Volume 06 · Issue 04 · 2026

The deep integration of artificial intelligence (AI) in the healthcare sector is revolutionizing medical services, clinical decision-making, and academic research and education in medical institutions. However, its autonomous evolutionary characteristics based on continuous learning and iterative updates pose fundamental challenges to traditional, static regulatory frameworks of medical devices. The "Expert consensus on the application and governance of artificial intelligence in medical institutions (2026)," developed by over 40 leading Chinese medical and scientific research institutions, involved experts from fields such as medicine, hospital management, medical informatics, health policy, law, and medical ethics. Drawing on the latest Chinese and international regulations and best practices, the expert panel systematically compiled a comprehensive governance guide for AI applications across the entire lifecycle of medical institutions, focusing on six thematic pillars: Access evaluation, clinical application, patient rights protection, data governance, risk management, and competency enhancement. For each domain, the consensus details critical factors, including tiered access, multidisciplinary review, real-world validation, cross-validation of competing algorithms, delineation of human-machine collaboration liabilities, tiered informed consent, algorithmic traceability and explainability, dynamic risk monitoring, circuit-breaker mechanisms, and progressive AI competency training systems, that medical institutions should consider during AI implementation and management. This consensus aims to establish a compliance baseline and development direction for AI applications in medical institutions that is characterized by safety, efficacy, fairness, and interpretability, thereby ensuring that technological evolution remains within a legally and ethically sound framework. Ultimately, it serves to promote equitable access to high-quality medical resources and achieve substantial improvements in national health standards.

Surgery is a real-time, embodied and safety-critical form of care in which decisions, anatomy, physical action and patient outcomes are tightly coupled. Artificial intelligence (AI) has improved the recognition of instruments, anatomy, workflow, technical performance and multimodal surgical context, but these capabilities have limited clinical value if they remain confined to retrospective benchmarks, narrow procedural settings and isolated perception tasks. The central challenge is to determine whether surgical models can support safer operative care under uncertainty, across procedures, institutions, devices and patient contexts, without obscuring surgical responsibility. This is not a scaling problem alone, but a problem of representation, evidence and accountability. Generalist surgical foundation models may provide a shared basis for clinically useful surgical understanding, but surgery demands a more constrained formulation than general biomedical AI, one that is anchored in operative accountability, real-time decision-making and patient-level consequences. Clinical-grade surgical intelligence should accordingly be defined as a qualification standard, not a capability threshold, for models intended to assist surgeons. First, surgery should be treated as a generalization problem across sensory-action regimes, longitudinal perioperative episodes and heterogeneous clinical environments. Second, model development should progress from machine-readable perception to temporally anchored procedural understanding, evidence-grounded reasoning, operating-room multimodal intelligence and surgeon-supervised action support. Third, clinical-grade claims should be earned through evidence that extends beyond benchmark accuracy to external and cross-regime validity, real-time reliability under intraoperative uncertainty, human-facing utility, patient-level relevance and lifecycle governance. We therefore reframe generalist surgical foundation models not as autonomous substitutes or increasingly capable pretraining assets, but as components of a clinically accountable intelligence layer for safer operative care. The clinical value of these models will depend on whether broader representations, stronger evidence and explicit accountability mature together.

Critical care medicine is at a pivotal stage of transformation from symptom-based approaches toward systems science. Building on a first-principles understanding of the mechanisms underlying critical illness, we propose the entropic critical illness theory (ECIT). Grounded in the second law of thermodynamics, ECIT posits that critical illness arises when the body’s capacity to regulate entropy production is impaired, leading to systemic disorder. This process is manifested by the concurrent escalation of host response entropy and hemodynamic entropy, resulting in disruptions in blood flow and oxygen delivery, ischemia, and hypoxia at the level of the critical unit, which is defined as the terminal microcirculatory–mitochondrial functional unit, ultimately culminating in multi-organ dysfunction. This study delineates the theoretical foundations of ECIT and outlines an entropy-based critical care framework that may inform future multimodal entropy-informed monitoring, risk stratification, and AI-assisted precision intervention in critical care medicine.

