a biomarker for AD diagnosis  but limited utility for predicting cognitive severity (MMSE). 

• These findings support the Dual-Process Model of Language3, where cortical systems process novel/non-formulaic language and subcortical system process formulaic language.

• LLM-based analysis offers a scalable way to detect formulaic language patterns, extending prior manual annotation work and supporting the Dual-Process Model of Language.