Learning-based Model Research Articles

Fault Diagnosis of Semi-Supervised Electromechanical Transmission Systems Under Imbalanced Unlabeled Sample Class Information Screening

In the health monitoring of electromechanical transmission systems, the collected state data typically consist of only a minimal amount of labeled data, with a vast majority remaining unlabeled. Consequently, deep learning-based diagnostic models encounter the challenge of scarcity in labeled data and abundance in unlabeled data. Traditional semi-supervised deep learning methods based on pseudo-label self-training, while alleviating the issue of labeled data scarcity to some extent, neglect the reliability of pseudo-label information, the accuracy of feature extraction from unlabeled data, and the imbalance in sample selection. To address these issues, this paper proposes a novel semi-supervised fault diagnosis method under imbalanced unlabeled sample class information screening. Firstly, an information screening mechanism for unlabeled data based on active learning is established. This mechanism discriminates based on the variability of intrinsic feature information in fault samples, accurately screening out unlabeled samples located near decision boundaries that are difficult to separate clearly. Then, combining the maximum membership degree of these unlabeled data in the classification space of the supervised model and interacting with the active learning expert system, label information is assigned to the screened unlabeled data. Secondly, a cost-sensitive function driven by data imbalance is constructed to address the class imbalance problem in unlabeled sample screening, adaptively adjusting the weights of different class samples during model training to guide the training of the supervised model. Ultimately, through dynamic optimization of the supervised model and the feature extraction capability of unlabeled samples, the recognition ability of the diagnostic model for unlabeled samples is significantly enhanced. Validation through two datasets, encompassing a total of 12 experimental scenarios, demonstrates that in scenarios with only a small amount of labeled data, the proposed method achieves a diagnostic accuracy increment exceeding 10% compared to existing typical methods, fully validating the effectiveness and superiority of the proposed method in practical applications.

Development and validation of a deep learning-enhanced prediction model for the likelihood of pulmonary embolism

BackgroundPulmonary embolism (PE) is a common and potentially fatal condition. Timely and accurate risk assessment in patients with acute deep vein thrombosis (DVT) is crucial. This study aims to develop a deep learning-based, precise, and efficient PE risk prediction model (PE-Mind) to overcome the limitations of current clinical tools and provide a more targeted risk evaluation solution.MethodsWe analyzed clinical data from patients by first simplifying and organizing the collected features. From these, 37 key clinical features were selected based on their importance. These features were categorized and analyzed to identify potential relationships. Our prediction model uses a convolutional neural network (CNN), enhanced with three custom-designed modules for better performance. To validate its effectiveness, we compared this model with five commonly used prediction models.ResultsPE-Mind demonstrated the highest accuracy and reliability, achieving 0.7826 accuracy and an area under the receiver operating characteristic curve of 0.8641 on the prospective test set, surpassing other models. Based on this, we have also developed a Web server, PulmoRiskAI, for real-time clinician operation.ConclusionThe PE-Mind model improves prediction accuracy and reliability for assessing PE risk in acute DVT patients. Its convolutional architecture and residual modules substantially enhance predictive performance.

Performance of Two Deep Learning-based AI Models for Breast Cancer Detection and Localization on Screening Mammograms from BreastScreen Norway.

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Purpose To evaluate cancer detection and marker placement accuracy of two artificial intelligence (AI) models developed for interpretation of screening mammograms. Materials and Methods This retrospective study included data from 129 434 screening examinations (all female, mean age 59.2, SD = 5.8) performed between January 2008 and December 2018 in BreastScreen Norway. Model A was commercially available and B was an in-house model. Area under the receiver operating characteristic curve (AUC) with 95% confidence interval (CIs) were calculated. The study defined 3.2% and 11.1% of the examinations with the highest AI scores as positive, threshold 1 and 2, respectively. A radiologic review assessed location of AI markings and classified interval cancers as true or false negative. Results The AUC was 0.93 (95% CI: 0.92-0.94) for model A and B when including screen-detected and interval cancers. Model A identified 82.5% (611/741) of the screen-detected cancers at threshold 1 and 92.4% (685/741) at threshold 2. For model B, the numbers were 81.8% (606/741) and 93.7% (694/741), respectively. The AI markings were correctly localized for all screen-detected cancers identified by both models and 82% (56/68) of the interval cancers for model A and 79% (54/68) for B. At the review, 21.6% (45/208) of the interval cancers were identified at the preceding screening by either or both models, correctly localized and classified as false negative (n = 17) or with minimal signs of malignancy (n = 28). Conclusion Both AI models showed promising performance for cancer detection on screening mammograms. The AI markings corresponded well to the true cancer locations. ©RSNA, 2025.

