Seasonal aerosol variations at the Land-Ocean boundary: Insights from a global AERONET network analysis.

Tele-monitoring is a useful platform for remote monitoring of cardiac patients, where compression plays a significant role in reducing the link burden and memory utilization of the source device. This paper describes a new approach for lossless ECG compression based on a deep-learning method via an adaptive autoregressive integrated moving average (ARIMA) model. Raw ECG signals were denoised and preprocessed to generate beat-cells for further processing. The ARIMA model uses the individual cardiac cycles to generate model parameters, which are then compressed. In this research, the optimal model hyperparameters were predicted by a deep autoencoder followed by a multilayer perceptron neural network (MLPNN) regressor combination. The predictor was tuned offline via particle swarm optimization (PSO), which produced the reference data for MLPNN tuning. The technique uses 46 records of mitdb under PhysioNet, including 10 major abnormal beats: H, A, V, P, L, R, a, f, F and j. Because of the adaptive nature, compression quality is high with negligible loss. No deviations in the clinical features of the reconstructed beats are found. The mean CR and PRD% values were 41.51 and 0.209%, respectively, which are superior to those reported in published research on ECG compression. The proposed adaptive ECG compression model can be useful for real-time telemonitoring applications, efficient storage and transmission of streamlined data of critical patients under continuous monitoring.

Complex data streams are highly dynamic, large-scale, and prone to continuous distributional shifts, posing significant challenges for neural network-based classification systems-particularly in maintaining accuracy and efficiency without frequent retraining. To address these issues, this study proposes an intelligent incremental learning framework that integrates a Dynamic Grasshopper Optimization Algorithm (DGOA) with a Multilayer Perceptron (MLP) neural network for real-time hyperparameter optimization. The proposed system functions as an adaptive intelligent system, utilizing DGOA-an enhanced form of the traditional Grasshopper Optimization Algorithm equipped with dynamic parameter control and online swarm reconfiguration-to autonomously adjust to evolving data characteristics. Unlike conventional GOA, the dynamic variant modifies its search behavior and population dynamics in real time, enabling continuous learning without restarting the optimization process. Through this mechanism, the model incrementally tunes critical hyperparameters such as the learning rate and momentum, resulting in improved accuracy and generalization on unseen data. The main contribution of this research lies in developing a fully online, swarm-intelligence-driven hyperparameter optimization strategy tailored for big data streams. Experimental evaluations on the Australian electricity market dataset demonstrate that the DGOA-based MLP achieves a classification accuracy of 89.5%, outperforming Grid Search (84.2%), Random Search (83.5%), PSO (86.7%), GA (87.1%), ACO (86.9%), and standard GOA (87.8%). Additionally, DGOA reduces the average computational time to 120s and converges in only 30 iterations while achieving the lowest final loss (0.21), highlighting its superior efficiency and convergence stability.

The digital transformation of the financial services sector, accelerated by the emergence of neobanks and advanced online platforms, has markedly increased its exposure to sophisticated cyberthreats. High-profile incidents, such as coordinated attacks on financial institutions in Iraq, have demonstrated the severe operational, economic, and reputational consequences that can arise from delayed threat detection and inadequate mitigation. Traditional cybersecurity measures, including firewalls, antivirus software, and signature-based intrusion detection systems, remain constrained by their dependence on known attack signatures, thereby leaving financial networks susceptible to zero-day exploits, AI-driven intrusions, and complex multi-vector threats. This study proposes and evaluates a supervised machine learning intrusion detection and prevention model aimed at proactively securing financial networks at a network level. To simulate realistic network conditions and generate representative traffic data, a banking environment was constructed using GNS3. To address class imbalance within the dataset, the Synthetic Minority Oversampling Technique (SMOTE) was employed, thereby improving the detection of minority-class attack instances. Several machine learning algorithms, including Support Vector Machine, Multi-Layer Perceptron Neural Network, and Long Short-Term Memory, were assessed using key performance metrics to determine their effectiveness. The Decision Tree model demonstrated superior performance, achieving an accuracy rate of 99.98%, perfect precision and recall, zero false positives, and only thirteen false negatives. These results underscore its capacity to deliver highly accurate, real-time threat detection while minimising operational disruptions caused by false alarms. Its transparent decision-making process enhances explainability, supports regulatory compliance, and fosters institutional trust, factors that are critical in financial cybersecurity. The findings validate the viability of interpretable, high-performance machine learning models for the real-time detection and mitigation of advanced cyberthreats, including Distributed Denial-of-Service (DDoS) attack patterns. Future research should prioritise scaling the simulation framework to encompass more complex financial network topologies, integrating adaptive online learning capabilities, and incorporating explainable artificial intelligence (XAI) techniques to investigate whether enhanced model interpretability improves threat detection accuracy and analyst response times.