Background Centrally located non-small cell lung cancer (NSCLC) traditionally requires sleeve resection or pneumonectomy. We investigated whether neoadjuvant chemoimmunotherapy (NeoCIT) could enable conversion to standard lobectomy and evaluated an artificial intelligence (AI)-assisted pure-vision and hyperspectral imaging (HSI) system for intraoperative margin assessment. Methods This multicenter cohort study enrolled patients from six tertiary referral centers in China between January 2019 and December 2025. The retrospective cohort (n=56, January 2019–December 2024) and prospective exploratory cohort (n=28, January–December 2025) included adults with histologically confirmed centrally located NSCLC, who were initially indicated for sleeve resection and received NeoCIT followed by surgical resection. Primary endpoints were event-free survival (EFS) and overall survival (OS), analyzed using Kaplan-Meier methods and multivariable Cox regression. The prospective cohort additionally underwent intraoperative bronchial margin assessment using an AI-assisted system combining HSI (600–950 nm and 16 spectral bands) with a multi-layer perceptron-based deep learning classifier for real-time pixel-wise tissue classification. Results In the retrospective cohort, 57.1% of patients achieved major pathologic response (MPR) or pathologic complete response (pCR) and 71.4% of them transitioned from sleeve lobectomy to lobectomy. The postoperative complication rate was significantly lower in the lobectomy group (10.0%) compared to the sleeve lobectomy group (38.5%, P=0.031). Two-year EFS was 69.9% for lobectomy versus 80.2% for sleeve resection (HR=0.76, P=0.665), with similar OS (84.0 vs. 90.9%, HR=0.68, P=0.713). Multivariable analysis identified pCR/MPR, age, Eastern Cooperative Oncology Group status, and pathological nodal stage after neoadjuvant therapy stage as independent prognostic factors. For the prospective cohort, the AI-assisted pure-vision and HSI system accurately assessed negative margins in real-time, showing high sensitivity (100%) and specificity (96.2%) compared to frozen-section pathology. Conclusion NeoCIT safely enables surgical conversion from sleeve resection to lobectomy in centrally located NSCLC without compromising oncologic outcomes, and AI-assisted HSI provides accurate real-time intraoperative bronchial margin assessment.

Background Spondyloarthritis (SpA) is a rare chronic inflammatory disease consisting of subtypes such as ankylosing spondylitis (AS), psoriatic arthritis (PsA), reactive arthritis (ReA), inflammatory bowel disease-associated arthritis (IBDA), and juvenile spondyloarthritis (JSpA). Early diagnosis and precise subtype identification are crucial for improving prognosis. However, in clinical practice, the overlapping clinical manifestations among subtypes, the lack of a single specific diagnostic biomarker, and the extremely limited clinical data for rare subtypes (e.g., IBDA, and JSpA) make early differential diagnosis challenging. Consequently, patients often face the risks of delayed diagnosis, overdiagnosis, and misclassification, thereby adversely affecting individualized treatment decisions and long-term outcomes. Currently, although general-purpose medical large language models (LLMs) have shown potential in certain domains, their performance remains inadequate in the vertical, disease-specific scenario of SpA subtyping, which demands high-quality data. Faced with scarcity and extreme imbalance across multiple categories, these models struggle to meet the practical clinical need for precise classification. Methods This study proposes a paradigm shift from passively "relying on data" to actively "creating data" . Based on electronic health records (EHRs) of SpA inpatients hospitalized at our institution between 2000 and 2019, we constructed a real-world dataset with five subtype groups (AS, PsA, ReA, IBDA, and JSpA). The dataset encompasses critical disease features such as sacroiliac joint imaging results, Human Leukocyte Antigen B27 (HLA-B27) status, and C-reactive protein levels. A human-machine comparative trial was conducted using the McNemar test. We achieved this through a three-stage framework: (1) Template extraction: High-quality real-world records were extracted from EHRs to establish a diagnostic benchmark for SpA subtypes. (2) Synthetic data generation: We developed a knowledge-guided synthetic data engine that deeply integrates the assessment of spondyloarthritis international society (ASAS) classification criteria and expert priors into a LLM, enabling the targeted generation of clinical case pairs covering rare subtypes with logical consistency. (3) Adaptive optimization: An adaptive weighted sampling strategy was introduced, which dynamically adjusts the training data distribution based on the initial model’s accuracy across categories. This optimizes the model’s learning on weak links, ultimately yielding the final selected model (RobotGPT-SpA). Results The experiments demonstrated that the proposed method significantly enhances the performance of SpA subtype classification. RobotGPT-SpA achieved an overall accuracy of 0.843, exhibiting balanced and superior diagnostic capabilities across all five subtypes. Notably, it improved performance by 30% on the data-scarce IBDA subtype and by 27% on ReA. Compared with general-purpose medical LLMs with larger parameter counts, our domain-specific model demonstrated higher clinical utility and precision, thereby validating the superiority of the "domain-specific fine-tuning + high-quality data" strategy. In the differential diagnostic comparative experiments, the RobotGPT-SpA model significantly outperformed junior and midlevel physicians, demonstrating higher accuracy, consistency, as well as better sensitivity and specificity. Conclusion This study confirms that knowledge-guided data engineering is the key to building high-performance, deployable artificial intelligence (AI) models for specialized diseases. Through systematic data generation and balancing strategies, this approach effectively overcomes the application bottlenecks of medical AI in rare and complex chronic diseases. It provides a reliable technical pathway and a universal methodology for achieving early and precise SpA subtype diagnosis in resource-constrained scenarios.