Data-knowledge hybrid driven intelligent prediction method of tunnel excavation profiles geometric deformation

ABSTRACT Applying digital twin technology to the tunnel excavation process enables intelligent tunnel construction, which is critical for ensuring construction safety and advancing intelligent building methods. However, the complex construction environment, involving numerous environmental factors and dynamic changes, poses challenges for traditional deformation prediction and analysis methods. These methods fail to systematically integrate the elements of the excavation scene to achieve accurate prediction. Therefore, this paper proposes a data-knowledge hybrid driven intelligent prediction method of tunnel excavation profile geometric deformation. A dynamic data-driven tunnel excavation knowledge extraction method is designed, along with the construction of a multi-process coupled tunnel excavation knowledge base. Additionally, a deep learning-based excavation profile deformation prediction model is developed, integrating the data-knowledge hybrid driven method to improve prediction performance. The experimental results demonstrate that the method proposed in this paper can significantly integrate and analyse the deformation elements of the tunnel excavation scene. It achieves intelligent prediction of the tunnel excavation profile geometric deformation, with reductions of 57.46%, 76.00%, and 53.12% in MAE, MSE, and RMSE, respectively, compared with the traditional prediction model. This method effectively enhances the deformation prediction accuracy of the tunnel excavation profile twin model.

Reinforcement Learning-Based Event-Triggered Model Predictive Control for Electric Vehicle Active Battery Cell Balancing

Abstract To extend the operation window of batteries, active cell balancing has been studied in the literature. However, such an advancement presents significant computational challenges on real-time optimal control, especially when the number of cells in a battery increases. This article investigates the use of reinforcement learning (RL) and model predictive control (MPC) to effectively balance battery cells while at the same time keeping the computational load at a minimum. Specifically, event-triggered MPC is introduced as a way to reduce real-time computation. Different from the existing literature where rule-based or threshold-based event-trigger policies are used to determine the event instances, deep RL is explored to learn and optimize the event-trigger policy. Simulation results demonstrate that the proposed framework can keep the cell state-of-charge variation under 1% while using less than 1% computational resources compared to conventional MPC.

Impact of contrast enhancement phase on CT-based radiomics analysis for predicting post-surgical recurrence in renal cell carcinoma.

To investigate the effect of CT enhancement phase on radiomics features for predicting post-surgical recurrence of clear cell renal cell carcinoma (ccRCC). This retrospective study included 144 patients who underwent radical or partial nephrectomy for ccRCC. Preoperative multiphase abdominal CT scans (non-contrast, corticomedullary, and nephrographic phases) were obtained for each patient. Automated segmentation of renal masses was performed using the nnU-Net framework. Radiomics signatures (RS) were developed for each phase using ensembles of machine learning-based models (Random Survival Forests [RSF], Survival Support Vector Machines [S-SVM], and Extreme Gradient Boosting [XGBoost]) with and without feature selection. Feature selection was performed using Affinity Propagation Clustering. The primary endpoint was disease-free survival, assessed by concordance index (C-index). The study included 144 patients. Radical and partial nephrectomies were performed in 81% and 19% of patients, respectively, with 81% of tumors classified as high grade. Disease recurrence occurred in 74 patients (51%). A total of 1,316 radiomics features were extracted per phase per patient. Without feature selection, C-index values for RSF, S-SVM, XGBoost, and Penalized Cox models ranged from 0.43 to 0.61 across phases. With Affinity Propagation feature selection, C-index values improved to 0.51-0.74, with the corticomedullary phase achieving the highest performance (C-index up to 0.74). The results of our study indicate that radiomics analysis of corticomedullary phase contrast-enhanced CT images may provide valuable predictive insight into recurrence risk for non-metastatic ccRCC following surgical resection. However, the lack of external validation is a limitation, and further studies are needed to confirm these findings in independent cohorts.

