Digital linguistics biomarkers of coherence and cohesion for the early detection of mild cognitive impairment and dementia

Linguistic deficits emerge early in the course of dementia and are already observable in prodromal stages such as Mild Cognitive Impairment (MCI). Advances in language technology have enabled the automatic analysis of verbal production as a low-cost and non-invasive tool for screening cognitive decline, supporting timely intervention. The integration of digital devices into clinical practice further allows the use of voice as a Digital Linguistic Biomarker (DLB), a measurable indicator of cognitive integrity that can be leveraged for dementia detection and disease monitoring. To date, most DLB research has focused on acoustic, lexical, and syntactic features, while the macrolinguistic level has remained relatively underexplored. However, coherence and cohesion—key properties involved in discourse planning and execution, integrating new information with semantic knowledge and continuously monitoring the communicative context—are also affected by cognitive deterioration. This doctoral project aims to define and automatically extract DLBs of cohesion and coherence from two corpora of Italian spontaneous speech produced by individuals with MCI, early dementia, and healthy controls. The proposed methodology led to the definition of local and global coherence indicators based on text embedding representations, as well as three classes of cohesive devices: reference, lexical iteration, and connectives. In addition, task-specific DLBs were developed to formalize the use of Information Units in a picture description task and the mention of thematic roles in a procedural narrative. The effectiveness of the proposed DLBs in distinguishing cognitively impaired individuals from healthy controls was assessed through significance and distributional analyses, revealing disease-related linguistic patterns. Finally, Machine Learning classification experiments were conducted to discriminate individuals with MCI from healthy controls, addressing the critical boundary between physiological and pathological aging.