ABSTRACT Crack evaluation in the fracturing pump chamber is essential to ensure the safe conduct of oil and gas fracturing construction. However, the traditional manual evaluation methods have bottlenecks such as downtime for dismantling, grave danger, low efficiency, etc., and detecting crack development in real-time is challenging. This paper proposes an online crack evaluation framework based on an improved multilayer perceptron (MLP), which aims to effectively ensure the safety of equipment construction and personnel, and provide real-time feedback on crack development information. The framework includes (1) Adaptive threshold noise reduction and wavelet packet transform. It effectively resists external complex interference and obtains damage-sensitive signal features as an energy spectrum. (2) MLP network based on arithmetic whale optimisation. It realises high-precision identification of the cavity’s crack state and reduces computation and parameters compared with other advanced networks. (3) The first online crack evaluation system in the fracturing industry. Through oilfield engineering field tests, it can realise online evaluation and early warning of crack development in the pump body cavity during fracturing, with a comprehensive accuracy rate of 87.06%, further validating the engineering feasibility and broad application prospect of the proposed framework.

This study introduces a new perspective into deep learning in the light of a multimodal approach: cardiovascular events can be predicted, using real-time data of physiological signals in collaboration with metadata related to the patient. Electronic Health Records (EHR) are digital versions of patients’ medical histories, while Multilayer Perceptron (MLP) and Convolutional Neural Network (CNN) are deep learning architectures designed for processing structured data and spatial/temporal patterns, respectively. A hybrid neural network model is designed that allows taking, as input from the CNN, the 12-lead ECG signals, while an MLP processes patient demographic and clinical features. It is designed to simultaneously process temporal ECG patterns and static patient characteristics for all-rounded cardiovascular risk assessment. In this work, our dataset consisted of 17,441 ECG recordings per patient, each being a 12-channel signal sampled on 500-time points and patient metadata like age, sex, and weight. Our architecture has two specialised components: the proposed SignalCNN to process the waveforms including two convolutional layers with batch normalization and dropout as regularization and MetaMLP processing patient metadata. These combined features are then fed into a classifier to enable multi-label prediction of five common cardiovascular conditions. The model yielded very promising results and performed very robustly with an overall validation accuracy of 85.19% after 15 epochs of training. The training was improving smoothly for both training and validation metrics, while the validation loss decreased from 0.4298 to 0.3484, which is indicative of good generalization. The model was very stable in its training without showing any hint of overfitting thanks to strategic dropout and batch normalization. This work will contribute to cardiovascular healthcare with a real-time, automated system that can be used for the early detection of cardiac events. The approach is multimodal, offering more nuanced predictions by including instantaneous physiological signals, together with patient-specific factors. This may enable earlier and more accurate clinical assessment of cardiovascular risk.

This study investigates the natural convection flow of Williamson fluid between two concentric cylinders while affected by the radiation effect and magnetic field. The inner cylinder remains fixed while the outer cylinder rotates. Additionally, magnetic field is oriented radially, which influences the flow of the fluid. Applying a proper transformation, one transform the non-linear partial differential equations of the Williamson fluid model into ordinary differential equations. Artificial neural networks (ANN) facilitate the computation of solutions to these nonlinear ordinary differential equations. Trial functions employ a multilayer perceptron neural network with tunable parameters, including weights and biases. The governing equations are satisfied by determining the trial solution’s changeable parameters by applying the Adam (adaptive moment estimation algorithm) optimization technique. Compared to the analytical solutions, the ANN’s result demonstrates good accuracy. Moreover, graphs show how pertinent parameters affect the velocity and temperature profiles. The temperature and velocity profiles get smaller as the magnetic parameter value increases. Furthermore, the temperature and velocity profiles increase as the Hall parameter value rises.

