Fusion of Traditional and Large Language Models for Linguistic Analysis in Alzheimer's Disease Detection

Language-based markers hold potential for the detection of Alzheimer's disease (AD), offering a non-invasive and accessible alternative to traditional methods. Previous studies have leveraged traditional language models (TLM) or large language models (LLM) for AD detection, but their effectiveness is limited by the lack of annotated datasets for training. To address this limitation, this work proposes a fusion of TLM and LLM for classification. Transcripts of speech from 156 AD and non-AD subjects are preprocessed and used to extract linguistic features for TLM classifiers, including random forest (RF), which achieves a test accuracy of 89.6%. In parallel, fine-tuning a pre-trained BERT model on the same dataset results in a test accuracy of 87.5%. The outputs from RF and BERT are then fused using a multilayer perceptron, which achieves a test accuracy of 91.7%. Additionally, word cloud analysis highlights distinct patterns in filler word usage, such as increased reliance on "uh" in AD speech versus "um" in non-AD speech, reflecting linguistic differences linked to cognitive decline. The proposed fusion of traditional and large language models improves AD detection and presents a scalable framework for language-based diagnostic systems.