Objective To investigate the scientific outputs of metagenomic next-generation sequencing (mNGS) for diagnosis of infection of unknown origin in intensive care units (ICUs), and to decipher the latest advances, frontier trends and spatiotemporal evolution of research hotpots in mNGS for diagnosis of infection of unknown origin in ICUs. Methods Publications pertaining to the application of mNGS in diagnosis of infection of unknown origin in ICUs were retrieved from Web of Science Core Collection (WOSCC) from January 1, 2015 to December 31, 2024. The software Scimago Graphica 1.0.30 was employed to generate the network maps of collaboration relationships between countries, international collaborative relationships, author collaborations, institutional collaborative relationships, and a heatmap of journals, and the software VOSviewer 1.6.18 was used to create a heatmap of keywords, and maps of keyword co-occurrence clustering and keyword clustering timelines. In addition, the keyword burst map was created using the software CiteSpace 6.3.R3. Results A total of 1 707 publications were included in the final analysis, and the number of publications appeared an overall tendency towards a rise from 2015 to 2024, with the largest number of publications seen in 2024 (545 publications). The largest number of publications was recorded in China (1 390 publications), followed by in USA (190 publications) and United Kingdom (31 publications), and China led the global research in this field, with 81% of global related researches linked with China. Frontiers in Cellular and Infection and Microbiology published the largest number of articles (212 publications, 12.42%), and Joseph Derisi was the most productive author (33 publications). Author collaborations occurred within groups; however, there was a lack of close inter-group collaborations, with University of California, San Francisco and Chan Zuckerberg Biohub-based group seen as the largest collaborative group. High-frequency co-occurrence keywords included mNGS, infection, diagnosis, case report, community-acquired pneumonia and bronchoalveolar lavage fluid, and the 100 most common high-frequency co-occurrence keywords were assigned into four clusters. Keyword clustering timeline analysis revealed that the research hotspots in this field shifted from virus sequencing and sequence alignment to severe pulmonary infections, and keyword burst analysis showed identification, mNGS and virus as top three keywords with the highest burst intensity. Conclusions mNGS was mainly used for identification of viruses among patients with infections of unknown origins in ICUs from 2015 to 2024, and future research priority shifted to pathogen detection for severe pulmonary infections.

Background In 2021, chronic obstructive pulmonary disease (COPD) emerged as the forth leading cause of death in the world. However, the impact of air pollutants on COPD is still inconsistent across current studies. Objective To analyze the relationship between ambient sulfur dioxide (SO₂) exposure and hospital admissions for COPD in Lanzhou, and to examine the modified effects of SO₂ across different genders, age groups, and seasons. Methods A total of 20718 hospital admission records for COPD were collected from Lanzhou between 2016 and 2020, alongside synchronized data on SO₂ concentration and meteorological variables. A generalized additive model integrated with a distributed lag nonlinear model was used to assess the relationship and the lag effects between SO₂ exposure and COPD admissions. Potential confounders, including meteorological factors, day-of-the-week effect, and holidays, were controlled for, with a maximum lag period set at 7 d. Two-pollutant models and sensitivity analyses were conducted to ensure robustness. Results are expressed as relative risks (RR) with 95% confidence intervals (95%CI). Results During the study period, the average daily number of COPD admissions was 11.34±7.03. The average mass concentration of SO₂ was (23.72±12.35) μg·m⁻³. SO₂ exposure was positively correlated with COPD admissions; specifically, each 10 μg·m⁻³ increase in SO₂ concentration was associated with an RR of 1.440 (95%CI: 1.317, 1.573). Stratified analyses found that at a cumulative lag of 7 d, the RRs were higher among females, individuals aged ≥65 years, and during the cold season. Conversely, no significant increase in COPD admission risk related to SO₂ was observed during the warm season. For every 10 μg·m⁻³ increase in SO₂ concentration, the RR was 1.483 (95%CI: 1.311, 1.678) for females, 1.441 (95%CI: 1.311, 1.583) for those aged > 65 years， and 1.595 (95%CI: 1.313, 1.937) during the cold season. Conclusion Short-term exposure to ambient SO₂ exposure in Lanzhou is associated with an increased risk of COPD admission. The association is more pronounced among females, the elderly (aged> 65 years) and during the cold season.

