Background: Retrieval-augmented generation (RAG) is increasingly used to improve large language models in the medical and nursing domains. However, a comprehensive understanding of its specific architecture and applications in medical and nursing reasoning remains limited. Objective: We aimed to summarize the current state, existing limitations, and future development directions of RAG in the medical and nursing domains. Methods: The PubMed, Web of Science, IEEE Xplore, and arXiv databases were searched for relevant articles using queries that combined terms related to RAG, medical, and nursing domains, covering the period from November 1, 2022, to May 31, 2025. This review was conducted following the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews) guidelines. Results: A total of 917 articles were retrieved, of which 67 met the inclusion criteria. Most studies focused on the medical domain (63/67, 94%), while only a few addressed nursing applications (4/67, 6%). The RAG frameworks included in this review were categorized into 5 functional types: text-based RAG (36/67, 54%), knowledge graph–enhanced RAG (17/67, 25%), agentic RAG (6/67, 9%), multimodal RAG (2/67, 3%), and plug-and-play RAG (6/67, 9%). On the basis of the Simon decision-making process theory, we divided the RAG workflow into 4 stages: intent recognition, knowledge retrieval, knowledge integration, and generation. Only 26 studies included explicit reasoning support, and few were aligned with real-world clinical workflows. Only 12 studies attempted to address ethical considerations related to RAG. Conclusions: We identified 4 key shifts in recent RAG development: shifting from surface-level matching toward contextualized intent recognition, from vague semantics toward logic-driven dynamic retrieval, from passive toward active knowledge retrieval, and from simple aggregation toward coherent context construction. However, most RAG systems in the medical and nursing domains have not yet introduced reasoning methods, and those that have are still predominantly reliant on data‑driven associations without causal modeling. This highlights the need to integrate causal mechanisms for more effective and domain-relevant reasoning in health care. Trial Registration: OSF Registries 10.17605/OSF.IO/WBSV5; https://osf.io/wbsv5

Background: Nonadherence to medication among patients with cardiovascular diseases undermines the desired therapeutic outcomes. eHealth interventions emerge as promising strategies to effectively tackle this issue. Objective: The aim of this study was to conduct a network meta-analysis (NMA) to compare and rank the efficacy of various eHealth interventions in improving medication adherence among patients with cardiovascular diseases (CVDs). Methods: A systematic search strategy was conducted in PubMed, Embase, Web of Science, Cochrane, China National Knowledge Infrastructure Library (CNKI), China Science and Technology Journal Database (Weipu), and WanFang databases to search for randomized controlled trials (RCTs) published from their inception on January 15, 2024. We carried out a frequentist NMA to compare the efficacy of various eHealth interventions. The quality of the literature was assessed using the risk of bias tool from the Cochrane Handbook (version 2.0), and extracted data were analyzed using Stata16.0 (StataCorp LLC) and RevMan5.4 software (Cochrane Collaboration). The certainty of evidence was evaluated using the Grading of Recommendations, Assessment, Development, and Evaluations (GRADE) approach. Results: A total of 21 RCTs involving 3904 patients were enrolled. The NMA revealed that combined interventions (standardized mean difference [SMD] 0.89, 95% CI 0.22-1.57), telephone support (SMD 0.68, 95% CI 0.02-1.33), telemonitoring interventions (SMD 0.70, 95% CI 0.02-1.39), and mobile phone app interventions (SMD 0.65, 95% CI 0.01-1.30) were statistically superior to usual care. However, SMS compared to usual care showed no statistical difference. Notably, the combined intervention, with a surface under the cumulative ranking curve of 79.3%, appeared to be the most effective option for patients with CVDs. Regarding systolic blood pressure and diastolic blood pressure outcomes, the combined intervention also had the highest probability of being the best intervention. Conclusions: The research indicates that the combined intervention (SMS text messaging and telephone support) has the greatest likelihood of being the most effective eHealth intervention to improve medication adherence in patients with CVDs, followed by telemonitoring, telephone support, and app interventions. The results of these network meta-analyses can provide crucial evidence-based support for health care providers to enhance patients’ medication adherence. Given the differences in the design and implementation of eHealth interventions, further large-scale, well-designed multicenter trials are needed. Trial Registration: INPLASY 2023120063; https://inplasy.com/inplasy-2023-12-0063/

