Learning Framework Research Articles

MetaFusion-FL: A Cross Modality Federated Meta Learning Framework for Robust and Explainable Healthcare System

Mpox is a re-emerging zoonotic viral disease that attracted the attention of the whole world because of its spreading transmission and clinical similarity with other skin diseases. It is highly important that this identification is fast and accurate, even in remotely located areas or resource-limited settings. However, the conventional centralized deep learning models exhibit severe limitations regarding data privacy, modality variation, and scalability across varied clinical environments. To this end, this paper presents MetaFusion-FL, a new federated meta-learning framework that combines cross-modality image analysis based on a hybrid Transformer-Capsule model with Hierarchical Attention-Based Multimodal Fusion (HAMFM). The model can work on multi-source images as input, namely smartphone images, dermoscopic images, and clinical images, which are processed locally at edge hospitals without raw data transmission. Reptile federated meta-learning strategy guarantees quick personalization of models and global generalization. When evaluated on a wide dataset, MetaFusion-FL has a higher classification accuracy of 99.46%, precision of 99.52%, recall of 99.40%, and F1-score of 99.46% compared to other current models, including ViT-RLXGBFL (99.12%) and ResViT-FLBoost (98.78%). The framework is also resistant to image noise and is consistent and stable across federated clients. Besides, SHAP and Grad-CAM++ explanations are used to ensure interpretability in a clinical context. MetaFusion-FL is therefore a leap in the development of AI-based, privacy-preserving, and generalizable skin disease classification, particularly Mpox.

Adaptive Deep Learning Strategies for Formaldehyde Monitoring in Industrial Air Quality

Inhaling formaldehyde a chemical that is widely used in many different industries can have serious health consequences. In order to precisely detect formaldehyde levels in industrial air quality environments, this study makes use of deep learning techniques. Using sensor data gathered from high-risk industrial areas the study focuses on variables like air quality index temperature and humidity. The data is processed by Convolutional Neural Networks (CNNs), which identify trends linked to increases in formaldehyde concentrations. To improve model accuracy preprocessing of the data is done including feature scaling and outlier elimination. The model's performance is assessed using evaluation metrics like Mean Squared Error (MSE), sensitivity, specificity, and prediction accuracy. Results show that when compared to conventional regression models the CNN-based model considerably lowers false positives while achieving a high prediction accuracy. Rapid reaction to hazardous formaldehyde levels is made possible by the deep learning frameworks' real-time monitoring capability which lowers possible health hazards. To improve long-term prediction accuracy and trend identification future research will investigate the use of recurrent neural networks (RNN) for time-series analysis.

An Efficient Deep Learning Framework for Accurate Disease Classification

One of the leading causes of memory loss and thinking problems in older adults is a condition that affects human function over time. Detecting this condition early is important for better care and treatment. However, even with the latest technology in artificial intelligence (AI) and deep learning, the results are not convincing because the dynamic nature of the datasets. This study introduces a new deep learning approach that includes a tool called Grad-CAM, which helps explain how the AI makes decisions. Our goal is to build a reliable and understandable system that uses a special type of AI model called a convolutional neural network (CNN) to analyze online dataset images. The model includes techniques to reduce errors and handle different types of data, while Grad-CAM provides visual feedback showing what the model is focusing on. The system achieved 95% accuracy, performing better than other well-known models like Xception (94.40%) and InceptionV3 (93.20%). Overall, this work offers a highly accurate and transparent tool to support early detection of memory-related conditions, assist professionals in planning care, and open new possibilities for research in AI-supported health applications.

The Rapid Social Immersive Learning (RaSIL) framework: bridging the gap for MSW students without prior formal social work training

ABSTRACT The Rapid Social Immersive Learning (RaSIL) framework is an innovative pedagogical intervention aimed at rapidly fostering critical competencies in social work students with no prior formal social work training. The RaSIL framework integrates experiential learning through intensive 12-hour immersive fieldwork, enabling students to engage directly with marginalized communities and confront real-world social challenges. Grounded in Kolb’s theory of experiential learning, the framework emphasizes the development of compassion, social intelligence, and professional readiness in a compressed time frame. The study employed a single-arm pre-post design to assess the impact of the RaSIL framework on 37 first-year Master of Social Work (MSW) students in India. Social intelligence and compassion were measured using the Tromso Social Intelligence Scale (TSIS) and Compassion Scale (CS) before and after the intervention. Results demonstrated significant improvements in social intelligence and compassion scores, with multiple regression analyses highlighting key socio-demographic factors influencing these outcomes. The findings suggest that the RaSIL Framework can effectively cultivate essential competencies in social work students, contributing to their professional development within a constrained educational timeframe. This study offers valuable insights for educators and policymakers seeking to innovate social work education through experiential and immersive approaches.

