Uncovering Linguistic Patterns: A Machine Learning Exploration for Early Dementia Detection in Speech Transcripts

Dementia, a prevalent neurodegenerative disorder, demands early detection for effective intervention and support. This paper leverages machine learning and deep learning models to diagnose dementia based on linguistic patterns in speech transcripts. Utilizing datasets from DementiaBank, namely the Pitt Corpus and ADReSS Challenge, we pre-process data, extract linguistic features, and apply 10-fold cross-validation. Classical models (Logistic Regression, Random Forest, K-Nearest Neighbors, Support Vector Machine, Multinomial Naive Bayes) demonstrate varied accuracies, with Random Forest leading at 0.90. Transfer learning using BERT models (bert-base-uncased, distilbert-base-uncased) showcases promising results, with bert-base-uncased-LR3 achieving 0.79 accuracy. This paper underscores the effectiveness of machine learning and transfer learning in dementia detection.