Background Telemedicine has emerged as a critical modality in healthcare delivery, especially for elderly patients with complex medication regimens. Pharmacists play a central role in optimizing pharmacotherapy, and the integration of artificial intelligence tools, particularly large language models (LLMs), may enhance telepharmacy services by providing timely, accurate, and empathetic responses. Despite growing theoretical applications, limited evidence exists on how LLMs perform in pharmacy-relevant teleconsultations. However, little is known about whether LLMs can reliably address patient-specific pharmacotherapy concerns during teleconsultations. This study specifically investigated how different LLMs perform when applied to simulated, pharmacy-relevant scenarios in elderly care. Methods This observational, cross-sectional study was conducted between May and June 2025. Ten simulated patient case scenarios representing common pharmacotherapeutic concerns in elderly populations were used as the dataset. Three LLMs (ChatGPT-4.0, DeepSeek-V3, and Gemini 2.5 Flash) were tested. Each LLM was prompted with identical patient scenarios, and responses were independently evaluated by two authors. A structured rubric assessed eight primary parameters: clarity, tone, empathy, patient-centeredness, accuracy, actionability, risk mitigation, and safety accuracy. Scores (0-3 per domain, max 21) with a negative score for safety accuracy and errors were estimated. Readability assessments were carried out using Flesch-Kincaid (FK) grade level and Flesch reading ease (FRE) scores, along with public reach (%). Descriptive statistics and Kruskal-Wallis H test with Bonferroni corrections were used for comparative analysis. Results All three LLMs successfully generated responses for all 10 case scenarios. Their outputs demonstrated high levels of empathy, patient-specific guidance, and clinical appropriateness and excelled in emotional tone and structural clarity. Overall, Gemini scores (20.0 ± 0.7) were statistically significantly higher compared to ChatGPT (18.6 ± 1.1; P=0.002) and DeepSeek (19.0 ± 0.8; P=0.031). Regarding the reading assessment, FK grade levels were significantly different with Gemini (8.9 ± 0.6) compared with DeepSeek (7.5 ± 1.1; P=0.002), as was public reach ((79.3 ± 3.7) vs. (84.0 ± 1.9); P=0.001). All their responses were graded excellent except DeepSeek for case 1 and ChatGPT for case 7, where their responses were good. DeepSeek stood out for its tiered, instructional formatting and inclusion of safety tips such as pill organizers and warning signs, while Gemini delivered detailed pharmacologic explanations and emphasized urgent care needs when appropriate. Variability was noted in verbosity and response structuring, with Gemini being more narrative and ChatGPT being more conversational. All LLMs displayed accurate drug-related knowledge and patient-centered communication aligned with best practices in telepharmacy. Conclusions LLMs consistently generated accurate, empathetic, and actionable responses to simulated telepharmacy scenarios. Their use may serve as a supportive tool in delivering high-quality virtual pharmacy consultations. Rigorous real-world validation and comparative trials against pharmacist-delivered care are required before LLMs can be responsibly integrated into clinical workflows.