Machine learning-driven prediction of risk factors for postoperative re-fractures in elderly OVCF patients with underlying diseases: model development and validation

BackgroundPostoperative re-fractures in elderly osteoporotic vertebral compression fracture (OVCF) patients with comorbidities pose a major clinical challenge, with rates up to 52%. Traditional risk models overlook complex underlying diseases interactions in elderly patients. This study pioneers a machine learning (ML) framework for this high-risk group, integrating multidimensional factors to predict re-fractures and identify novel predictors.MethodsWe analyzed 560 OVCF patients with comorbidities who underwent percutaneous vertebroplasty (PVP). Fourteen characteristic variables—including scoliosis, chronic kidney disease (CKD), mental disorders, and cardiovascular comorbidities—were selected using feature engineering. Six ML models [Random Forest (RF), XGBoost, support vector machine (SVM), etc.,] were trained and validated. Model performance was rigorously assessed via AUC-ROC, precision-recall curves, and decision curve analysis (DCA). SHapley Additive exPlanations (SHAP) values provided interpretable risk quantification.ResultsThe RF model achieved superior predictive performance (test AUC = 0.88, sensitivity = 0.77, specificity = 0.87), outperforming conventional approaches. Notably, we identified scoliosis (SHAP = 0.14), mental disorders (0.12), and CKD (0.10) as the three top risk factors, with biomechanical and comorbidity interactions playing pivotal roles. DCA confirmed high clinical utility, with RF providing the greatest net benefit across risk thresholds.ConclusionThis pioneering study establishes ML as a transformative tool for re-fracture prediction in OVCF patients with underlying diseases, uncovering previously underappreciated risk factors. Our findings highlight the critical need for integrated management of spinal deformity, mental health, and renal function in this vulnerable population. This ML framework offers a paradigm shift in personalized risk stratification and postoperative care.

AuBoKT: an auxiliary boosted knowledge tracing model

Purpose This research aims to introduce a knowledge tracing (KT) method that evaluates students’ knowledge mastery state dynamically and precisely by analyzing their historical interaction data. Design/methodology/approach The proposed KT method is called Auxiliary Boosted Knowledge Tracing (AuBoKT). First, this paper presents a novel difficulty evaluation approach that takes into account individual abilities and the number of problem solvers, providing a more accurate estimation of exercise difficulty. In addition, this paper extracts various auxiliary features to mimic the learning process, enriching the information available for modeling students’ knowledge states. Moreover, this paper proposes a sequential neural network-based performance prediction model, which not only predicts students’ performance on given exercises but also implicitly models their knowledge state. Findings Extensive experiments on three public real-world data sets are conducted. The experimental results highlight the significance and effectiveness of each component in our approach. Originality/value This research addresses classical test theory’s limitations in exercise difficulty assessment by introducing a multi-concept fusion mechanism for comprehensive KT. This paper proposes AuBoKT, a deep learning framework leveraging auxiliary features to model fine-grained student interactions while dynamically integrating educational forgetting/learning theories for improved knowledge state tracking accuracy.

Emotions and Reputation Learning by Audience Networks: A Research Agenda in Bureaucratic Politics

ABSTRACTAudiences that observe and interact with government agencies play a crucial role in shaping these agencies' reputations. However, existing research often treats these audience networks as monolithic, overlooking the inherent diversity in their cognitive and emotional processing of reputational information. This approach fails to account for the variations in how audiences experience and evaluate agencies. To address this gap, we propose a new research agenda focused on the role of emotions in bureaucratic politics. We introduce a novel theoretical framework of Reputation Learning, informed by Affect‐as‐Information Theory and Affective Intelligence Theory, to explore the downstream effects of emotions as content and as process in shaping judgment formation and information processing. Specifically, we identify emotion‐based components of bureaucratic reputation and examine how emotions influence audience decision‐making processes and perceptions of government agencies. We conclude by outlining four key contributions of this framework to advancing the study of emotions in bureaucratic politics.