Background: Health-related rumors and misconceptions are spreading at an alarming rate, fueled by the rapid development of the internet and the exponential growth of social media platforms. This phenomenon has become a pressing global concern, as the dissemination of false information can have severe consequences, including widespread panic, social instability, and even public health crises. Objective: The aim of the study is to compare the accuracy of rumor identification and the effectiveness of health science popularization between 2 generated large language models in Chinese (GPT-4 by OpenAI and Enhanced Representation through Knowledge Integration Bot [ERNIE Bot] 4.0 by Baidu). Methods: In total, 20 health rumors and misconceptions, along with 10 health truths, were randomly inputted into GPT-4 and ERNIE Bot 4.0. We prompted them to determine whether the statements were rumors or misconceptions and provide explanations for their judgment. Further, we asked them to generate a health science popularization essay. We evaluated the outcomes in terms of accuracy, effectiveness, readability, and applicability. Accuracy was assessed by the rate of correctly identifying health-related rumors, misconceptions, and truths. Effectiveness was determined by the accuracy of the generated explanation, which was assessed collaboratively by 2 research team members with a PhD in nursing. Readability was calculated by the readability formula of Chinese health education materials. Applicability was evaluated by the Chinese Suitability Assessment of Materials. Results: GPT-4 and ERNIE Bot 4.0 correctly identified all health rumors and misconceptions (100% accuracy rate). For truths, the accuracy rate was 70% (7/10) and 100% (10/10), respectively. Both mostly provided widely recognized viewpoints without obvious errors. The average readability score for the health essays was 2.92 (SD 0.85) for GPT-4 and 3.02 (SD 0.84) for ERNIE Bot 4.0 (P=.65). For applicability, except for the content and cultural appropriateness category, significant differences were observed in the total score and scores in other dimensions between them (P<.05). Conclusions: ERNIE Bot 4.0 demonstrated similar accuracy to GPT-4 in identifying Chinese rumors. Both provided widely accepted views, despite some inaccuracies. These insights enhance understanding and correct misunderstandings. For health essays, educators can learn from readable language styles of GLLMs. Finally, ERNIE Bot 4.0 aligns with Chinese expression habits, making it a good choice for a better Chinese reading experience.

Epilepsy in children and adolescents harms cognitive development and quality of life, necessitating early risk identification to improve outcomes. Yet, current predictive models yielded inconsistent results, demanding a thorough evaluation of their accuracy and effectiveness to guide future research and inform evidence-based clinical strategies. This review aimed to integrate existing research findings on epilepsy prediction models for children and adolescents. China National Knowledge Infrastructure, Wanfang Database, SinoMed, China Science and Technology Journal Database, PubMed, Embase, CINAHL, and Web of Science were searched from inception to August 31, 2025. The Prediction Model Risk of Bias Assessment Tool was used to assess the risk of bias and applicability. The areas under the curve (AUC) with 95% confidence intervals were pooled using random-effects meta-analysis. The study was registered with PROSPERO (CRD42025637913). A total of 27 studies were included in this review. Sixteen studies were conducted in China. Twenty-five studies were at high risk of bias. The pooled AUC for 14 training models was 0.794 (95% CI: 0.747–0.840). For 17 validation models, the pooled AUC was 0.726 (95% CI: 0.659–0.792). Clinical features + EEG outperformed combinations with MRI in training (0.855 vs 0.725) and validation (0.743 vs 0.655). Non-machine learning models surpassed machine learning (training: 0.838 vs 0.717; validation: 0.778 vs 0.654), but the difference might not be statistically significant as the 95% CIs are overlapped in the validation; and external validation yielded higher AUC (0.807) than internal validation (0.634), though with extreme heterogeneity (I² = 90.93%). Current research showed uneven regional distribution. Models based on clinical features + EEG warrants further exploration. Predictor selection predominantly relies on univariate analysis, lacking standardized and scientific methodologies. Most studies carry a high risk of bias and rarely undergo validation, limiting their practical applicability. Validating existing models is crucial for identifying flaws and enhancing future research. 10.13039/501100004735 Natural Science Foundation of Hunan Province (grant No. 2024JJ8254).