[摘 要] 目的：探讨复发/转移性鼻咽癌（NPC）接受含PD-1单抗免疫治疗的临床特征和预后影响因素。方法：回顾性分析2019年3月至2024年7月期间南部战区总医院确诊的95例NPC患者的临床资料和外周血生化及免疫学指标。预后分析采用Kaplan-Meier曲线，组间比较使用Log-rank检验，采用Cox比例风险模型进行单因素和多因素分析。结果：95例患者中男性81例，女性14例，中位年龄49.72岁（16~74岁），Ⅳ期91例（95.79%），所有患者均采用免疫治疗，联合或不联合化疗方案治疗，中位无进展生存期（mPFS）为10.5个月，客观缓解率（ORR）70.53%，疾病控制率（DCR）89.47%，接受含铂治疗方案患者PFS相对更长，且差异有统计学意义。紫杉醇 + 顺铂 + 氟尿嘧啶（TPF）对比吉西他滨 + 顺铂（GP）和紫杉醇 + 顺铂（TP）显示出更长的PFS，但差异无统计学意义。不同PD-1单抗治疗组间的PFS未显示出有统计学意义的差异。单因素及多因素Cox回归分析结果显示，肿瘤复发状态、初始血浆EBV感染状态、治疗周期数、基线外周血SII是复发/转移性NPC患者接受PD-1抑制剂治疗疗效预测的独立相关因素（均P < 0.05），并且非复发患者、初始血浆EBV DNA阳性、接受 ≥ 4治疗周期、基线外周血SII < 772.81的患者接受PD-1抑制剂治疗预后相对更好。结论：在接受PD-1抑制剂治疗的复发/转移性NPC患者中，非复发患者、初始血浆EBV DNA阳性、≥ 4治疗周期且外周血SII < 772.81者PFS相对更长，可早期识别免疫治疗效果不佳患者并精准干预。

Background As one of the most severe occupational diseases in China, pneumoconiosis is significantly burdened by its complications, which adversely affects patients' quality of life. Objective To identify the influencing factors of complications in pneumoconiosis and to construct an early prediction model for pneumoconiosis complications, providing theoretical guidance for clinical diagnosis, treatment, and rehabilitation. Methods A case-control study was conducted using data from the Chongqing 5G Pneumoconiosis Rehabilitation Management Information Platform. A total of 1872 pneumoconiosis patients with complications, who received rehabilitation at various rehabilitation stations in Chongqing from January 2021 to December 2024, were enrolled as the case group. Concurrently, 2348 patients without complications from the same platform and period were included as the control group. Univariate analysis, LASSO regression, and logistic regression were employed to identify influencing factors for complications. An extreme gradient boosting (XGBoost) model was trained using ten-fold cross-validation to predict pneumoconiosis complications. The SHapley Additive exPlanations (SHAP) method was applied for model visualization. Results The analysis of influencing factors revealed that increased age, silicosis type, advanced stage of pneumoconiosis, dependence in activities of daily living, abnormal muscle strength, increased dyspnea index, presence of cough, viscous sputum, impaired cardiac function, severe abnormality in the 6-minute walk test (6MWT), and increased Borg score were risk factors for complications (P<0.05). The XGBoost model demonstrated an area under the curve (AUC) of receiver operating characteristic of 0.718, and a Brier score of 0.211 for the calibration curve. Decision curve analysis demonstrated a superior net benefit within the threshold probability range of 0.2 to 1.0 compared to "all-patient intervention" and "no-patient intervention" strategies, indicating good predictive performance and significant clinical utility. SHAP visualization identified the top six important features as Borg score, age, 6MWT, muscle strength, pneumoconiosis stage, and cough symptoms. Conclusion Multiple factors influence the occurrence of complications in pneumoconiosis patients. Clinical attention should be focused on elderly patients with decreased cardiopulmonary reserve, pronounced respiratory symptoms, and diminished muscle strength. The XGBoost model demonstrates satisfactory discrimination, calibration, and clinical applicability in predicting pneumoconiosis complications and may serve as a useful reference for clinical decision-making.

Renal Fanconi syndrome (FS) is a rare renal manifestation of primary Sjögren's syndrome (pSS). This study aims to analyze the clinical and prognostic characteristics of patients with pSS-associated renal FS (pSS-FS) and provide insights for clinical management.