Exploring CBC Data for Anemia Diagnosis: A Machine Learning and Ontology Perspective

Background: Anemia, a common health disorder affecting populations globally, demands timely and accurate diagnosis for treatment to be effective. The aim of this paper is to detect and classify four types of anemia: hgb, iron-deficiency, folate-deficiency, and B12-deficiency anemia. Methods: This paper proposes an ontology-enhanced machine learning (ML) framework to classify types of anemia from CBC data obtained from Kaggle, which contains 15,300 patient records. It evaluates the effects of classical versus deep classifiers on imbalanced and oversampled training samples. Tests include KNN, SVM, DT, RF, CNN, CNN+SVM, CNN+RF, and XGBoost. Another interesting contribution is the use of ontological reasoning via SPARQL queries to semantically enrich clinical features with categories like “Low Hemoglobin” or “Macrocytic MCV”. These semantic features were then used in both classical (SVM) and deep hybrid models (CNN+SVM). Results: Ontology-enhanced and CNN hybrid models perform competitively when paired with ROS or ADASYN, but their performance degrades significantly on the original dataset. There were tremendous performance gains with ontology-enhanced models in that Onto-CNN+SVM achieved an F1-score (1.00) for all the four types of anemia under ROS sampling, while Onto-SVM exhibited more than 20% improvement in F1-scores for minority categories like folate and B12 when compared to baseline models, except XGBoost. Conclusions: Ontology-driven knowledge coalescence has been shown to improve classification results; however, XGBoost consistently outperformed all other classifiers across all data conditions, making it the most robust and reliable model for clinically relevant decision-support systems in anemia diagnosis.

DRFW-TQC: Reinforcement Learning for Robotic Strawberry Picking with Dynamic Regularization and Feature Weighting

Strawberry harvesting represents a labor-intensive agricultural operation where existing end-effector pose control algorithms frequently exhibit insufficient precision in fruit grasping, often resulting in unintended damage to target fruits. Concurrently, deep learning-based pose control algorithms suffer from inherent training instability, slow convergence rates, and inefficient learning processes in complex environments characterized by high-density fruit clusters and occluded picking scenarios. To address these challenges, this paper proposes an enhanced reinforcement learning framework DRFW-TQC that integrates Dynamic L2 Regularization for adaptive model stabilization and a Group-Wise Feature Weighting Network for discriminative feature representation. The methodology further incorporates a picking posture traction mechanism to optimize end-effector orientation control. The experimental results demonstrate the superior performance of DRFW-TQC compared to the baseline. The proposed approach achieves a 16.0% higher picking success rate and a 20.3% reduction in angular error with four target strawberries. Most notably, the framework’s transfer strategy effectively addresses the efficiency challenge in complex environments, maintaining an 89.1% success rate in eight-strawberry while reducing the timeout count by 60.2% compared to non-adaptive methods. These results confirm that DRFW-TQC successfully resolves the tripartite challenge of operational precision, training stability, and environmental adaptability in robotic fruit harvesting systems.

Advanced Deep Lung Care Net: A Next Generation Framework for Lung Cancer Prediction

Lung cancer is a leading cause of cancer-related death globally, necessitating innovative diagnostic methods to enhance early detection and improve treatment effectiveness. This study presents "Advanced DeepLungCareNet," an enhanced deep learning framework designed to predict and classify lung cancer from medical imaging data with greater accuracy and reliability. The approach improves diagnostic efficacy by employing convolutional neural networks (CNNs) and incorporating sophisticated image processing algorithms. The study utilized the IQ-OTH/NCCD Lung Cancer Dataset from Kaggle, which includes a diverse collection of annotated medical images, such as computed tomography (CT) scans and X-rays. Data preprocessing included normalization, augmentation, and segmentation to improve input quality for the neural network. The model architecture has been refined with deeper convolutional layers, optimized pooling techniques, and sophisticated feature extraction algorithms, enabling the detection of minute anomalies and patterns in the imaging data. The performance evaluation metrics, including accuracy, precision, recall, F1-score, and AUC-ROC, illustrate the superiority of "Advanced DeepLungCareNet" over existing state-of-the-art models. The framework achieved exceptional sensitivity and specificity, reducing false positives and false negatives, which is crucial for clinical reliability. The model demonstrated remarkable accuracy in detecting lung cancer from CT scans, making it a valuable tool for assisting healthcare professionals in early diagnosis. This study emphasizes the transformative potential of "Advanced DeepLungCareNet" in clinical environments, offering a robust solution for the early diagnosis and risk evaluation of lung cancer. Future attempts will focus on integrating multi-modal datasets, incorporating real-world clinical data, and exploring transfer learning approaches to enhance and validate the model's effectiveness across various healthcare situations.