Allostatic load scores (ALSs) quantify the cumulative physiological burden of sustained stress across neuro-endocrine, metabolic, cardiovascular and inflammatory domains; however ALS is heterogeneous in nature. Using domain-specific urinary metabolomic signatures may improve evaluation accuracy, provide an innovative alternative to stress characterization, identify early domain-specific perturbations and allow comparative investigations. We designed and evaluated a novel, multilayered perceptron neural network (MLP-NN) method to investigate metabolic perturbations reflecting domain-specific alterations in low and high sustained stress, measured by ALS, and described ALS domain-specific metabolomic profiles. Data from 955 South Africans were used. ALS was calculated from dehydroepiandrosterone (DHEA), adrenocorticotropic hormone (ACTH), cortisol, interleukin-6 (IL-6), C-reactive protein (CRP), waist circumference (WC), glycated haemoglobin (HbA1c), blood-pressure and high-density lipoprotein cholesterol. Urinary amino acids and acylcarnitines (N = 32 metabolites) were analysed using liquid chromatography-tandem-mass-spectrometry. MLP-NN assessed metabolite contribution to the allostatic load (AL) domains, controlling for confounders, identifying the main metabolites, per AL domain. The median ALS was 3, with high stress (ALS ≥ 4) observed in 30% of participants. Significant differences were observed across all 32 metabolites between high and low ALS groups (all P < 0.05). MLP-NN revealed distinct domain-specific metabolomic patterns in low and high ALS. In low ALS the neuro-endocrine, cardiovascular and metabolic domains showed metabolomic signatures reflective of normal physiology. However in high ALS, metabolomic profiles reflected compensatory mechanisms linking neurotransmitter synthesis, redox balance and energy metabolism, mainly in the neuroendocrine, inflammatory and metabolic domains. This novel MLP-NN-based approach identified unique urinary profiles reflective of higher AL, independent of traditional confounders. This non-invasive approach may serve as an alternative for assessing AL, retaining domain-specificity, yet allowing comparative studies without the heterogeneity of traditional ALS. KEY POINTS: Allostatic load scores (ALSs) are heterogenous, complicating cross-study comparisons and clinical inferences. Determining allostatic load domain-specific metabolomic profiles using neural networks may identify early changes in specific domains. This study designed and evaluated a novel neural network-based model determining allostatic-load-domain-specific metabolomic signatures in high and low sustained stress. This novel neural-network-based approach, combinedly analysing metabolomic data and ALS domain-specific patterns, identified unique urinary profiles, independent of traditional confounders. This non-invasive approach may serve as an alternative for assessing AL, retaining the domain-specificity, identifying early domain-related perturbations related to sustained stress and allowing comparative studies.

Abstract This study introduces a novel network-based crime risk score, the Street Segment Risk Score (SSRS), designed to enhance crime hotspot predictions on street networks. The SSRS evaluates the risk of individual street segments by incorporating the dynamic spatial influence of nearby urban features on local crime patterns. Our dataset comprises all reported incidents of robbery ( n = 2,016) and theft ( n = 31,493) from 2015 to 2018 in Chicago’s Central Side (CS). We developed both daily and intraday crime hotspot prediction models that integrate the SSRS and compared their performance—with and without the SSRS—using two graph-based deep learning algorithms, Graph WaveNet (GWNet) and the Spatiotemporal Graph Convolutional Neural Network (STGCN); a traditional deep learning model, Long Short-Term Memory (LSTM); and two baseline methods, Multilayer Perceptron (MLP) and Spatiotemporal Network Kernel Density Estimation (STNetKDE). Results indicate that incorporating the SSRS improves daily robbery hotspot prediction accuracy by up to 5.3% and intraday theft prediction accuracy by as much as 33%. The proposed SSRS demonstrates strong potential to support more precise, street-level security interventions by enhancing daily and intraday crime hotspot predictions.