Background Neoadjuvant chemoradiotherapy (nCRT) has become the standard preoperative treatment for patients with locally advanced rectal cancer (LARC). Accurately predicting patient prognosis can help formulate individualized treatment strategies. We aimed to establish and validate a clinical radiomics model based on magnetic resonance imaging and clinical characteristics to predict the prognosis of patients with LARC who underwent radical surgery after nCRT. Methods This retrospective study included 234 patients with LARC who underwent radical surgery after nCRT at the Affiliated Hospital of Qingdao University between December 2019 and October 2023. The patients were randomly divided into training (164 patients) and testing (70 patients) sets. Imaging and clinical data were collected and analyzed. Overall, 1,172 radiomic features were extracted from preoperative pelvic magnetic resonance T2-weighted imaging (MRI T2WI). After screening, the radiomic feature score (Rad-Score) of the tumor and mesorectal regions was calculated, and three radiomics models were developed. A clinical model was constructed using the identified clinical predictors. The Rad-Score and clinical predictors were integrated using the random forest algorithm to establish a clinical radiomics model. The accuracy of the clinical radiomic model was evaluated using the area under the curve (AUC), calibration, and decision curves analysis. Results A total of 8 tumor radiomic features and 12 mesenteric radiomic features that were highly correlated with patient prognosis from pelvic MRI T2WI of patients after nCRT. We then established a hybrid radiomics model using all the radiomic features. The AUCs of the hybrid radiomics model for the training and testing sets were 0.832 (95% CI: 0.762–0.902) and 0.764 (95% CI: 0.646–0.881), respectively. The clinical radiomics model combining 20 radiomic features and 4 clinical predictors achieved an AUC of 0.874 (95% CI: 0.811–0.936) in the training set and 0.813 (95% CI: 0.710–0.916) in the testing set. Conclusion The clinical radiomics model based on MRI developed in this study has a high predictive performance for the prognosis of patients with LARC who underwent radical surgery after nCRT, thereby helping clinicians perform nCRT in patients with rectal cancer.

Background Mass gatherings pose significant risks of mass casualty incidents, where static triage protocols often fail to balance individual patient urgency with global emergency resource constraints, leading to preventable mortality. We aimed to evaluate a dynamic diversion strategy optimized using machine learning to improve survival outcomes and emergency efficiency in such scenarios. Methods We conducted a retrospective simulation study based on real-world emergency data from two major events in Hangzhou, China: The 2024 Hangzhou Marathon and Linping Ice and Snow World fire incident. Data encompassing patient severity such as respiratory rate, pulse rate, and motor response (RPM) scores, ambulance logistics, and real-time hospital bed capacity from multiple trauma centers were used to construct the simulation environment. We compared a dynamic diversion framework (using a value decomposition network-based multiagent approach) against standard clinical protocols (nearest-hospital and designated-hospital strategies). Primary outcomes included overall survival rate, average waiting time for critical care, and total length of hospital stays. Results In simulated mass casualty incident scenarios with rapid casualty influx and limited resources, our dynamic strategy significantly outperformed conventional methods. Compared to two traditional baselines, our approach boosts average patient survival rates by 6.8% and 8.3% during casualty surges, and by 13.8% and 8.8% during gradual influxes. Under resource-sufficient conditions, the respective gains are 12.2% and 8.7%. These results indicate that our method can effectively integrate and use dynamically changing emergency medical resources to achieve efficient triage and transportation decisions when dealing with mass casualties. Conclusion Our diversion decision-making method can effectively use the large amount of dynamic emergency medical information generated during mass gatherings. It enables diversion decisions based on patient injuries and emergency medical resources, thereby providing emergency diversion doctors with more personalized transfer decision-making schemes.

Treatment of locally advanced adenocarcinoma of the rectum has evolved rapidly in recent decades. Multimodality therapy, which combines neoadjuvant chemoradiotherapy (CRT), total mesorectal excision (TME), and adjuvant chemotherapy, demonstrates improved outcomes compared to historical approaches. Although surgery remains a core component, low anterior resection and abdominoperineal resection may have a severe impact on the quality of life, including bowel, urinary, and sexual dysfunction. Recently, patients with no evidence of tumor after completion of treatment may be managed using a nonsurgical, watch-and-wait (WW) approach. This paradigm involves extensive surveillance using clinical, endoscopic, and radiologic methods. Artificial intelligence (AI) is increasingly being applied at multiple stages of rectal cancer care to improve treatment precision and decision-making. AI can integrate clinical data to predict treatment response, improve early detection of disease, and optimize therapy. The WW approach can better identify ideal candidates, facilitate surveillance, and avoid overtreatment or undertreatment. This review explores key concepts in the WW approach and the role of AI in its management. AI can support the individualization of cancer care by assessing treatment response, and predicting outcomes to facilitate shared decision-making. These strategies could help expand utilization of the WW approach while increasing organ preservation.