RLsite: Integrating 3D-CNN and BiLSTM for RNA-Ligand Binding Site Prediction

Abstract Accurate identification of RNA-ligand binding sites is essential for elucidating RNA function and advancing structure-based drug discovery. Here, we present RLsite, a novel deep learning framework that integrates energy-, structure- and sequence-based features to predict nucleotide-level binding sites with high accuracy. RLsite leverages energy-based three-dimensional representations, obtained from atomic probe interactions using a pre-trained ITScore-NL potential, and models their contextual features through a 3D convolutional neural network (3D-CNN) augmented with self-attention. In parallel, structure-based features, including network properties, Laplacian norm, and solvent-accessible surface area, together with sequence-based evolutionary constraint scores, are mapped along the RNA sequence and used as sequential descriptors. These descriptors are modeled using a bidirectional long short-term memory (BiLSTM) network enhanced with multi-head self-attention. By effectively fusing these complementary modalities, RLsite achieves robust and precise binding site prediction. Extensive evaluations across four diverse RNA-ligand benchmark datasets demonstrate that RLsite consistently outperforms state-of-the-art methods in terms of precision, recall, Matthews correlation coefficient (MCC), area under the curve (AUC), and overall robustness. Notably, on a particularly challenging test set composed of RNA structures containing junctions, RLsite surpasses the second-best method by 7.3% in precision, 3.4% in recall, 7.5% in MCC, and 10.8% in AUC, highlighting its potential as a powerful tool for RNA-targeted molecular design.

A Proposed Deep Learning Framework for Air Quality Forecasts, Combining Localized Particle Concentration Measurements and Meteorological Data

Air pollution in urban areas has increased significantly over the past few years due to industrialization and population increase. Therefore, accurate predictions are needed to minimize their impact. This paper presents a neural network-based examination for forecasting Air Quality Index (AQI) values, employing two different models: a variable-depth neural network (NN) called slideNN, and a Gated Recurrent Unit (GRU) model. Both models used past particulate matter measurements alongside local meteorological data as inputs. The slideNN variable-depth architecture consists of a set of independent neural network models, referred to as strands. Similarly, the GRU model comprises a set of independent GRU models with varying numbers of cells. Finally, both models were combined to provide a hybrid cloud-based model. This research examined the practical application of multi-strand neural networks and multi-cell recurrent neural networks in air quality forecasting, offering a hands-on case study and model evaluation for the city of Ioannina, Greece. Experimental results show that the GRU model consistently outperforms the slideNN model in terms of forecasting losses. In contrast, the hybrid GRU-NN model outperforms both GRU and slideNN, capturing additional localized information that can be exploited by combining particle concentration and microclimate monitoring services.

EquiCPI: SE(3)-Equivariant Geometric Deep Learning for Structure-Aware Prediction of Compound-Protein Interactions.

Accurate prediction of compound-protein interactions (CPI) remains a cornerstone challenge in computational drug discovery. While existing sequence-based approaches leverage molecular fingerprints or graph representations, they critically overlook the three-dimensional (3D) structural determinants of binding affinity. To bridge this gap, we present EquiCPI, an end-to-end geometric deep learning framework that synergizes first-principles structural modeling with SE(3)-equivariant neural networks. Our pipeline transforms raw sequences into 3D atomic coordinates via ESMFold for proteins and DiffDock-L for ligands, followed by physics-guided conformer reranking and equivariant feature learning. At its core, EquiCPI employs SE(3)-equivariant message passing over atomic point clouds, preserving symmetry under rotations, translations, and reflections, while hierarchically encoding local interaction patterns through tensor products of spherical harmonics. The proposed model is evaluated on BindingDB (affinity prediction) and DUD-E (virtual screening). EquiCPI achieves performance on par with or exceeding the state-of-the-art deep learning competitors.