Intraocular pressure prediction method combining finite element simulations and multi-layer perceptron neural network

The goal of this study was to evaluate the performance of traditional gradient boosting (GB) and neural network models on diverse tabular datasets that differ in scale, class balance, and feature composition (numerical, categorical, or mixed). We focused on six representative datasets: adult census income, bank marketing, credit card fraud, breast cancer diagnosis, diabetes, and in-vehicle coupon recommendation, each with distinct challenges related to dimensionality, sample size, and heterogeneity. We benchmark the predictive performance of XGBoost and LightGBM (gradient boosting models) against Multilayer Perceptrons (MLP), Tabular Transformers, and tabular prior-data fitted network (TabPFN), using metrics such as accuracy, F1 score, ROC-AUC, and log loss. To ensure transparency and interpretability, we applied SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanation (LIME) to all models and evaluated the explanation quality using stability, fidelity, and consistency criteria. Our findings confirm that gradient boosting models consistently achieve the best balance of performance, calibration, and interpretability across heterogeneous and imbalanced datasets. SHAP-based insights show that gradient boosting (GB) models provide more stable and interpretable feature attributions, making them well suited for high-stakes domains such as finance and healthcare. These results emphasize the practical advantages of gradient boosting methods for structured data tasks and highlight the interpretability limitations of deep learning models when applied to tabular datasets. Future work will explore hybrid architectures and pretraining strategies to close this performance gap.

Emotion recognition from speech plays a crucial role in enhancing human–computer interaction by enabling systems to interpret and respond to users’ emotional states. This study develops and evaluates a Speech Emotion Recognition (SER) system using three machine learning techniques; Support Vector Machines (SVM), Multilayer Perceptron (MLP), and Convolutional Neural Networks (CNNs). The system is trained and tested on the RAVDESS dataset, which contains 1,440 professionally recorded audio samples representing a wide range of emotions. Our approach involves careful preprocessing of the audio signals, extraction of key acoustic features, and comparative performance evaluation of the three models using standard metrics. Results show that each model exhibits unique strengths and limitations, with CNNs achieving the most robust feature learning and generalization. The study underscores the importance of diverse feature representation for accurate emotion classification and provides insight into how different model architectures handle emotional nuances in speech. Identified challenges such as dataset diversity, feature selection, and computational complexity are discussed, along with recommendations for future research to improve SER systems’ real-world adaptability. This work contributes to ongoing efforts toward developing emotionally aware technologies that can enhance natural human–machine communication.

Liver disease poses a severe threat to human health if not detected early. Existing diagnostic methods are usually time-consuming, expensive, and require expertise, which is often unavailable in healthcare facilities. This study introduces a hybrid AI-based diagnostic framework that integrates both Deep Learning (DL) and Machine Learning (ML) techniques to support the early and accurate detection of liver disease. The proposed hybrid model integrates a MultiLayer Perceptron Neural Network (MLPNN) with a soft Voting Classifier, which includes Extreme Gradient Boosting (XGB) and Light Gradient Boosting Machine (LGBM). To enhance the predictive performance of the model, advanced feature engineering techniques were employed, including formulating medically pertinent ratios and balancing the data using SMOTE-Tomek resampling. The proposed hybrid model achieved an accuracy of 95.49%, demonstrating remarkable generalization capabilities across the dataset. The proposed model is strong and reliable, as demonstrated by the confusion matrix, classification report, and ROC-AUC curve results.