ObjectiveTo analyze the application trends and research hotspots of action observation therapy (AOT) in the field of stroke rehabilitation over the past decade. MethodsLiteratures on AOT in stroke rehabilitation published from January, 2016 to December, 2025 were retrieved from the Web of Science Core Collection database. CiteSpace 6.4.R1 was used for visual analysis. ResultsA total of 463 articles were included. The annual publication volume showed a fluctuating upward trend. The country with the highest number of publications was China, the most productive institution was Chang Gung University and Consiglio Nazionale delle Ricerche, and the most prolific author was Avanzini Pietro. Mirror neuron system, motor imagery, upper limb and facilitation were identified as high-frequency keywords and bursting words. ConclusionIn the past decade, the number of publications on AOT in stroke rehabilitation has generally increased. Researches are focusing on the synergy of sensory-closed-loop multimodal technologies, reconstruction of fine upper limb function and neural facilitation mechanisms.

ObjectiveTo develop a machine learning (ML)-based prediction model for neurological rehabilitation outcomes of stroke patients. MethodsA total of 420 stroke patients admitted to the Fuzhou First People's Hospital from October, 2022 to October, 2024 were enrolled as the training set. According to the modified Rankin Scale (mRS) scores three months after discharge, the patients were divided into prognosis group (n = 289) and poor prognosis group (n = 131). An additional 180 stroke patients hospitalized in the same hospital from November, 2024 to April, 2025 were selected as the validation set. Univariate analysis, least absolute shrinkage and selection operator regression, and multivariate logistic regression were performed to identify independent influencing factors for the prognosis of neurological function recovery. Using the screened independent influencing factors as feature variables, six ML models were established, including logistic regression, linear discriminant analysis, naive Bayes, support vector machine, random forest and extreme gradient boosting (XGBoost). The area under the receiver operating characteristic curve (AUC), confusion matrix indicators (accuracy, precision, recall and F1-score), calibration curve and decision curve analysis were adopted to evaluate the predictive efficacy, calibration degree and clinical net benefit of each model, with external validation conducted in the validation set. The SHapley Additive exPlanations framework was used to interpret the optimal model, and bar charts were applied to visualize the feature importance of the best model. ResultsAge, National Institutes of Health Stroke Scale (NIHSS) score, collateral circulation grading, fasting plasma glucose (FPG), lymphocyte percentage (LYMPH%), and homocysteine (Hcy) were independent risk factors for poor neurological rehabilitation prognosis (P < 0.05). For the XGBoost model, the AUC of the training and validation sets were 0.963 (95%CI 0.947 to 0.979) and 0.825 (95%CI 0.764 to 0.885), respectively, while the accuracy was 88.81% and 77.22%, the precision was 92.86% and 68.42%, the recall was 69.47% and 47.27%, and the F1-score was 79.48% and 55.91%, optimal in both calibration and clinical net benefit. The feature importance ranking for the XGBoost model from high to low was NIHSS score, age, collateral circulation grading, FPG, Hcy and LYMPH%. ConclusionThe interpretable XGBoost ML model exhibits excellent predictive efficacy and favorable clinical applicability in predicting neurological rehabilitation outcomes for stroke patients.