Accelerating Data Set Population for Training Machine Learning Potentials with Automated System Generation and Strategic Sampling.

Machine Learning Interatomic Potentials (MLIPs) offer a powerful way to overcome the limitations of ab initio and classical molecular dynamics simulations. However, a major challenge is the generation of high-quality training data sets, which typically require extensive ab initio calculations and intensive user intervention. Here, we introduce Strategic Configuration Sampling (SCS), an active learning framework to construct compact and comprehensive data sets for MLIP training. SCS introduces the usage of workflows for the automated generation and exploration of systems, collections of MD simulations where geometries and run conditions are set up automatically based on high-level, user defined inputs. To explore nontrivial atomic environments, initial geometries can be assembled dynamically via collaging of structures harvested from preceding runs. Multiple automated exploration workflows can be run in parallel, each with its own resource budget according to the computational complexity of each system. Besides leveraging the MLIP models trained iteratively, SCS also incorporates pretrained models to steer the exploration MD, thereby eliminating the need for an initial data set. By integrating widely used software, SCS provides a fully open-source, automatic, active learning framework for the generation of data sets in a high-throughput fashion. Case studies demonstrate its versatility and effectiveness to accelerate the deployment of MLIP in diverse materials science applications.

SelfLoc: Robust Self-Supervised Indoor Localization with IEEE 802.11az Wi-Fi for Smart Environments

Accurate and scalable indoor localization is a key enabler of intelligent automation in smart environments and industrial systems. In this paper, we present SelfLoc, a self-supervised indoor localization system that combines IEEE 802.11az Round Trip Time (RTT) and Received Signal Strength Indicator (RSSI) data to achieve fine-grained positioning using commodity Wi-Fi infrastructure. Unlike conventional methods that depend heavily on labeled data, SelfLoc adopts a contrastive learning framework to extract spatially discriminative and temporally consistent representations from unlabeled wireless measurements. The system integrates a dual-contrastive strategy: temporal contrasting captures sequential signal dynamics essential for tracking mobile agents, while contextual contrasting promotes spatial separability by ensuring that signal representations from distinct locations remain well-differentiated, even under similar signal conditions or environmental symmetry. To this end, we design signal-specific augmentation techniques for the physical properties of RTT and RSSI, enabling the model to generalize across environments. SelfLoc also adapts effectively to new deployment scenarios with minimal labeled data, making it suitable for dynamic and collaborative industrial applications. We validate the effectiveness of SelfLoc through experiments conducted in two realistic indoor testbeds using commercial Android devices and seven Wi-Fi access points. The results demonstrate that SelfLoc achieves high localization precision, with a median error of only 0.55 m, and surpasses state-of-the-art baselines by at least 63.3% with limited supervision. These findings affirm the potential of SelfLoc to support spatial intelligence and collaborative automation, aligning with the goals of Industry 4.0 and Society 5.0, where seamless human–machine interactions and intelligent infrastructure are key enablers of next-generation smart environments.

Enhancing Solid Oxide Fuel Cells Development through Bayesian Active Learning

Abstract Ensuring the sustainable operation of solid‐oxide fuel cells (SOFCs) requires an understanding of the components' lifespan. Multiphase‐field simulation studies play a major role in understanding the underlying microstructural changes and the resulting property alterations in SOFCs over time. The primary challenge in such simulations lies in identifying a suitable model and defining its parametrization. This study presents an Active Learning framework combined with Bayesian Optimization to identify optimal model parameters to simulate the aging of nickel‐gadolinium doped ceria (Ni‐GDC) anodes. The study overcomes incompleteness and inconsistency of literature data, and navigates the complex, high‐dimensional parameter space, by leveraging experimental microstructure data and the power of the AL framework. The successful parameter search enables simulation studies of Ni‐GDC anode aging and performance during long‐term SOFC operation. This approach improves the accuracy of phase‐field simulations and offers a versatile tool for broader applications in SOFC development, predicting material behavior under various operational conditions.