The selection of hazard factors is an important factor affecting the accuracy of landslide susceptibility mapping (LSM). The systematic development of an integrated input framework, incorporating both static and time-varying factors, as well as comparative studies of different input frameworks, remains at a preliminary stage. The degree of fit between each data-driven method and landslide-prone environment cannot be known in advance, so the best modeling method can only be determined through comparative studies. Therefore, the Pearson correlation coefficient method and collinearity diagnostics were used to screen the hazard factors, and three hazard factor combinations, considering both static and time-varying factors, were established. A total of 4498 landslide grids and 4498 non-landslide grids were determined, among which 70% (3149 landslide grids and 3149 non-landslide grids) were training samples, and the remaining 30% (1349 landslide grids and 1349 non-landslide grids) were verification samples. The three combinations were input to five models (Support Vector Machine, Random Forest, Convolutional Neural Network-Random Forest, Convolutional Neural Network-Support Vector Machine and Deep Belief Network-Multilayer Perceptron). The results show that the LSM results of different combinations and models are quite varied, and the combination No.3 and the Deep Belief Network-Multilayer Perceptron are the best. The study area is divided into extremely low susceptible areas, low susceptible areas, medium susceptible areas, high susceptible areas and extremely high susceptible areas, and the extremely high susceptible areas mainly distribute in the northwest, south and east. The other models overestimate the distance from the fault and underestimate the distance from the road. The extreme tendency of LSM results of the combinations No.1 and No.2 are strong, and they are easy to produce error estimation areas, which overestimate the elevation and underestimate the distance from the river. The LSM results of the Convolutional Neural Network-Support Vector Machine are closer to those of the benchmark, which underestimates the distance from the road and overestimates the distance from the fault. This study selected the best combination and model through comparative studies and revealed the degree of influence of each hazard factor on landslide susceptibility, greatly improving LSM accuracy, which can provide a scientific basis for land use planning.

The Domain Name System (DNS) plays a critical role in the functioning of the Internet, providing essential resolution services for nearly all user activities. In this work, we examine the hypothesis that individual users exhibit recurrent and distinctive patterns in their DNS query behavior, which can be leveraged to create unique and robust user fingerprints. Building on a publicly available dataset of real DNS traffic collected from a large-scale network, we evaluate the feasibility of user identification solely based on these behavioral DNS traces, independent of IP address stability. We conducted a comparative study of several machine learning models - including Naive Bayes, Random Forests, XGBoost, Multilayer Perceptrons, and Convolutional Neural Networks - on their ability to classify users based on domain category frequencies and derived statistical features. After extensive data preprocessing, dimensionality reduction, and feature selection, our best-performing model (CNN) achieves a classification accuracy of 86.7% across 1727 classes (unique IP addresses). The results confirm the viability of DNS-based user fingerprinting, even in the presence of dynamic IP addresses. Our approach opens new avenues for applications in network forensics and anomaly detection, while also raising important questions about privacy and ethical use of passive traffic analysis.

The persistent increase in healthcare expenditure has become a major challenge for the sustainability of public financing worldwide. Therefore, identifying the characteristics of at-risk population and their predictability for healthcare use is crucial to inform targeted policy and interventions to curb with increasing healthcare use and expenditure. Drawing on three waves of the HILDA survey that included a ‘health module’, the study applied four machine learning (ML) methods–Random Forest, Gradient Boosting Decision Trees, Extreme Gradient Boosting, and multilayer perceptron neural networks and conventional logistic regression for prediction of non-emergency healthcare use (specifically primary and tertiary inpatient hospital care). Predictive performance for the classifiers was evaluated using accuracy, sensitivity, and specificity measures, and area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and Matthews Correlation Coefficient (MCC). Calibration of the models was assessed using Brier score, which measures the mean squared difference between predicted probabilities and observed outcomes with lower values indicating better calibration. Finally, Local Interpretable Model-agnostic Explanations (LIME) was conducted to explain the model’s predictive behaviour, while SHAP results are provided for each wave along with a representative SHAP plot as a demonstration which uses the probability-contribution scale. Based on 47,899 observations and 741 variables, our model identified socio-economic factors (age, socio-economic status, private insurance status) and health-related variables (e.g. previous contact with healthcare service) and having a designated doctor to see when sick or for health advice were strong predictors of healthcare use. Between the different ML techniques, Gradient Boosting Decision Trees provided better prediction performance on healthcare use compared with logistic regression across all three waves. Although the standard logistic regression produced AUC of 0.69, had 71% positive predictive value (PPV), and 52% negative predictive value (NPV), with 86% sensitivity and 30% specificity, the ML models produced AUC in the range of 0.68 to 0.76, PPV of 75% to 77%, and NPVs of 61% to 63% with sensitivity ranging between 0.86 and 0.89, specificity between 0.40 and 0.44 and brier scores ranging between 0.11 and 0.28. The novelty of using ML techniques on a large, nationally representative longitudinal household survey data that covers a range of different domains provided more robust estimates on factors influencing future healthcare use (primary and inpatient elective care) which are important to inform resource allocation decisions and priority setting.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-28968-6.

