A LLMs-Assisted Framework for Parkinson's Disease Assessment Based on PPMI Dataset

The assessment of diseases such as Parkinson's disease, which are chronic neurodegenerative conditions, is an extremely complex and time-consuming process. Self-assessment scales or interviews in the outpatient often lead to the loss of key information. Large Language Models (LLMs) show abilities in capturing subtle linguistic differences and handling long texts, enabling the extraction of a significant amount of key information while ensuring data integrity and user privacy. We focus attention on people with Parkinson's disease (PwP). In response to these issues, we propose a framework in this paper. Firstly, we use large language model (LLM) with chain-of-thought (CoT) prompt to rephrase patient self-report texts from real-outpatient structured data as a supervised fine-tuning (SFT) corpus. Secondly, we fine-tune bidirectional encoder representation from transformers (BERT) with Low-Rank Adaptation (LoRA) to enable it to under-stand and extract the semanteme of self-report, thus predicting each MDS-UPDRS item. Finally, we designed experiments on a large amount of test data to evaluate the effectiveness of the framework. The results indicate that the accuracy on this task has been improved to 95.36%, which is a 6.7% increase compared to the best-performing model.