Bronchiolar adenoma (BA) is a rare type of lung tumor characterized by bilayered epithelial cells having a continuous basal layer and a luminal layer. It resembles mucinous adenocarcinoma (MA) on frozen section, with difficulty in distinguishing the basal layer. Immunohistochemistry is the best choice for verifying the diagnosis. This study aimed to comprehensively characterize three cases of BA-combined carcinoma using clinical, histopathological, and genetic features. BA and carcinoma sections were subjected to next-generation sequencing, respectively. It was hypothesized that while different mutation forms matched different regions, BA and lung adenocarcinoma shared the same gene mutation when they co-occurred in the same location. BA with extensive carcinoma is extremely rare and presents diagnostic challenges due to its overlap with conditions such as MA. Because of its distinctive morphological characteristics, BA may be regarded as a low-grade malignancy, particularly during a confusing evaluation. A multifaceted examination of clinical, radiological, immunohistochemical, and genetic data is necessary for an accurate diagnosis.

Intelligent prescription recommendation has become an important research direction in traditional Chinese medicine (TCM), offering new opportunities to support clinical decision-making and promote the modernization of TCM practice. With the rapid development of artificial intelligence (AI), a variety of computational approaches have been proposed to learn prescription patterns from clinical data and generate personalized treatment recommendations. However, despite increasing research activity, systematic and comprehensive reviews of AI-driven methods for TCM prescription recommendation remain limited. In this study, we present a comprehensive review of computational approaches for herbal prescription recommendation (HPR) in TCM. Existing methods are systematically categorized into several major paradigms, including traditional machine learning methods, topic model methods, sequential generative methods, deep learning and graph-based methods, and large language model–based frameworks. In addition to summarizing methodological developments, we also review commonly used public datasets and evaluation metrics in this field. Furthermore, representative models with publicly available implementations are experimentally evaluated on multiple benchmark datasets to provide a comparative analysis of their performance on the HPR task. Finally, we discuss the key challenges that hinder the practical deployment of intelligent prescription recommendation systems, including data heterogeneity, limited interpretability, and insufficient integration of TCM domain knowledge. Future research directions are outlined to facilitate the development of more reliable, interpretable, and clinically applicable AI-assisted HPR systems for TCM.

Objective To investigate the clinical safety and efficacy of predictive extracorporeal membrane oxygenation (ECMO) support during liver transplantation in high-risk recipients. Methods A single-center retrospective study was conducted. A total of 35 liver transplant recipients who underwent predictive ECMO assistance during surgery at the Second Affiliated Hospital of Guangxi Medical University from January 2021 to April 2025 were included. Baseline characteristics, ECMO application, ECMO operating parameters and hemodynamic conditions during ECMO operation were analyzed. The cumulative survival rate was calculated using the Kaplan-Meier method and survival curves were plotted. Results Veno-arterial ECMO in 35 recipients was established after induction of general anesthesia. The median ECMO operation time was 8.5 h, and the success rate of weaning from ECMO was 97%. The cardiac index (CI) during the non-liver period and the new liver period was higher than that before cannulation, and CI increased during the non-liver period and before weaning (all P<0.001). The mean invasive arterial pressure and central venous pressure remained stable. No ECMO-related complications occurred in the recipients. The average survival time was (25 ± 3) months. The main causes of death were severe pulmonary infection and multiple organ failure. Conclusions For liver transplant recipients with severe cardiopulmonary dysfunction, hemodynamic disorders, or multiple organ failure risk factors, implementing predictive ECMO support during liver transplantation through active risk prediction is a safe and effective circulatory guarantee strategy.

