Transcript-Based Cognitive Decline Detection Using Word-Character TF-IDF Fusion

Abstract

Alzheimer is a progressive neurodegenerative condition which impacts linguistic ability. The research demonstrates a transcript-based framework to assist in detecting cognitive decline. The framework uses conventional textual representations based on term frequency-inverse document frequency (TF–IDF) features and machine learning based classifiers. The study tests Linear Support Vector Machine (SVM), Logistic Regression, Random Forest, Multinomial Naive Bayes, and K-Nearest Neighbors (KNN), and also evaluates a feature-fusion strategy using multiple representations. We conducted experiments on the ADReSS 2020 dataset, which contains 108 speech transcripts. The transcripts are balanced and consist of 54 dementia and 54 control participants. For feature extraction, we investigated two approaches: (i) TF–IDF at the word level with unigram and bigram features; (ii) TF–IDF at the character level with character n-grams (3–5 characters) for identifying subword structures and disfluencies in spontaneous speech. For performance evaluation, metrics such as accuracy, precision, recall and macro F1 score were based on stratified 5-fold cross-validation. The linear SVM classifier operating on the fusion representation gave the best performance among the linear classifiers. The accuracy of the fusion representation was 87.0% with a corresponding macro F1 score of 0.868. In the case of the character-level representation, an accuracy of 86% (F1 = 0.858) was achieved, while in the case of the word-level representation, an accuracy of 82.4% (F1 = 0.822) was achieved. These results show that lightweight TF–IDF representations with classical machine learning models can reveal linguistic markers of cognitive decline in speech transcripts without requiring acoustic analysis or heavy deep learning architectures.