Deep Learning For Detection of Foliar Diseases in Soybeans Based on the Mask R-CNN Model

Objective: The objective of this study is to apply the Mask R-CNN model to detect and classify foliar diseases in soybean (Glycine max L.), assisting in the early diagnosis of powdery mildew, Asian soybean rust, and target spot. Theoretical Framework: It was based on concepts of deep learning, machine learning, and instance segmentation, using ResNet-50 and ResNet-101 architectures, given the relevance of soybean and the impacts caused by foliar diseases. Method: An initial dataset of 525 images was used, which was expanded to 2,408 through data augmentation techniques (resizing and rotation). The model integrated region proposal networks (RPN) and region of interest alignment (RoIAlign) to improve the segmentation and classification of images with foliar diseases. For model evaluation, the metrics mAP, IoU, and F1-Score were used. Results and Discussion: The results revealed that the ResNet-101 architecture outperformed ResNet-50, achieving 85% accuracy in detecting foliar diseases in soybean. This outcome reinforces the applicability of artificial intelligence in the early diagnosis of foliar diseases in agriculture, particularly in soybean cultivation. Research Implications: The research highlights the feasibility of deep learning-based models for agricultural management. Future studies may explore methods that eliminate the need for polygonal annotations and integrate technologies such as drones and IoT devices for greater scalability and efficiency. Originality/Value: The research highlights the feasibility of deep learning-based models for agricultural management. Future studies may explore methods that eliminate the need for polygonal annotations and integrate technologies such as drones and IoT devices for greater scalability and efficiency.

Image recognition technology for bituminous concrete reservoir panel cracks based on deep learning.

Detecting cracks in asphalt concrete slabs is challenging due to environmental factors like lighting changes, surface reflections, and weather conditions, which affect image quality and crack detection accuracy. This study introduces a novel deep learning-based anomaly model for effective crack detection. A large dataset of panel images was collected and processed using denoising, standardization, and data augmentation techniques, with crack areas labeled via LabelImg software. The core model is an improved Xception network, enhanced with an adaptive activation function, dynamic attention mechanism, and multi-level residual connections. These innovations optimize feature extraction, enhance feature weighting, and improve information transmission, significantly boosting accuracy and robustness. The improved model achieves a 97.6% accuracy and a Matthews correlation coefficient of 0.98, remaining stable under varying lighting conditions. This method not only provides a fresh approach to crack detection but also greatly enhances detection efficiency.

The Dipeptidyl Peptidase-4 Inhibitor Saxagliptin as a Candidate Treatment for Disorders of Consciousness: A Deep Learning and Retrospective Clinical Analysis.

