Currently, the incidence and mortality rates of gastrointestinal (GI) diseases remain high, with over 10 million patients worldwide dying from GI diseases annually. In China, one-fifth of the population suffers from GI diseases, the highest globally, posing significant challenges to healthcare resources [1], particularly concerning GI cancers. Global data indicates that GI cancers account for 26.3 % of cancer cases and 35.4 % of deaths, showing an upward trend [2]. Given the substantial economic burden that GI diseases impose on patients, families, and society as a whole, early and effective management of these diseases becomes critically important.

GI diseases have a wide range of characteristics. Specifically, GI diseases are influenced by multiple factors, and their clinical manifestations are highly variable, posing significant challenges to diagnosis and treatment. Secondly, due to the chronic nature of GI diseases, they often require long-term or even lifelong treatment and management [3]. Additionally, the interaction of the gut-brain axis makes GI diseases susceptible to emotional states. This is particularly true for patients with functional gastrointestinal disorders, where emotional states have a profound impact on disease progression and treatment outcomes. As the central part of the healthcare process, it is crucial for patients and their families to clearly understand the characteristics of the disease and the interventions required, which is essential for achieving a good prognosis [4,5].

However, due to a scarcity of medical resources, patients' diagnostic and treatment needs are difficult to meet, and there is a lack of disease management and prevention knowledge, making it challenging to alleviate anxiety. Research indicates that the majority of GI disease patients spend less than 5 h per year in face-to-face communication with healthcare professionals [6]. In most cases, patients need to manage their conditions themselves. However, patients with GI diseases generally have low health literacy in self-management. A survey of 80 patients with Inflammatory Bowel Disease (IBD) shows that only 66 % correctly answer questions about medication dosage, 15 % about potential side effects, and 16.7 % about the impact of smoking on their disease [7]. While the internet serves as a source of health information for patients, the complexity of professional literature and potential misinformation can further confuse patients about their conditions [8]. Therefore, helping patients accurately understand their own diseases and enhancing their self-management capabilities is a new challenge in healthcare [9].

With the advancement of natural language processing (NLP) technology, medical large language models (LLMs), such as BenTsao and HuatuoGPT [10,11], have shown immense potential in various medical fields, including disease prevention, diagnosis, treatment, monitoring, and patient support [12,13]. However, their application in healthcare faces several challenges. Firstly, medical LLMs typically use online medical question-answering data as the source of medical knowledge. The accuracy and authority of these training data are often insufficient, leading to inaccuracies and unreliability in the models. Additionally, most current medical LLMs focus on general medicine and have not yet conducted in-depth research in the field of GI diseases. Last but not least, existing LLMs often employ efficient fine-tuning techniques like Low-Rank Adaptation (LoRA) technology [14]. LoRA indirectly trains changes in some dense layers of neural networks by optimizing the rank decomposition matrices of dense layers while keeping the pre-trained weights frozen. Although LoRA reduces the number of trainable parameters by introducing low-rank matrices, excessive parameter reduction can slow down the convergence speed, while insufficient reduction may lead to overfitting. Meanwhile, existing improvements further reduce memory consumption by freezing low-rank matrices (such as matrix A), but at the cost of the model’s ability to learn from new data and adapt.

To address the above challenges, we develop a Chinese medical LLM tailored for GI diseases, named GutGPT (where “Gut” denotes the gastrointestinal system, and “GPT” refers to the Generative Pre-trained Transformer architecture), using fine-tuning techniques. Firstly, we construct a high-quality gastrointestinal disease question-answering dataset consisting of 191,615 entries. The data is collected and curated through both manual and automated methods, leveraging real-world doctor-patient dialogues, clinical guideline data, medical knowledge graph information, and medical licensing examination questions. Subsequently, we propose LoRA with self-attention mechanism parameter sharing, effectively balances trainable parameter selection while optimizing model adaptability and minimizing memory requirements. By sharing low-rank matrices across different layers, our model significantly reduces trainable parameters and memory usage while maintaining the unique tuning capabilities of each layer. This enhances model robustness and transferability while achieving a better balance between parameter efficiency and model adaptability. The integration of a high-quality question-answering dataset with LoRA technology with self-attention mechanism parameter sharing not only improves the model's diagnostic accuracy and ability to provide personalized treatment recommendations but also aligns the model with medical professionals, enabling it to conduct medical diagnoses akin to doctors and imbuing it with humanistic care capabilities. Furthermore, we evaluate the performance of our medical large language model from two perspectives: expert evaluation and public benchmark datasets. Compared to 13 general LLMs and 3 medical LLMs, our approach achieves state-of-the-art performance. In expert evaluations, GutGPT improves diagnostic accuracy by 9.59 % over the baselines. On two public medical question-answering datasets, CMB and CMExam, it achieves an average accuracy improvement of 22.47 %. Our contributions can be summarized as follows:(1)

We propose a medical LLM specifically for GI diseases. To the best of our knowledge, this is the first LLM tailored for GI diseases, providing insights for the development of intelligent medical assistants in the future.

(2)

We propose LoRA technology with self-attention mechanism parameter sharing, which effectively balances the selection of trainable parameters while optimizing model adaptability and minimizing memory requirements.

(3)

We develop an instruction-tuned dataset for GI diseases, which includes real-world doctor-patient dialogues, medical knowledge graphs, medical guidelines, and Chinese medical licensing exam data. This integration of diverse data sources ensures that our model possesses accurate and authoritative knowledge in the field of GI diseases.

Statement of significance