ObjectiveThe frequency difference limen (FDL) serves as a fundamental metric utilized for effectively quantifying the precise perceptual capabilities of the central auditory system. However, traditional measurement methods rely heavily on the active behavioral responses of subjects and are consequently highly susceptible to the negative influence of confounding subjective factors. Furthermore, existing research paradigms frequently employ uniform stimulus configurations that overlook critical individual perceptual differences. Based on brain-computer interface (BCI) technology, this comprehensive study aims to establish an objective and quantitative evaluation method for auditory frequency discrimination by systematically analyzing and decoding the specific neural responses elicited at the exact threshold state. MethodsWe designed a personalized rapid serial auditory presentation (RSAP) paradigm customized based on each individual’s precise FDL. A cohort of eleven healthy participants was recruited to evaluate the paradigm using pure-tone sequences at a baseline frequency of 4 000 Hz. This experimental paradigm simulates a realistic auditory perception environment through the continuous presentation of acoustic stimuli, thereby allowing for an in-depth investigation into the specific neural representations evoked by weak frequency deviations at the threshold state. Given that auditory stimulus-evoked response features exhibit complex and differentiated spatiotemporal distribution patterns across multiple frequency domains, this study further deeply integrates the cross-scale feature interaction module with the dynamic spatiotemporal attention allocation strategy, innovatively proposing the Multi-Scale Spatial-Temporal Dual Attention Network (MS-STAMNet). Specifically, the network constructs parallel processing branches with multiple receptive fields and introduces a dynamic adaptive weighting strategy to precisely localize core neural activity signals, further deeply integrating multi-scale information through cross-branch feature information interaction to achieve robust single-trial decoding of weak auditory evoked responses. ResultsThe comprehensive electrophysiological data analysis demonstrated that subtle auditory frequency deviation stimuli presented at the threshold level successfully elicited pronounced N2 and P3 event-related potential features, reflecting pre-attentive mismatch detection and subsequent cognitive evaluation, which were prominently distributed over the frontal, central, and temporal regions of the scalp. In the complex time-frequency domain, the extracted neural response characteristics exhibited distinct, statistically significant event-related synchronization within both the low-frequency δ and θ frequency bands, which was simultaneously accompanied by a widespread, prominent event-related desynchronization within the higher α band. A comparative analysis of model performance demonstrated that MS-STAMNet achieved an average unweighted average recall (UAR) of (69.67±6.12)% and area under the curve (AUC) of 0.761 8±0.07, significantly outperforming the established baseline models such as EEGNet and PLNet. Furthermore, a distinct dissociation phenomenon was verified between neural decoding and behavioral performance through regression analysis (R²=0.016, P=0.709), indicating that this model can effectively capture the implicit features of subtle frequency deviations, even when they fail to trigger explicit conscious responses. Additionally, attention weight visualization analysis further reveals the highly accurate focus of the network on key features concentrated over the bilateral temporal and fronto-parietal regions. ConclusionThis study systematically and comprehensively uncovers the multi-dimensional spatiotemporal evolutionary patterns of complex neural responses processing subtle acoustic variations under long-sequence threshold auditory stimulation. Concurrently, it verifies the efficacy and robustness of the proposed MS-STAMNet architecture in accurately deciphering weak, single-trial electroencephalogram signals amidst complex background noise. Ultimately, these neurophysiological and algorithmic findings lay a solid theoretical and methodological foundation for the objective and quantitative evaluation of individual auditory cognitive capabilities in clinical applications, transcending the fundamental limitations of traditional behavioral paradigms and providing robust technical support for future auditory research and related clinical assessments.

Electroencephalography (EEG) is a non-invasive neurophysiological monitoring technique. It records the electrical activity of the cerebral cortex using electrodes placed on the scalp surface. Owing to its high safety, portability, and millisecond-level temporal resolution, EEG has been widely utilized in a variety of fields, including clinical diagnosis, brain-computer interfaces (BCIs), and cognitive neuroscience research. However, due to its microvolt-level amplitude, EEG is highly susceptible to various artifacts, including electrooculographic (EOG), electrocardiographic (ECG), electromyographic (EMG), and power line interference (PLI). These artifacts can obscure genuine neural activity and introduce spurious electrophysiological features. Consequently, they may compromise EEG signal quality, thereby reducing the reliability of downstream analyses. To address this issue, numerous EEG artifact removal methods have been developed, including both traditional denoising techniques and deep learning-based approaches. Traditional EEG denoising methods have long served as the primary solutions for artifact removal. Representative approaches include filtering, regression, and blind source separation. Although these methods have demonstrated effectiveness in specific scenarios, they suffer from several inherent limitations. Filtering assumes that artifacts and EEG signals can be separated in the frequency domain, but many artifacts, such as EOG and EMG, overlap with EEG spectra, which may lead to the loss of valuable neural information. Regression methods require high-quality artifact references to estimate and subtract contaminations, limiting their effectiveness in reference-free scenarios. Blind source separation can remove artifacts without external references, but it typically requires the number of EEG channels to exceed the number of sources, restricting its application in single- or low-channel EEG recordings. Deep learning-based EEG denoising methods address these limitations effectively. First, they learn the nonlinear mapping between contaminated and clean EEG directly from data in an end-to-end manner. This approach does not rely on assumptions about spectral separability, thereby preserving neural activity more completely. Second, the reference information is incorporated during the training phase, allowing the trained model to perform artifact removal independently without external references. Third, deep learning models can be flexibly designed to accommodate various recording setups, achieving robust denoising for both high-density and single-channel EEG. Collectively, these advantages enable deep learning-based methods to overcome the main challenges of traditional approaches, providing more accurate and reliable EEG signal recovery. The superior denoising performance of deep learning-based EEG denoising methods has attracted increasing attention in EEG artifact removal research. As a result, many deep learning-based denoising methods have been developed and successfully applied in neural engineering areas. However, a systematic review of the techniques and applications in this field is still lacking. To address this gap, this paper reviews recent advances in deep learning-based EEG denoising from four perspectives: technical principle, benchmark dataset, denoising model, and evaluation method. Representative applications in neural signal analysis and BCI decoding are also summarized. Furthermore, the advantage, existing challenge, and future research direction of deep learning-based EEG denoising are discussed. This review aims to provide valuable theoretical insights and technical guidance for researchers. It is also expected to promote further advances and broader applications of deep learning-based EEG denoising techniques.