Despite advancements in the neuroscience of consciousness, no new medications for disorders of consciousness (DOC) have been discovered in more than a decade. Repurposing existing US Food and Drug Administration (FDA)-approved drugs for DOC is crucial for improving clinical management and patient outcomes. To identify potential new treatments among existing FDA-approved drugs, we used a deep learning-based drug screening model to predict the efficacy of drugs as awakening agents based on their three-dimensional molecular structure. A retrospective cohort study from March 2012 to October 2024 tested the model's predictions, focusing on changes in Glasgow Coma Scale (GCS) scores in 4047 patients in a coma from traumatic, vascular, or anoxic brain injury. Our deep learning drug screens identified saxagliptin, a dipeptidyl peptidase-4 inhibitor, as a promising awakening drug for both acute and prolonged DOC. The retrospective clinical analysis showed that saxagliptin was associated with the highest recovery rate from acute coma among diabetes medications. After matching patients by age, sex, initial GCS score, coma etiology, and glycemic status, brain-injured patients with diabetes on incretin-based therapies, including dipeptidyl peptidase-4 inhibitors and glucagon-like peptide-1 analogues, recovered from coma at significantly higher rates compared to both brain-injured patients with diabetes on non-incretin-based diabetes medications (95% confidence interval of 1.8-14.1% higher recovery rate, P = 0.0331) and brain-injured patients without diabetes (95% confidence interval of 2-21% higher recovery rate, P = 0.0272). Post matching, brain-injured patients with diabetes on incretin-based therapies also recovered at a significantly higher rate than patients treated with amantadine (95% confidence interval for the difference 2.4-25.1.0%, P = 0.0364). A review of preclinical studies identified several pathways through which saxagliptin and other incretin-based medications may aid awakening from both acute and chronic DOC: restoring monoaminergic and GABAergic neurotransmission, reducing brain inflammation and oxidative damage, clearing hyperphosphorylated tau and amyloid-β, normalizing thalamocortical glucose metabolism, increasing neural plasticity, and mitigating excitotoxic brain damage. Our findings suggest incretin-based medications in general, and saxagliptin in particular, as potential novel therapeutic agents for DOC. Further prospective clinical trials are needed to confirm their efficacy and safety in DOC.

Deep learning and explainable AI for classification of potato leaf diseases

The accurate classification of potato leaf diseases plays a pivotal role in ensuring the health and productivity of crops. This study presents a unified approach for addressing this challenge by leveraging the power of Explainable AI (XAI) and transfer learning within a deep Learning framework. In this research, we propose a transfer learning-based deep learning model that is tailored for potato leaf disease classification. Transfer learning enables the model to benefit from pre-trained neural network architectures and weights, enhancing its ability to learn meaningful representations from limited labeled data. Additionally, Explainable AI techniques are integrated into the model to provide interpretable insights into its decision-making process, contributing to its transparency and usability. We used a publicly available potato leaf disease dataset to train the model. The results obtained are 97% for validation accuracy and 98% for testing accuracy. This study applies gradient-weighted class activation mapping (Grad-CAM) to enhance model interpretability. This interpretability is vital for improving predictive performance, fostering trust, and ensuring seamless integration into agricultural practices.

Machine Learning-Based Prediction of Drug Solubility in Lipidic Environments: The Sol_ME Tool for Optimizing Lipid-Based Formulations with a Preliminary Apalutamide Case Study.

Lipid-based formulations are essential for enhancing drug solubility and bioavailability, yet selecting optimal lipid excipients for specific drugs remains challenging. This study introduces Sol_ME, a machine learning-based model designed to predict drug solubility in lipidic environments, thereby streamlining the formulation process.The Sol_ME model uses PubChem® fingerprints, focusing on solubility correlations with lipid excipients, minimizing reliance on traditional parameters like LogP and molecular weight. The model was trained on a dataset of 1,379 drug-solvent entries and applied to the formulation of Apalutamide, a BCS Class II drug. Experimental validation was performed with 35 drug-solvent combinations to assess the accuracy of predicted solubilities.Sol_ME achieved a high predictive accuracy with a correlation coefficient of 0.998. The model successfully identified Cinnamon oil as the optimal excipient for Apalutamide, further refining the formulation with Vanillin. This reduced formulation volume by 75%, enabling the development of a single-unit 240mg soft gelatin capsule. Experimental validation showed 80% alignment between predicted and actual solubilities.The Sol_ME model demonstrates significant potential to optimize lipid-based drug formulation, offering a data-driven approach that enhances efficiency. The success of Apalutamide formulation highlights its practical utility. Future work will expand the dataset and extend the model to solid lipid systems, broadening its application in drug delivery technologies.