Background: The rapid increase in the number of Mobile Phone Base Stations (MPBS) has raised global concerns about the potential adverse health effects of exposure to Radiofrequency Electromagnetic Fields (RF-EMF). The application of machine learning techniques can enable healthcare professionals and policymakers to proactively address concerns surrounding RF-EMF exposure near MPBS. Objective: The current study aimed to investigate the potential of machine learning models for the prediction of health symptoms associated with RF-EMF exposure in individuals residing near MPBS.Material and Methods: This analytical study utilized Support Vector Machine (SVM) and Random Forest (RF) algorithms, incorporating 11 predictors related to participants’ living conditions. A total of 699 adults participated in the study, and model performance was assessed using sensitivity, specificity, accuracy, and the Area Under Curve (AUC). Results: The SVM-based model demonstrated strong performance, with accuracies of 85.3%, 82%, 84%, 82.4%, and 65.1% for headache, sleep disturbance, dizziness, vertigo, and fatigue, respectively. The corresponding AUC values were 0.99, 0.98, 0.920, 0.89, and 0.81. Compared to the RF model and a previously developed model, the SVM-based model exhibited higher sensitivity, particularly for fatigue, with sensitivities of 70.0%, 83.4%, 85.3%, 73.0%, and 69.0% for these five health symptoms. Particularly for predicting fatigue, sensitivity and AUC were significantly improved (70% vs. 8% and 11.1% for SVM, Multilayer Perceptron Neural Network (MLPNN), and RF, respectively, and 0.81 vs. 0.62 and 0.64, for SVM, MLPNN, and RF, respectively). Conclusion: Machine learning methods, specifically SVM, hold promise in effectively managing health symptoms in individuals residing near or planning to settle in the vicinity of MPBS.

Accurate thickness measurement of aluminum sheets is critical for industries such as aerospace and automotive but is challenged by traditional methods' dependency on known alloy compositions. This study proposes a novel x-ray-based system to determine thickness across four aluminum alloys (1050, 3105, 5052, and 6061) with thicknesses ranging from 1 to 45mm, independent of composition. Using Monte Carlo N-particle simulations, an optimized multi-layer perceptron (MLP) neural network, and ant colony optimization (ACO) for feature selection, the approach achieves precise predictions with reduced computational complexity. The model demonstrated high accuracy, with a mean relative error (MRE) of 1.06% on test data, outperforming conventional methods. This scalable, calibration-free system offers a robust solution for real-time thickness measurement in diverse industrial applications.

Electron spectra measurements in linear accelerators via neural network reconstruction from percentage depth dose (PDD) data.

To address the challenges of developing machine learning frameworks for Electronic Health Records (EHRs)-based predictive tasks, such as the intricate occurrence mechanism of clinical events, patient diversity, and the inherent limitations of real-world data like data incompleteness and class imbalance, we propose the Knowledge-augmented Patient Network embedding-based Dynamic model Selection (KPNDS) framework, focusing on two key aspects: dynamically selecting the most suitable model for each individual and integrating biomedical knowledge into the framework. KPNDS utilizes graph machine learning algorithms to generate patient embeddings from a knowledge-augmented network which integrates data from a diverse range of data sources including EHRs, drug-related information, toxicogenomics data and other relevant information to enrich the understanding of patients. A meta-learning based framework is adopted to dynamically select the optimal classifiers based on the latent patient representations to perform individualized risk prediction. Multi-Layer Perceptron, Transformer and Kolmogorov-Arnold Networks are used as meta-classifiers to enhance the selection of the optimal classifiers for each patient. The KPNDS framework was validated for the early prediction of drug-induced liver injury (DILI) in pediatric patients. Experimental results show that it outperforms common baseline models and dynamic ensemble selection methods. The KPNDS framework effectively integrates domain knowledge, graph-based machine learning and dynamic model selection strategies, thereby enhancing predictive performance. The KPNDS framework seamlessly integrates knowledge-augmented networks with dynamic model selection techniques, which has the potential to enable more accurate risk assessment and personalized medicine in complex scenarios, highlighting a novel approach to integrating external knowledge with data-driven models.