Patient-Reported and Technology-Assisted Monitoring in Orthotic Management of Adolescents With Idiopathic Scoliosis: Scoping Review.

Adverse drug events (ADEs) are defined as harmful reactions that are unrelated to the intended therapeutic effects of medications when they are administered at standard dosages and according to standard regimens. ADEs constitute a significant global public health concern. They are among the main causes of hospitalization and mortality in both developed and developing countries []. Meta-analyses reveal that approximately 5%‐10% of patients in health care institutions experience ADEs []. Furthermore, the economic burden attributable to ADEs within health care systems worldwide exceeds US $42 billion annually []. The precise identification and comprehensive evaluation of ADEs, including elucidating their underlying etiologies, are imperative for mitigating harm and augmenting the quality of clinical care []. Consequently, the surveillance and management of ADEs have emerged as critical public health priorities, wherein the accurate extraction of ADE-related information can significantly enhance drug safety and foster rational pharmacotherapy. A substantial proportion of ADE information remains embedded within unstructured, narrative clinical notes, presenting formidable challenges to conventional manual review and extraction methodologies, which are often inefficient and labor-intensive [,]. The advent of large language model (LLM)–based tools, such as ChatGPT, offers a promising way to efficiently and accessibly identify and retrieve ADE information []. Trained on vast corpora of textual data, LLMs demonstrate remarkable abilities in cross-domain text comprehension, logical inference, and human-like natural language generation []. Empirical studies have demonstrated their utility across diverse medical applications, including disease diagnosis and management [], patient education and counseling [], clinical text analysis [], and postoperative risk stratification []. Despite growing evidence attesting to the value of LLMs in multiple medical domains, their potential in pharmacovigilance, particularly in the detection of ADEs in clinical notes, remains insufficiently explored. Furthermore, as LLMs are primarily pretrained on publicly available datasets without domain-specific clinical fine-tuning, they are prone to generating “hallucinations,” a phenomenon warranting heightened vigilance in health care contexts [,]. For instance, Williams et al [] found that GPT-4 and GPT-3.5-turbo models produced fabricated patient visit summaries at an alarming rate of up to 42%. Retrieval-augmented generation (RAG) architectures enhance the performance of LLMs by integrating external information retrieval mechanisms, thus improving response accuracy and practical applicability []. By dynamically combining domain-specific knowledge bases with user queries, RAG provides comprehensive, high-quality contextual information, reducing the likelihood of erroneous model outputs and effectively addressing the hallucination challenge [,]. Additionally, RAG affords access to the latest reliable knowledge without the substantial costs of extensive model fine-tuning []. Recent studies have highlighted RAG’s superiority over standard LLMs in biomedical tasks, including question answering, text and image generation, and clinical scenario interpretation []. For instance, Li et al [] showed that combining RAG with LLMs substantially improves the accuracy and reliability of COVID-19 fact-checking, successfully overcoming the inherent hallucination and context-inaccuracy issues []. Nonetheless, the efficacy of RAG depends on the availability of trustworthy, unbiased data, and the quality of external domain knowledge significantly affects performance []. Currently, ADE-related knowledge is fragmented and lacks structured representation, posing significant barriers to the deployment of RAG and LLM frameworks for ADE extraction. Therefore, there is a compelling need to construct a high-quality, structured Chinese ADE corpus.