A machine learning-based model for predicting paroxysmal and persistent atrial fibrillation based on EHR

BackgroundThere is no effective way to accurately predict paroxysmal and persistent atrial fibrillation (AF) subtypes unless electrocardiogram (ECG) observation is obtained. We aim to develop a predictive model using a machine learning algorithm for identification of paroxysmal and persistent AF, and investigate the influencing factors.MethodsWe collected demographic data, medication use, serological indicators, and baseline cardiac ultrasound data of all included subjects, totaling 50 variables. The diagnosis of AF subtypes is confirmed by ECG observation for at least more than 7 days. Variable selection was performed by spearman correlation analysis, recursive feature elimination, and least absolute shrinkage and selection operator regression. We built a prediction model for AF using three machine learning methods. Finally, the significance of each variable was analyzed by Shapley additive explanations method.ResultsAfter screening, we found the optimal variable set consisting of 10 variables. The model we built achieved good predictive performance (AUC = 0.870, 95%CI 0.858 to 0.882), and had specificity of 0.851 (95%CI 0.844 to 0.858) and sensitivity of 0.716 (95%CI 0.676 to 0.755). Good predictive performance was stably achieved in different age subgroups and different gender subgroups. LA and NT-proBNP were the two most important variables for predicting paroxysmal and persistent AF in all models, except for the female subgroup aged less than 60 years.ConclusionsOur model makes it possible to predict paroxysmal and persistent AF based on baseline data at admission. Early and individualized intervention strategies based on our model may help to improve clinical outcomes in AF patients.

Deep learning-based AI model for sinusitis diagnosis

Problem statement While CT (Computed Tomography) is commonly used, its diagnostic accuracy for chronic sinusitis remains uncertain. Moreover, the high cost of CT examinations limits its use as a routine diagnostic method. There is an urgent need to develop an AI-assisted diagnostic model for sinusitis. Objective The primary aim of this study is to develop an AI-assisted diagnostic model for sinusitis that can improve diagnostic accuracy and accessibility compared to traditional CT methods. Methodology This study utilized a retrospective approach, focusing on patients diagnosed with chronic sinusitis via CT and normal patients admitted to the People's Hospital between January 2018 and January 2019. A total of 5000 sinus CT images were collected. All cases underwent T (targeted) coronal plain scans in the hospital's CT room, ensuring complete CT images. In constructing the chronic sinusitis classification model based on deep learning, 5000 CT images of soft tissue windows and sinuses were gathered. This included 1000 CT images for each of the four groups diagnosed with sphenoid sinusitis, frontal sinusitis, ethmoid sinusitis, and maxillary sinusitis, along with 1000 images from normal cases (250 images per group). The sigmoid function replaced the softmax function, and the binary cross-entropy function was used to assess the model's predictive accuracy. Results The model achieved an accuracy of 85.8%, outperforming doctors with low (71.7%), medium (78.4%), and senior (73.4%) qualifications. The model demonstrated high accuracy, superior feature extraction, and resolution capabilities.

HepatoPredict Accurately Selects Hepatocellular Carcinoma Patients for Liver Transplantation Regardless of Tumor Heterogeneity

Background/Objectives: Hepatocellular carcinoma (HCC) is a major cause of cancer-related deaths rising worldwide. This is leading to an increased demand for liver transplantation (LT), the most effective treatment for HCC in its initial stages. However, current patient selection criteria are limited in predicting recurrence and raise ethical concerns about equitable access to care. This study aims to enhance patient selection by refining the HepatoPredict (HP) tool, a machine learning-based model that combines molecular and clinical data to forecast LT outcomes. Methods: The updated HP algorithm was trained on a two-center dataset and assessed against standard clinical criteria. Its prognostic performance was evaluated through accuracy metrics, with additional analyses considering tumor heterogeneity and potential sampling bias. Results: HP outperformed all clinical criteria, particularly regarding negative predictive value, addressing critical limitations in existing selection strategies. It also demonstrated improved differentiation of recurrence-free and overall survival outcomes. Importantly, the prognostic accuracy of HP remained largely unaffected by intra-nodule and intra-patient heterogeneity, indicating its robustness even when biopsies were taken from smaller or non-dominant nodules. Conclusions: These findings support the usage of HP as a valuable tool for optimizing LT candidate selection, promoting fair organ allocation and enhancing patient outcomes through integrated analysis of molecular and clinical data.