Organoid-on-a-chip technology represents a promising interdisciplinary advancement that merges two cutting-edge biomedical platforms: stem cell-derived organoids and microfluidics-based organ-on-a-chip systems. Organoids are self-organizing three-dimensional (3D) cell cultures that mimic the key structural and functional features of in vivo organs. However, traditional organoid culture systems are often static, lacking dynamic environmental cues and suffering from limitations such as batch-to-batch variability, low stability, and low throughput. Organ-on-a-chip platforms, by contrast, utilize microfluidic technologies to simulate the dynamic physiological microenvironment of human tissues and organs, enabling more controlled cell growth and differentiation. By integrating the advantages of organoids and organ-on-a-chip technologies, organoid-on-a-chip systems transcend the limitations of conventional 3D culture models, offering a more physiologically relevant and controllable in vitro platform. In organoid-on-a-chip systems, stem cells or pre-formed organoids are cultured in micro-engineered environments that mimic in vivo conditions, enabling precise control over fluid flow, mechanical forces, and biochemical cues. Specifically, these platforms employ advanced strategies including bio-inspired 3D scaffolds for structural support, precise spatial cell patterning via 3D bioprinting, and integrated biosensors for real-time monitoring of metabolic activities. These synergistic elements recreate complex extracellular matrix signals and ensure high structural fidelity. Based on structural complexity, organoid-on-a-chip systems are classified into single-organoid and multi-organoid types, forming a trajectory from unit biomimicry to systemic simulation. Single-organoid chips focus on highly biomimetic units by integrating vascular, immune, or neural functions. Multi-organoid chips simulate inter-organ crosstalk and systemic homeostasis, advancing complex disease modeling and PK/PD evaluation. This emerging technology has demonstrated broad application potential in multiple fields of biomedicine. Organoid-on-a-chip systems can recapitulate organ developmentin vitro, facilitating research in developmental biology. They mimic organ-specific physiological activities and mechanisms, showing promising applications in regenerative medicine for tissue repair or replacement. In disease modeling, they support the reconstruction of models for neurodegenerative, inflammatory, infectious, metabolic diseases, and cancers. These platforms also enable in vitro drug testing and pharmacokinetic studies (ADME). Patient-derived chips preserve genetic and pathological features, offering potential for precision medicine. Additionally, they reduce species differences in toxicology, providing human-relevant data for environmental, food, cosmetic, and drug safety assessments. Despite progress, organoid-on-a-chip systems face challenges in dynamic simulation, extracellular matrix (ECM) variability, and limited real-time 3D imaging, requiring improved materials and the integration of developmental signals. Current bottlenecks also include the high technical threshold for automation and the lack of standardized validation frameworks for regulatory adoption. Meanwhile, the concept of a “human-on-a-chip” has been proposed to mimic whole-body physiology by integrating multiple organoid modules. This approach enables systemic modeling of drug responses and toxicity, with the potential to reduce animal testing and revolutionize drug development. Future advancements in bio-responsive hydrogels and flexible biosensors will further empower these platforms to bridge the gap between bench-side research and personalized clinical interventions. In conclusion, organoid-on-a-chip technology offers a transformative in vitro model that closely recapitulates the complexity of human tissues and organ systems. It provides an unprecedented platform for advancing biomedical research, clinical translation, and pharmaceutical innovation. Continued development in biomaterials, microengineering, and analytical technologies will be essential to unlocking the full potential of this powerful tool.

ObjectiveThe accuracy of Y-chromosome haplogroup assignment is crucial for tracing paternal lineage in male samples. With the advancement of high-throughput sequencing technologies, high-density Y-SNP genotyping from whole-genome or array-based data has become a standard method for determiningY-chromosome haplogroups. This study systematically evaluated the performance of 4 commonly used high-density SNP genotyping systems—namely, the Global Screening Array (GSA), Chinese Genotyping Array (CGA), Affymetrix array, and the 1240K capture panel—for haplogroup assignment. This work provides a reference for data comparison across different systems. MethodsWe extracted genotype data for the 4 Y-SNP panels from 30× whole-genome sequencing (WGS) data of 1 590 male samples from the 1000 Genomes Project. Additionally, GSA array genotype data from 384 relative pairs (spanning 1st- to 12th-degree relationships) from 109 Chinese Han families were collected. Haplogroup assignment was performed using Y-LineageTracker v1.3.0 software. We assessed the concordance and resolution of haplogroup assignments between the four Y-SNP panels and the WGS data. The consistency and resolution of haplogroup assignments were also evaluated for both the 1000 Genomes Project samples and the 109 family samples collected in this study. Furthermore, the impact of varying numbers of Y-SNPs on haplogroup assignment was examined. ResultsThe GSA and CGA panels demonstrated superior resolution and discrimination of haplogroup subclades compared with the other two panels. The haplogroup assignments from the GSA, CGA, and 1240K panels showed high concordance with WGS data, with consistency rates exceeding 88.70%, whereas the Affymetrix platform exhibited a significantly lower consistency rate of 61.89%. Specifically, the GSA and CGA panels consistently demonstrated superior performance compared with the other two panels in the