Weikai Li

Background: Major depressive disorder (MDD) is a serious psychological disorder marked by persistent feelings of sadness and a notable decline in daily functioning. Functional Brain Networks (FBNs) have become crucial biomarkers for the diagnosis of MDD, and Graph Convolutional Networks (GCNs) have demonstrated great potential in extracting FBN features. However, existing GCN methods typically depend on large‐scale labeled datasets and lack clinical interpretability in MDD identification, which limits their application. Methods: We introduce Universum Inspired Graph Contrast Learning (Uni‐GCL), a novel graph‐based self‐supervised learning framework to relieve the demand for training sample size. Additionally, to enhance the clinical interpretability of GCNs, we optimized the Gradient‐weighted Class Activation Mapping (Grad‐CAM) method to enable it to effectively pinpoint the crucial nodes within the GCN. Results: The Uni‐GCL method outperformed previous state‐of‐the‐art approaches. It improved accuracy by 3% at Site 21 and 5% at Site 1. Additionally, Uni‐GCL identified that brain regions FGoperc. R, CAU. L, FFG. R, AMYG. L, PUT. L, AMYG. R, PUT. R, THA. L, IFGtriang. R, FFG. L, INS. L, INS. R, HIP. L and HIP. R make significant contributions to the diagnosis of MDD. Conclusions: Uni‐GCL can effectively reduce the sample size for model training and improve the diagnostic performance of MDD. Meanwhile, the adapted Grad‐CAM provides an effective tool to enhance the interpretability of GCN.

The advent of microscopic‐resolution spatial transcriptomics () enables the mapping of tissue complexity at subcellular resolution. However, the massive increase in data volume introduces a critical “memory wall,” significantly limiting the applicability of existing computational methods on deployable hardware platforms. In addition, traditional single‐scale modeling approaches often struggle to disentangle weak biological signals from technical noise. To overcome these challenges, STWave is proposed as a scalable framework for spatial domain identification. Specifically, STWave first enables efficient computation during both training and inference by leveraging a decoupled patch‐based learning strategy, thereby mitigating hardware limitations without sacrificing global contextual information. Furthermore, STWave adopts a discrete wavelet transform to encode gene expression features across multiple scales, effectively capturing both global trends and fine‐grained details. Extensive experimental results show that STWave attains state‐of‐the‐art clustering performance while maintaining exceptional computational efficiency, enabling scalable analysis of ultra‐large spatial transcriptomics datasets under constrained computational resources. Across platforms including Visium HD, Xenium, and CosMx, the proposed STWave accurately resolves complex tissue structures, enabling the identification of distinct immune niches within the tumor microenvironment and revealing fine‐grained developmental substructures. These results support STWave as a robust and efficient tool for large‐scale analysis.

Background: Currently, studies have shown that this procedure of transcatheter closure of ventricular septal defect can lead to a range of adverse complications, especially for the threatening heart conduct block. However, it had not been fully analysed that whether early alternations of electrocardiogram, including vectorcardiogram, could determine the long-term persistent heart conduct block. Methods: A retrospective study was conducted to develop a Random Forest-based prediction model for the postoperative of the transcatheter interventional closure. The study encompassed data preprocessing, model training, and assessment of feature importance. Result: We assessed the predictive value of parameters derived from defect anatomical features, interventional closure data, and ECG+VCG changes measured preoperatively, on postoperative days 1 and 3, and at 1 month after closure. The area under receiver operating characteristic curve values of such evaluations on the test sets were 0.9974, 0.9592, 0.9977, 0.9985, 0.9957, and 0.9977, recalls were 98.35%, 83.39%, 99.57%, 98.04%, 97.80%, and 99.58%. In the prediction model, the importance of features was ranked. We found that the diameter and number of occluder used in surgery and the size of left atrium and ventricle were the primary factors for predicting arrhythmia after surgery. For electrocardiogram data, significant parameters include RV5 voltage and the R–T angle of the sagittal plane vectorcardiogram before the treatment, T-wave, R-wave, and QRS duration in early days after operation and RV1 voltage, R-wave axis during 1-month follow-up. Conclusion: Electrocardiogram information during the perioperative period of interventional closure could be used to effectively predict the occurrence and prognosis of postoperative arrhythmia.

Background: Major depressive disorder (MDD) poses a significant challenge to global mental health, highlighting the urgent need for effective diagnostic tools to enable early and accurate detection. Resting‐state functional magnetic resonance imaging (rs‐fMRI) has garnered considerable attention for evaluating MDD through functional connectivity networks (FCNs). While graph convolutional networks (GCNs) have advanced FCN analysis by capturing complex interregional connection patterns, existing GCN‐based approaches have predominantly focused on a single connection pattern, thereby neglecting the multifaceted information encoded in diverse connection profiles. Methods: We propose a Multiple Functional Connection Patterns Graph Convolutional Network (MFCP) that integrates three distinct connection patterns—sparse representation, Pearson correlation, and Granger causality mapping—to leverage their complementary information. The proposed framework employs multiple graph convolutional modules to integrate diverse connectivity information, thereby enhancing MDD‐related diagnostic features extracted from FCNs. We conducted primary experiments on the REST‐MDD dataset Site 20, comprising 533 subjects, to evaluate the proposed MFCP. Results: The MFCP framework achieved superior diagnostic performance with an accuracy of 87.74%, precision of 86.21%, recall of 90.91%, F1‐score of 88.50%, and an area under the ROC curve (AUC) of 0.9326. Comparative analysis revealed that integrating three connection patterns significantly outperformed single‐pattern approaches (accuracy range: 72.64%–81.13%) and two‐pattern combinations (accuracy range: 77.36%–83.02%). t‐SNE visualization confirmed enhanced class separability with increasing pattern integration, while Grad‐CAM analysis identified distinct discriminative brain regions across different connectivity patterns. Conclusions: The proposed MFCP framework effectively integrates multiple functional connection patterns to enhance diagnostic feature extraction, demonstrating strong effectiveness for MDD identification. These findings suggest that leveraging complementary connectivity information through multi‐pattern integration represents a promising approach for improving automated MDD diagnosis based on rs‐fMRI functional connectivity analysis.