Deep learning-based overall survival prediction in patients with glioblastoma: An automatic end-to-end workflow using pre-resection basic structural multiparametric MRIs

PurposeAccurate and automated early survival prediction is critical for patients with glioblastoma (GBM) as their poor prognosis requires timely treatment decision-making. To address this need, we developed a deep learning (DL)-based end-to-end workflow for GBM overall survival (OS) prediction using pre-resection basic structural multiparametric magnetic resonance images (Bas-mpMRI) with a multi-institutional public dataset and evaluated it with an independent dataset of patients on a prospective institutional clinical trial. Materials and methodsThe proposed end-to-end workflow includes a skull-stripping model, a GBM sub-region segmentation model and an ensemble learning-based OS prediction model. The segmentation model utilizes skull-stripped Bas-mpMRIs to segment three GBM sub-regions. The segmented GBM is fed into the contrastive learning-based OS prediction model to classify the patients into different survival groups. Our datasets include both a multi-institutional public dataset from Medical Image Computing and Computer Assisted Intervention (MICCAI) Brain Tumor Segmentation (BraTS) challenge 2020 with 235 patients, and an institutional dataset from a 5-fraction SRS clinical trial with 19 GBM patients. Each data entry consists of pre-operative Bas-mpMRIs, survival days and patient ages. Basic clinical characteristics are also available for SRS clinical trial data. The multi-institutional public dataset was used for workflow establishing (90% of data) and initial validation (10% of data). The validated workflow was then evaluated on the institutional clinical trial data. ResultsOur proposed OS prediction workflow achieved an area under the curve (AUC) of 0.86 on the public dataset and 0.72 on the institutional clinical trial dataset to classify patients into 2 OS classes as long-survivors (>12 months) and short-survivors (<12 months), despite the large variation in Bas-mpMRI protocols. In addition, as part of the intermediate results, the proposed workflow can also provide detailed GBM sub-regions auto-segmentation with a whole tumor Dice score of 0.91. ConclusionOur study demonstrates the feasibility of employing this DL-based end-to-end workflow to predict the OS of patients with GBM using only the pre-resection Bas-mpMRIs. This DL-based workflow can be potentially applied to assist timely clinical decision-making.

Radiomics Analysis of Different Machine Learning Models based on Multiparametric MRI to Identify Benign and Malignant Testicular Lesions.

To develop and validate a machine learning-based prediction model for the use of multiparametric magnetic resonance imaging(MRI) to predict benign and malignant lesions in the testis. The study retrospectively enrolled 148 patients with pathologically confirmed benign and malignant testicular lesions, dividing them into: training set (n=103) and validation set (n=45). Radiomics characteristics were derived from T2-weighted(T2WI)、contrast-enhanced T1-weighted(CE-T1WI)、diffusion-weighted imaging(DWI) and Apparent diffusion coefficient(ADC) MRI images, followed by feature selection. A machine learning-based combined model was developed by incorporating radiomics scores (rad scores) from the optimal radiomics model along with clinical predictors. Draw the receiver operating characteristic (ROC) curve and use the area under the curve (AUC) to evaluate and compare the predictive performance of each model. The diagnostic efficacy of the various machine learning models was evaluated using the Delong test. Radiomics features were extracted from four sequence-based groups(CE-T1WI+DWI+ADC+T2WI), and the model that combined Logistic Regression(LR) machine learning showed the best performance in the radiomics model. The clinical model identified one independent predictors. The combined clinical-radiomics model showed the best performance, whose AUC value was 0.932(95% confidence intervals(CI)0.868-0.978), sensitivity was 0.875, specificity was 0.871 and accuracy was 0.884 in validation set. The combined clinical-radiomics model can be used as a reliable tool to predict benign and malignant testicular lesions and provide a reference for clinical treatment method decisions.