A Robust Machine Learning Framework for Predicting Slag Eye Formation in Industrial Steelmaking Ladles

A Robust Machine Learning Framework for Predicting Slag Eye Formation in Industrial Steelmaking Ladles

DamageScope: An Integrated Pipeline for Building Damage Segmentation, Geospatial Mapping, and Interactive Web-Based Visualization

Effective post-disaster damage assessment is crucial for guiding emergency response and resource allocation. This study introduces DamageScope, an integrated deep learning framework designed to detect and classify building damage levels from post-disaster satellite imagery. The proposed system leverages a convolutional neural network trained exclusively on post-event data to segment building footprints and assign them to one of four standardized damage categories: no damage, minor damage, major damage, and destroyed. The model achieves an average F1 score of 0.598 across all damage classes on the test dataset. To support geospatial analysis, the framework extracts the coordinates of damaged structures using embedded metadata, enabling rapid and precise mapping. These results are subsequently visualized through an interactive, web-based platform that facilitates spatial exploration of damage severity. By integrating classification, geolocation, and visualization, DamageScope provides a scalable and operationally relevant tool for disaster management agencies seeking to enhance situational awareness and expedite post-disaster decision making.

Post-Traumatic Stress Disorder Diagnosis using Brain Cellular Resting-State Functional Magnetic Resonance Imaging with Stacked Deep Learning Framework

Background Post-traumatic stress disorder (PTSD) is caused by depression and stress affecting the brain's emotional, memory, and sensory processes. Materials and Methods This study investigates a stacked deep learning model for trauma-based PTSD disorder diagnosis using rs-fMRI scans. Twenty-eight individual subjects, fourteen PTSD, and fourteen healthy controls were used, and each subject had 140 Resting-State Functional MRI (rs-fMRI) scans. The selected subjects were assessed to obtain brain activation from twelve brain regions of interest. Results The boxplot was used to check the performance of twelve ROI brain regions. Different deep learning algorithms were used for classification through a 10-fold cross-validation approach. This study examines the efficacy of employing a stacked deep approach with two models in the realm of predictive modeling. Discussion The objective of the proposed tacking model is to enhance the overall prediction accuracy and durability by using the complementary attributes of each model. The stacked model achieved a 98.30% accuracy rate on the training dataset and 96.60% on the test dataset. Conclusion Using the proposed approach, we could detect PTSD at an early stage. The selected ROI regions could also discriminate healthy PTSD from infected regions due to trauma events such as violence, accidents, and terrorism.

Dual-Channel Catalytic Immunochromatography Empowered by Machine Learning: Ultrasensitive Detection of Escherichia coli O157:H7 via Magnetic CoFe2O4@HRP Nanocomposites.

Traditional immunochromatographic test strips face significant limitations in detecting trace levels of Escherichia coli O157:H7 due to insufficient sensitivity and reliability. To address this challenge, we developed a novel "three-In-One" nanoplatform based on magnetic CoFe2O4 NPs functionalized with horseradish peroxidase (HRP) for dual-channel lateral flow immunoassay (LFIA). The secondary catalytic channel, leveraging HRP-mediated oxidation of 3,3',5,5'-tetramethylbenzidine (TMB), enables signal amplification, achieving an unprecedented detection limit of 9 CFU/mL─a 100-fold improvement over conventional gold nanoparticle-based LFIA (930 CFU/mL) and a 10-fold enhancement compared to the noncatalyzed CoFe2O4 system (93 CFU/mL). The CoFe2O4@HRP nanocomposite demonstrates remarkable synergistic effects, combining the magnetic separation capability of CoFe2O4 with the catalytic activity of HRP. This integration not only enhances detection sensitivity but also improves the aqueous stability and antibody loading capacity. In real food sample analyses (pork and milk), the system exhibits excellent accuracy (recovery rate: 89.29-110.71%) and precision (RSD: 3.31-7.93%). To further optimize detection performance, we implemented a robust machine learning framework incorporating deep neural networks (DNN), random forest regression, and k-nearest neighbors algorithms. This predictive model achieved exceptional agreement with experimental results (R2 > 0.999), 100% classification accuracy at the order-of-magnitude level, and >95% of predictions within Bland-Altman agreement limits. This work establishes a new paradigm for foodborne pathogen detection by synergistically combining nanomaterial engineering with artificial intelligence, offering a novel paradigm in rapid, ultrasensitive, and quantitative diagnostics for food safety monitoring and clinical applications.