assignment of haplogroups O-M175 and H-L901, achieving complete concordance (100%) for both haplogroups. In contrast, the Affymetrix panel erroneously assigned all individuals belonging to haplogroup O-M175 to haplogroup K2-M526. Furthermore, its accuracy for haplogroup H-L901 was exceedingly low, at merely 1.41%. This poor performance was characterized by the misassignment of 98.59% of H-L901 samples—specifically, 1.41% to J-M304 and a predominant 97.18% to F-M89. For haplogroup R-M207, all four panels exhibited uniformly high levels of consistency, with concordance values exceeding 94.00%. Notably, for haplogroup E-M96, the 1240K and Affymetrix panels outperformed the GSA and CGA panels in terms of concordance, representing the first instance in which these two panels surpassed the latter. Conversely, for haplogroups J-M304, Q-M242, and I-M170, all 4 panels showed relatively elevated misclassification rates, with the Affymetrix array demonstrating the poorest overall performance. None of the four panels showed any discordant haplogroup assignments among the familial relative pairs analyzed. A positive correlation was observed between the number of Y-SNPs (ranging from 1 000 to 10 000) and classification consistency; however, classification consistency plateaued when the number of Y-SNPs exceeded 10 000. Furthermore, a random sampling analysis conducted on the GSA and CGA panels demonstrated that the haplogroup misclassification rate exhibited negligible fluctuation across the Y-SNP range of 500 to 1 000. Conversely, a marked enhancement in classification consistency was observed as the number of markers increased from 1 000 to 5 000, ultimately reaching a plateau within the interval of 5 000 to 8 000 markers. ConclusionThese findings indicate that the GSA and CGA panels provide high resolution and concordance, delivering reliable Y-haplogroup assignment for forensic investigations.

Organoid-on-a-chip technology represents a promising interdisciplinary advancement that merges two cutting-edge biomedical platforms: stem cell-derived organoids and microfluidics-based organ-on-a-chip systems. Organoids are self-organizing three-dimensional (3D) cell cultures that mimic the key structural and functional features of in vivo organs. However, traditional organoid culture systems are often static, lacking dynamic environmental cues and suffering from limitations such as batch-to-batch variability, low stability, and low throughput. Organ-on-a-chip platforms, by contrast, utilize microfluidic technologies to simulate the dynamic physiological microenvironment of human tissues and organs, enabling more controlled cell growth and differentiation. By integrating the advantages of organoids and organ-on-a-chip technologies, organoid-on-a-chip systems transcend the limitations of conventional 3D culture models, offering a more physiologically relevant and controllable in vitro platform. In organoid-on-a-chip systems, stem cells or pre-formed organoids are cultured in micro-engineered environments that mimic in vivo conditions, enabling precise control over fluid flow, mechanical forces, and biochemical cues. Specifically, these platforms employ advanced strategies including bio-inspired 3D scaffolds for structural support, precise spatial cell patterning via 3D bioprinting, and integrated biosensors for real-time monitoring of metabolic activities. These synergistic elements recreate complex extracellular matrix signals and ensure high structural fidelity. Based on structural complexity, organoid-on-a-chip systems are classified into single-organoid and multi-organoid types, forming a trajectory from unit biomimicry to systemic simulation. Single-organoid chips focus on highly biomimetic units by integrating vascular, immune, or neural functions. Multi-organoid chips simulate inter-organ crosstalk and systemic homeostasis, advancing complex disease modeling and PK/PD evaluation. This emerging technology has demonstrated broad application potential in multiple fields of biomedicine. Organoid-on-a-chip systems can recapitulate organ developmentin vitro, facilitating research in developmental biology. They mimic organ-specific physiological activities and mechanisms, showing promising applications in regenerative medicine for tissue repair or replacement. In disease modeling, they support the reconstruction of models for neurodegenerative, inflammatory, infectious, metabolic diseases, and cancers. These platforms also enable in vitro drug testing and pharmacokinetic studies (ADME). Patient-derived chips preserve genetic and pathological features, offering potential for precision medicine. Additionally, they reduce species differences in toxicology, providing human-relevant data for environmental, food, cosmetic, and drug safety assessments. Despite progress, organoid-on-a-chip systems face challenges in dynamic simulation, extracellular matrix (ECM) variability, and limited real-time 3D imaging, requiring improved materials and the integration of developmental signals. Current bottlenecks also include the high technical threshold for automation and the lack of standardized validation frameworks for regulatory adoption. Meanwhile, the concept of a “human-on-a-chip” has been proposed to mimic whole-body physiology by integrating multiple organoid modules. This approach enables systemic modeling of drug responses and toxicity, with the potential to reduce animal testing and revolutionize drug development. Future advancements in bio-responsive hydrogels and flexible biosensors will further empower these platforms to bridge the gap between bench-side research and personalized clinical interventions. In conclusion, organoid-on-a-chip technology offers a transformative in vitro model that closely recapitulates the complexity of human tissues and organ systems. It provides an unprecedented platform for advancing biomedical research, clinical translation, and pharmaceutical innovation. Continued development in biomaterials, microengineering, and analytical technologies will be essential to unlocking the full potential of this powerful tool.