Implementing cloud computing in drug discovery and telemedicine for quantitative structure-activity relationship analysis

This work aims to use cutting-edge machine learning methods to improve quantitative structure-activity relationship (QSAR) analysis, which is used in drug development and telemedicine. The major goal is to examine the performance of several predictive modeling approaches, including random forest, deep learning-based QSAR models, and support vector machines (SVM). It investigates the potential of feature selection techniques developed in chemoinformatics for enhancing model accuracy. The innovative aspect is using cloud computing resources to strengthen computational skills, allowing for managing massive amounts of chemical information. This strategy produces accurate and generalizable QSAR models. By using the cloud's scalability and constant availability, remote healthcare apps have a workable answer. The goal is to show how these methods may improve telemedicine and the drug development process. Utilizing cloud computing equips researchers with a flexible set of tools for precise and timely QSAR analysis, speeding up the discovery of bioactive chemicals for therapeutic use. This new method fits well with the dynamic nature of pharmaceutical study and has the potential to transform the way drugs are developed and delivered to patients via telemedicine.

Optimizing data-sampling period in a machine learning-based surrogate model for powder mixing simulations

Optimizing data-sampling period in a machine learning-based surrogate model for powder mixing simulations

Enhancing federated learning-based social recommendations with graph attention networks

Social recommendations are crucial in information services, particularly for mitigating information overload on the Internet. Graph Attention Networks (GATs) have emerged as an effective solution for the social recommendation problem, harnessing the benefits of representation learning from user-item and user-user graphs. However, traditional GAT-based social recommendation models often necessitate the centralized storage and aggregation of all the users’ social connections and item interactions. This centralization raises significant privacy concerns. To address these issues, we propose a federated learning-based social recommendation system using GATs. Specifically, we first design a structured-enhanced graph clustering algorithm to obtain soft clustering assignments for items. Based on these assignments, the similarity between pairwise items and their corresponding correlations are computed. Second, we employ graph attention networks to infer the local heterogeneity of social relationships, user-item interactions, and item relationships within the federated learning framework for each client. Using the soft clustering assignments, we derive hard clustering indicators and sample pseudo items. Local gradients of embedding and model parameters are then computed and uploaded to the central server for aggregation under the local differential privacy mechanism. Experimental results over real-world data sets indicate that the proposed model is more effective in privacy protection while maintaining acceptable performance compared with the central learning framework. Additionally, the proposed model is more stable in addressing the cold-start issue for items than state-of-the-art federated learning-based social recommendation models.

Machine learning techniques for independent gait recovery prediction in acute anterior circulation ischemic stroke

ObjectiveThis study aimed to develop and validate a machine learning-based predictive model for gait recovery in patients with acute anterior circulation ischemic stroke.MethodsBetween May and November 2023, 237 patients with acute anterior circulation ischemic stroke were enrolled. Patients were randomly divided into training and validation sets at a 7:3 ratio. Thirty-one medical characteristics were collected, and the Least Absolute Shrinkage and Selection Operator (LASSO) regression was applied to screen predictor variables. Predictive models were developed using the Random Survival Forest (RSF) and COX regression methods. The optimal model was identified based on C-index values. The SHapley Additive exPlanations (SHAP) method was employed to interpret the RSF model globally and locally.ResultsTen predictors were identified through LASSO regression, including age, gender, periventricular white matter hyperintensities (PVWMH), Montreal Cognitive Assessment (MoCA), National Institutes of Health Stroke Scale (NIHSS), enlarged perivascular spaces in basal ganglia (BG-EPVS), lacunes, parietal infarction, basal ganglia infarction, and Timed Up & Go (TUG) test score. The C-index values of the COX regression and RSF models were 0.741 and 0.761 in the training set and 0.705 and 0.725 in the validation set, respectively. SHAP analysis of the RSF model identified BG-EPVS, TUG, MoCA, age, and PVWMH as the top five most influential predictors of gait recovery.ConclusionThe RSF model demonstrated superior performance to the COX regression model in predicting gait recovery, offering a reliable tool for clinical decision-making regarding stroke patients’ prognoses.