Conception et évaluation d’un outil d’extraction, d’analyse et de classification de marqueurs linguistiques de la maladie d’Alzheimer

Abstract

Alzheimer’s is a neurodegenerative disease that leads to a progressive deterioration in cognitive functions and can be observed by a decline in language functions. Many researchers have turned to the use of automated computer processes to analyze language biomarkers and detect signs of the disease in a non-invasive way. Over the past five years, LiNCS has contributed to this research by developing an application dedicated to extracting linguistic features and classifying participants. However, it remains essential to be able to extract and analyze relevant measures, manipulate and visualize them, and rely on a reusable and reliable classifier. This thesis presents a solution that builds on previous LiNCS work and proposes a redesigned and improved application. Our approach implements a reliable system, based on optimized feature extraction methods, with relevant data manipulation and visualization tools. It integrates to the classifier significant features derived from the analysis of three categories less explored in the literature : coreference chains, pauses, and syntactic complexity. The application also features an optimized classification pipeline using an XGBoost algorithm as the basis for the predictive model. The results obtained using our system showed that short pauses at the beginning of sentences, ADVP, ADJP, VP, S and SBARQ phrase ratio measures, as well as FRAG, VP, NP and ADJP phrase density measures, correlated with diagnostic labels (F > 4, p-value < 0.05) and contributed to better classification. Our tool also enabled us to eliminate 17 noise-generating features. The final classification model achieved an F1 score of 80.9% and a recall of 80,5%, an improvement of 5% and 8,2% respectively. The design of our tool provides a powerful system for observing language decline and identifying features associated with the diagnosis of Alzheimer’s disease. It provides a solid basis for future research and development, and offers a tool that can be used by healthcare professionals to monitor patients’ progress.

Date	3 Sept 2025
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Sylvie Ratté (Supervisor) & Pierre André Ménard (Co-supervisor)