ObjectiveThe accuracy of Y-chromosome haplogroup assignment is crucial for tracing paternal lineage in male samples. With the advancement of high-throughput sequencing technologies, high-density Y-SNP genotyping from whole-genome or array-based data has become a standard method for determiningY-chromosome haplogroups. This study systematically evaluated the performance of 4 commonly used high-density SNP genotyping systems—namely, the Global Screening Array (GSA), Chinese Genotyping Array (CGA), Affymetrix array, and the 1240K capture panel—for haplogroup assignment. This work provides a reference for data comparison across different systems. MethodsWe extracted genotype data for the 4 Y-SNP panels from 30× whole-genome sequencing (WGS) data of 1 590 male samples from the 1000 Genomes Project. Additionally, GSA array genotype data from 384 relative pairs (spanning 1st- to 12th-degree relationships) from 109 Chinese Han families were collected. Haplogroup assignment was performed using Y-LineageTracker v1.3.0 software. We assessed the concordance and resolution of haplogroup assignments between the four Y-SNP panels and the WGS data. The consistency and resolution of haplogroup assignments were also evaluated for both the 1000 Genomes Project samples and the 109 family samples collected in this study. Furthermore, the impact of varying numbers of Y-SNPs on haplogroup assignment was examined. ResultsThe GSA and CGA panels demonstrated superior resolution and discrimination of haplogroup subclades compared with the other two panels. The haplogroup assignments from the GSA, CGA, and 1240K panels showed high concordance with WGS data, with consistency rates exceeding 88.70%, whereas the Affymetrix platform exhibited a significantly lower consistency rate of 61.89%. Specifically, the GSA and CGA panels consistently demonstrated superior performance compared with the other two panels in the assignment of haplogroups O-M175 and H-L901, achieving complete concordance (100%) for both haplogroups. In contrast, the Affymetrix panel erroneously assigned all individuals belonging to haplogroup O-M175 to haplogroup K2-M526. Furthermore, its accuracy for haplogroup H-L901 was exceedingly low, at merely 1.41%. This poor performance was characterized by the misassignment of 98.59% of H-L901 samples—specifically, 1.41% to J-M304 and a predominant 97.18% to F-M89. For haplogroup R-M207, all four panels exhibited uniformly high levels of consistency, with concordance values exceeding 94.00%. Notably, for haplogroup E-M96, the 1240K and Affymetrix panels outperformed the GSA and CGA panels in terms of concordance, representing the first instance in which these two panels surpassed the latter. Conversely, for haplogroups J-M304, Q-M242, and I-M170, all 4 panels showed relatively elevated misclassification rates, with the Affymetrix array demonstrating the poorest overall performance. None of the four panels showed any discordant haplogroup assignments among the familial relative pairs analyzed. A positive correlation was observed between the number of Y-SNPs (ranging from 1 000 to 10 000) and classification consistency; however, classification consistency plateaued when the number of Y-SNPs exceeded 10 000. Furthermore, a random sampling analysis conducted on the GSA and CGA panels demonstrated that the haplogroup misclassification rate exhibited negligible fluctuation across the Y-SNP range of 500 to 1 000. Conversely, a marked enhancement in classification consistency was observed as the number of markers increased from 1 000 to 5 000, ultimately reaching a plateau within the interval of 5 000 to 8 000 markers. ConclusionThese findings indicate that the GSA and CGA panels provide high resolution and concordance, delivering reliable Y-haplogroup assignment for forensic investigations.

