Large language models (LLMs) are advanced deep learning models with billions or even trillions of parameters, enabling powerful natural language processing and knowledge reasoning capabilities. Their applications in the medical domain have been rapidly expanding, spanning medical research, clinical diagnosis, drug development, and patient management. As a cornerstone of China’s healthcare system, traditional Chinese medicine (TCM) faces significant challenges, including difficulties in knowledge extraction, and lack of standardization. The emergence of TCM-focused LLMs presents a transformative opportunity, offering a novel technological framework to process vast amounts of TCM data, uncover hidden theoretical insights, and enhance both research and clinical applications. Despite the growing interest in artificial intelligence (AI)-driven medical solutions, systematic research on LLMs in the TCM domain remains limited. This article provides a comprehensive review of LLM development, detailing their underlying mechanisms, training methodologies, and key technological advancements. It further explores the unique characteristics and diverse application scenarios of existing TCM-LLMs. Additionally, this study also conducts a horizontal comparison of the differences between intelligent question-answering (QA) systems on general LLMs and QA systems on TCM-LLMs, discusses challenges and potential risks, and offers strategic recommendations for future development. By synthesizing current advancements and addressing critical gaps, this work aims to support the continued modernization and intelligent evolution of TCM, fostering its integration into contemporary healthcare systems.

Alzheimer’s disease (AD) is the leading cause of dementia, affecting millions worldwide and characterized by progressive neuronal loss, brain atrophy, and cognitive decline. Early diagnosis of AD is crucial for improving treatment outcomes and prolonging patients' quality of life, as timely interventions at the earliest stages can slow disease progression and enhance survival rates. However, accurate early detection remains challenging due to overlapping features across disease stages and the complexity of neuroimaging data. Traditional diagnostic approaches and conventional machine learning techniques often struggle to capture subtle variations in early-stage AD, limiting their diagnostic efficacy. To address these limitations, we propose a novel deep learning framework that enhances the diagnostic performance of a pre-trained ResNet-50 model by integrating a Convolutional Block Attention Module (CBAM) and Multi-Head Self-Attention (MHSA). Our framework leverages attention mechanisms to refine feature representations at an early processing stage, enabling precise differentiation between disease stages. The integration of Cross-Attention further improves the model's ability to extract discriminative features from neuroimaging data, optimizing predictive performance. Trained and evaluated on the OASIS Kaggle and MIRIAD datasets, the model achieves state-of-the-art results, including an accuracy and AUC of 99.17% and 99.00% on the OASIS dataset, and an accuracy of 98.59% and AUC of 98.44% on the MIRIAD dataset. Its near-perfect performance across all evaluated metrics establishes its potential as a transformative tool for early AD diagnosis.