Large Language Models (LLMs) have revolutionized natural language processing (NLP) with significant advancements in text generation. LLMs often struggle with complex domain-specific tasks, such as medical report analysis, despite their capabilities. This study focuses on enhancing LLM performance for medical applications, particularly in diagnosing and managing cardiogenic pulmonary edema (CPE). This research explores fine-tuning LLMs to develop a real-time CPE chatbot for Intensive Care Units (ICUs). The chatbot aims to provide diagnostic suggestions and explanations based on patient data. In the results, the LLaMa3-8B model performed better in predicting patients' CPE stage and keyword extraction. The accuracies achieved 72% and 86%.
Cardiogenic pulmonary edema encompasses a diversity of subtypes, each of which requires specific treatment strategies. Physicians must have the ability to rapidly identify edema subtypes in medical images to achieve timely intervention and mitigate lung impairment. This study compared six supervised classification models pre-trained and tested using the MIMIC-CXR dataset in the identification of cardiogenic pulmonary edema subtypes. Note that the same analytic methods were consistently applied across the six parameter sets. Fine-tuning the windowing to L: 2500 and W: 3000 resulted in five AUC values that surpassed 0.8, despite the fact that this did not result in peak accuracy across all test data categories. The predictive accuracy for vascular congestion reached 90%, while that of interstitial and alveolar edema reached 96%. This research holds significant potential for the early diagnosis and targeted treatment of cardiogenic pulmonary edema to enhance the standard of patient care.
Large language models have demonstrated significant potential in medical applications, particularly in the construction of structured knowledge databases through retrieval-augmented generation systems. However, the optimization of chunking strategies in medical data processing remains an open challenge, as excessive segmentation may lead to retrieval inefficiencies while overly large chunks risk losing critical information. This study systematically evaluates the impact of chunking strategies on medical database construction, focusing on Traditional Chinese epilepsy guidelines. Using Llama 2-70b to generate structured question-answer datasets, we employ Gemini and Nemotron-4-340B-Instruct to assess data accuracy, consistency, and relevance through a reciprocal evaluation framework. Our findings reveal that while chunk size influences retrieval precision, inference length and response generation have a more significant impact on processing efficiency, with a strong correlation between inference length and generation time $(r= 0.89)$. Furthermore, the evaluation models exhibit notable differences: Gemini assigns higher and more consistent scores with faster processing times, while NVIDIA applies stricter evaluation criteria, resulting in lower scores and longer validation durations. These results highlight the importance of model selection in medical AI applications, as different LLMs prioritize different aspects of data quality. Additionally, the study suggests that optimizing inference structure and refining chunking methods are crucial for enhancing retrieval performance. Future research should explore advanced entity-linking techniques and cost-efficient deployment strategies to improve the scalability and reliability of RAG-based medical knowledge systems.