Extracellular vesicles (EVs) are pivotal mediators of intercellular communication within the tumor immune microenvironment (TME). They are broadly categorized into exosomes, microvesicles, and apoptotic bodies based on their distinct biogenesis pathways. Exosomes originate from the endosomal system via multivesicular body fusion, microvesicles bud directly from the plasma membrane, and apoptotic bodies are released during programmed cell death. By shuttling diverse bioactive cargoes—including proteins, lipids, and nucleic acids such as mRNA, miRNA, and DNA—EVs exert dual modulatory effects on tumor initiation, progression, and immune evasion. Importantly, EVs exhibit remarkable compositional heterogeneity that is intrinsically linked to their cellular origin. Tumor-derived EVs (TDEVs) are typically enriched with immunosuppressive molecules like PD-L1, TGF‑β, and miR-21, which promote tumor immune escape and metastasis. In contrast, EVs derived from immune cells, such as dendritic cells or cytotoxic T lymphocytes, often carry immunostimulatory components including antigens, co-stimulatory molecules, and granzymes, thereby potentiating anti-tumor immunity. This review systematically delineates the biogenesis and molecular composition of EVs, with a particular emphasis on their dynamic regulatory functions within the TME. Specifically, we discuss how EVs mediate intricate crosstalk between immune and tumor cells, facilitating signal transfer that reshapes immune surveillance. For instance, TDEVs can induce macrophage polarization toward an M2-like pro-tumor phenotype, while also suppressing natural killer cell cytotoxicity and dendritic cell maturation. The clinical utility of EV-associated biomarkers in liquid biopsy is increasingly recognized. Circulating EVs carry tumor-specific molecular signatures that mirror the genetic and proteomic alterations of primary tumors, enabling non-invasive early diagnosis, molecular subtyping, and real-time monitoring of therapeutic responses. Their natural biocompatibility, low immunogenicity, and intrinsic ability to traverse biological barriers make them ideal candidates for drug delivery systems. This review explores cutting-edge applications, including the use of EVs in immune checkpoint blockade therapy—for instance, engineered EVs displaying anti-PD-1 antibodies or carrying siRNA to silence immunosuppressive genes. Moreover, EV-based tumor vaccines are being developed, leveraging dendritic cell-derived EVs loaded with tumor antigens to elicit potent T cell responses. The feasibility of loading EVs with therapeutic molecules such as chemotherapeutic agents, oncolytic viruses, or CRISPR-Cas9 components is also under active investigation. The advent of engineered EVs has further expanded their therapeutic potential. Through surface modification or cargo encapsulation, EVs can be tailored for targeted delivery and controlled release, enhancing precision immunotherapy. However, several hurdles impede clinical translation. Current isolation and purification methods, such as ultracentrifugation and size-exclusion chromatography, suffer from low yield and purity. Distinguishing EV subpopulations remains technically challenging due to overlapping size and marker expression. Moreover, the lack of standardized protocols for EV production, characterization, and quality control poses significant barriers to regulatory approval and clinical adoption. Looking forward, the convergence of multi-omics technologies with artificial intelligence offers a powerful approach to decipher EV heterogeneity and identify robust diagnostic signatures. Machine learning algorithms can integrate proteomic, transcriptomic, and lipidomic data from large patient cohorts to construct predictive models for cancer diagnosis and prognosis. Concurrently, advances in bioengineering are enabling the design of next-generation EVs with enhanced targeting specificity, on-demand drug release, and reduced off-target effects. Future efforts should also focus on establishing good manufacturing practice (GMP)‑compliant production processes and conducting rigorous preclinical and clinical evaluations. In summary, this review provides a comprehensive overview of EV biology, their multifaceted roles in the TME, and their transformative potential in cancer diagnostics and therapeutics. By addressing current challenges and leveraging emerging technologies, EV-based strategies are poised to revolutionize precision oncology.

