Detection of Alzheimer's Disease Using Fine-Tuned Large Language Models

Since there is no known cure for Alzheimer's disease (AD), early detection is essential to controlling its progression.Because of the high cost and invasiveness of traditional diagnostic techniques like MRIs and pathological testing, researchers are looking into less expensive alternatives that use machine learning (ML) and natural language processing (NLP). By evaluating their performance against traditional ML and deep learning (DL) techniques, this study explores the possibility of using fine-tuned open-source large language models (LLMs) to identify AD through linguistic analysis. To optimize models like Qwen1.5–7B and OLMo1.7–7B, we used supervised fine-tuning (SFT) with parameter-efficient techniques like LoRA and QLoRA on the Pitt Corpus dataset, which consists of speech transcripts from the “Cookie Theft” picture description task. The findings showed that LLMs performed noticeably better than conventional techniques; Qwen1.5–7B had an F1-score of 0.8824, which was higher than CNN (0.7987), LSTM (0.7689), and logistic regression (0.83). The study demonstrates how LLMs can detect subtle linguistic impairments in AD that are difficult for traditional models to identify, like syntactic errors and repetitions. The comparatively small dataset size and exclusive reliance on textual data are limitations, though, and it is recommended that future studies include multimodal inputs and more varied datasets. Despite limitations, the results highlight the potential of optimized LLMs as scalable, non-invasive methods for early AD detection, providing a way to enhance patient care and diagnostic precision. Through this study, a novel, accurate, and reliable method for early diagnosis of Alzheimer's disease patients can be provided.