Objective: Kawasaki disease (KD) is characterized as an acute systemic vasculitis predominantly affecting young children, with coronary artery lesions (CALs) representing the most serious complication. Therapeutic resistance to intravenous immunoglobulin (IVIG) remains a significant clinical challenge. Consequently, numerous investigations have sought to identify predictive risk factors for IVIG resistance (IVIGR) and CAL development. Limited research has systematically evaluated the prognostic utility of electrocardiographic (ECG) parameters in KD outcome prediction. This study was therefore undertaken to assess the contributory value of ECG analysis in determining KD prognosis and therapeutic responses. Methods: This prospective cohort study enrolled 255 hospitalized children diagnosed with KD at West China Second University Hospital between July 2022 and December 2024. Initially, univariate analysis was performed to identify risk factors differentiating IVIGR from non-IVIGR patients and CAL from non-CAL patients. Statistically significant parameters were subsequently incorporated into machine learning analyses. Random forest algorithms were employed to construct predictive models based on the following: (1) complete blood count parameters, (2) biochemical and metabolic profiles, (3) electrocardiographic features, and (4) a comprehensive multimodal model integrating all parameters. These models generated feature importance scores, providing hierarchical rankings that quantified the relative contribution of each predictor to outcome prediction. Results: Univariate analysis demonstrated that alterations in hematological parameters, biochemical and metabolic profiles, and electrocardiographic features were significantly associated with therapeutic responses to IVIG and CAL development. Machine learning analysis revealed that ECG parameters individually contributed modest predictive weight for KD prognosis. However, the integration of ECG features into the comprehensive model substantially enhanced the discriminatory capacity, elevating the area under the curve (AUC) to 0.92 for CAL prediction. For IVIGR prediction, ECG-exclusive models demonstrated suboptimal performance in early disease management. Nevertheless, the multimodal integration of ECG with inflammatory and metabolic biomarkers achieved a comparable AUC of 0.92 for IVIGR prediction. Conclusions: This study establishes that ECG parameter alterations are significantly associated with CAL development and IVIGR in KD patients. Although ECG features demonstrate a limited independent predictive capacity compared to inflammatory and metabolic biomarkers, their integration into comprehensive predictive models substantially enhances discriminatory performance. These findings underscore the complementary value of electrocardiographic assessment in multimodal risk stratification strategies for KD management, supporting the clinical utility of ECG analysis as an adjunctive prognostic tool when combined with conventional laboratory parameters.

Accurate cell type annotation across datasets is a key challenge in single-cell analysis. snRNA-seq enables profiling of frozen or difficult-to-dissociate tissues, complementing scRNA-seq by capturing fragile or rare cell types. However, cross-annotation between these two datasets remains largely unexplored, as existing methods treat them independently. We introduce ScNucAdapt, a method designed for cross-annotation between paired and unpaired scRNA-seq and snRNA-seq datasets. To address distributional and cell composition differences, ScNucAdapt employs partial domain adaptation. Experiments across both unpaired and paired scRNA-seq and snRNA-seq show that ScNucAdapt achieves robust and accurate cell type annotation, outperforming existing approaches. Therefore, ScNucAdapt provides a practical framework for the cross-domain cell type annotation between scRNA-seq and snRNA-seq data.

Test-time few-shot object detection (FSOD) represents an innovative approach for identifying novel categories using a limited number of support examples, obviating the need for model fine-tuning. Despite advancements, existing FSOD methods, including our prior work, continue to grapple with challenges posed by domain/category shift and limited data availability. Building upon our previous research on test-time FSOD, this article proposes a novel dynamic prototype fusion network (PFN) to overcome these limitations. To mitigate the impact of the distribution shift, a dynamic prototype refinement method is introduced that updates prototypes from supporting images in an adaptive manner. Further, limited samples are mitigated through exhaustive exploitation of information within support images. Specifically, we design a dual-level multiscale information integration approach that effectively fuses information across different network layers and image scales, enhancing the model’s discriminating capabilities. Additionally, a mask-based preprocessing technique harnesses segmentation labels on support samples, effectively suppressing the adverse impact of background noise on model accuracy. Notably, to align with the constraints of test-time scenarios, model parameters remain fixed during the configuration step, with only prototypes being updated each time users input novel supporting samples. As a result, our method achieves superior performance over existing state-of-the-art FSOD methods on multiple benchmarks, demonstrating remarkable potential in the realm of FSOD. The code is available at https://github.com/CatfishW/TIDEV2