ObjectiveGastric hemorrhage is one of the most common and life-threatening emergencies of the upper digestive tract. Early identification and continuous monitoring are essential for reducing rebleeding rates and mortality, particularly within the critical early hours after onset. Although endoscopy and radiological imaging can accurately localize bleeding sites, these approaches are invasive, resource-intensive, and unsuitable for continuous bedside monitoring. Electrical impedance tomography (EIT), as a noninvasive and radiation-free functional imaging technique, offers real-time visualization of conductivity distribution and has the potential for detecting intragastric bleeding based on the electrical contrast between blood and surrounding gastric tissues. In this study, a three-dimensional gastric EIT (3D-gEIT) framework is proposed to achieve noninvasive, real-time, and dynamic monitoring of gastric hemorrhage, with emphasis on spatial localization and quantitative volume assessment. MethodsA three-dimensional upper-abdominal simulation model incorporating the stomach, gastric wall, gastric contents, and surrounding tissues was established. Three electrode configurations, namely the dual layer ring, the four layer staggered ring, and the opposed dual plane array, were designed and systematically compared to evaluate their influence on depth sensitivity and spatial resolution. Based on the Tikhonov-Noser hybrid regularization scheme, a region-clustering constraint was introduced to develop the TK-Noser-RCC algorithm. This approach aggregates spatially adjacent elements with similar conductivity variations, thereby enhancing structural continuity and suppressing isolated noise artifacts. To validate the proposed framework, an upper-abdominal physical phantom was constructed using agar to simulate background tissue conductivity. Hemispherical high-conductivity inclusions with volumes ranging from 10 ml to 50 ml were attached to the inner gastric wall to mimic localized bleeding under different gastric filling states. Boundary voltages were acquired under a 120 kHz excitation current and reconstructed using the TK-Noser-RCC algorithm. Furthermore, an in vivo animal experiment was performed using a porcine model with adult-scale abdominal dimensions. A total of 100 ml of autologous blood was injected incrementally into the stomach to simulate progressive gastric hemorrhage, and time-difference EIT reconstruction was conducted at each injection stage to assess the dynamic system response under physiological conditions. ResultsSimulation results demonstrated that the opposed dual-plane electrode array achieved superior depth sensitivity distribution and spatial resolution. For a 40 ml hemorrhage model, the average ICC and SSIM improved by 55.9% and 38.8% compared with the dual-layer ring configuration, and by 64.0% and 39.5% compared with the four-layer staggered configuration. The proposed region-clustering constraint significantly enhanced reconstruction stability. Under added Gaussian noise of 40 dB and 30 dB, ICC values remained approximately 0.85, indicating effective artifact suppression and preservation of boundary integrity. In physical phantom experiments, reconstructed hemorrhage volumes increased approximately linearly with the preset hemispherical volumes, and the reconstructed high-conductivity regions closely matched the actual bleeding locations. Both empty-stomach and full-stomach conditions were evaluated, demonstrating that the opposed dual-plane configuration maintained stable imaging performance across varying gastric contents. In the animal experiment, reconstructed low-impedance regions expanded progressively with increasing injected blood volume. The spatial localization of the hemorrhage remained stable throughout the procedure, and no significant artifacts were observed. Quantitative analysis showed that reconstructed volume and average conductivity variation exhibited an approximately linear growth trend with injected blood volume, confirming the sensitivity of the system to dynamic intragastric conductivity changes. ConclusionThe proposed 3D-gEIT framework enables quantitative reconstruction of gastric hemorrhage volume and spatial distribution with improved depth sensitivity, structural continuity, and noise robustness compared with conventional EIT approaches. By integrating optimized electrode configuration and a region-clustering-constrained reconstruction algorithm, the system provides stable dynamic monitoring under both controlled phantom conditions and in vivo physiological environments. This method offers a noninvasive, real-time, and low-cost imaging strategy for early diagnosis, postoperative monitoring, and bedside surveillance of gastric bleeding.