Electroencephalogram (EEG) is a non-invasive, high temporal-resolution technique for monitoring brain activity. However, affected by the volume conduction effect, EEG has a low spatial resolution and is difficult to locate brain neuronal activity precisely. The surface Laplacian (SL) technique obtains the Laplacian EEG (LEEG) by estimating the second-order spatial derivative of the scalp potential. LEEG can reflect the radial current activity under the scalp, with positive values indicating current flow from the brain to the scalp (“source”) and negative values indicating current flow from the scalp to the brain (“sink”). It attenuates signals from volume conduction, effectively improving the spatial resolution of EEG, and is expected to contribute to breakthroughs in neural engineering. This paper provides a systematic overview of the principles and development of SL technology. Currently, there are two implementation paths for SL technology: current source density algorithms (CSD) and concentric ring electrodes (CRE). CSD performs the Laplace transform of the EEG signals acquired by conventional disc electrodes to indirectly estimate the LEEG. It can be mainly classified into local methods, global methods, and realistic Laplacian methods. The global method is the most commonly used approach in CSD, which can achieve more accurate estimation compared with the local method, and it does not require additional imaging equipment compared with the realistic Laplacian method. CRE employs new concentric ring electrodes instead of the traditional disc electrodes, and measures the LEEG directly by differential acquisition of the multi-ring signals. Depending on the structure, it can be divided into bipolar CRE, quasi-bipolar CRE, tripolar CRE, and multi-pole CRE. The tripolar CRE is widely used due to its optimal detection performance. While ensuring the quality of signal acquisition, the complexity of its preamplifier is relatively acceptable. Here, this paper introduces the study of the SL technique in resting rhythms, visual-related potentials, movement-related potentials, and sensorimotor rhythms. These studies demonstrate that SL technology can improve signal quality and enhance signal characteristics, confirming its potential applications in neuroscientific research, disease diagnosis, visual pathway detection, and brain-computer interfaces. CSD is frequently utilized in applications such as neuroscientific research and disease detection, where high-precision estimation of LEEG is required. And CRE tends to be used in brain-computer interfaces, that have stringent requirements for real-time data processing. Finally, this paper summarizes the strengths and weaknesses of SL technology and envisages its future development. SL technology boasts advantages such as reference independence, high spatial resolution, high temporal resolution, enhanced source connectivity analysis, and noise suppression. However, it also has shortcomings that can be further improved. Theoretically, simulation experiments should be conducted to investigate the theoretical characteristics of SL technology. For CSD methods, the algorithm needs to be optimized to improve the precision of LEEG estimation, reduce dependence on the number of channels, and decrease computational complexity and time consumption. For CRE methods, the electrodes need to be designed with appropriate structures and sizes, and the low-noise, high common-mode rejection ratio preamplifier should be developed. We hope that this paper can promote the in-depth research and wide application of SL technology.

Objective To assess the accuracy of pN staging prediction in papillary thyroid carcinoma (PTC) using ultrasound radiomics and deep neural networks (DNN). Methods A retrospective analysis was conducted on 375 patients with pathologically confirmed PTC, comprising 261 cases in the training set and 114 in the test set. Staging was categorized as pN0 (no cervical lymph node metastasis), pN1a (central neck lymph node metastasis), and pN1b (lateral neck lymph node metastasis). An ultrasound physician manually segmented the regions of interest (ROIs) for PTC, extracting 1899 radiomic features. Dimensionality reduction was performed using the least absolute shrinkage and selection operator (LASSO) regression. A DNN model for predicting PTC pN staging was developed using the H2O deep learning platform, trained on the training set, and validated on the test set to assess the accuracy of the optimal model. Results A total of 153 patients were in the pN0 stage, 131 patients in the pN1a stage, and 91 patients in the pN1b stage. LASSO regression selected 15 radiomic features for each PTC. The optimal DNN model, constructed using these 15 features, achieved accuracies of 85.82% on the training set and 81.57% on the test set. Conclusion Ultrasound radiomics of PTC demonstrates high accuracy in predicting pN staging and shows potential for automating N staging in patients.