Alzheimer’s disease (AD) is a gradually developing neurodegenerative condition marked by substantial cognitive decline, underscoring the importance of early and precise diagnosis for optimal management. EEG has emerged as a valuable, non-invasive modality for detecting AD-related alterations in brain dynamics. In this work, we propose an end-to-end explainable machine learning framework for patient-independent AD diagnosis. The temporal EEG signals are converted into graph representations, in which each time point is mapped to a node, allowing the graph to directly encode the intrinsic complexity of EEG signals. A comparative study of seven visibility graph (VG) variants for EEG-based AD classification is conducted. By extracting and analysing topological metrics, measures of connectedness, and higher-order statistics, our study highlights the effectiveness of VG-based approaches in capturing distinctive characteristics of AD without the need for prior decomposition into frequency bands. The extracted features are evaluated using SVM, Random Forest, XGBoost, AdaBoost, and Extra Trees. Despite its simplicity, SVM consistently achieves the highest accuracy across all VG variants. The study confirms patient-independent classification, achieving 95.2% and 95.8% accuracies for AD classification under EC conditions in Dataset-I and Dataset-II, respectively. When both datasets are combined, the model further improves to 97.7%, demonstrating promising patient-independent classification performance under the evaluated limited-data conditions. To ensure transparency, ANOVA is applied as a statistical test to validate feature significance, and SHAP is used to explain the model’s decisions. The results underscore the effectiveness and reliability of the proposed framework for AD classification and its possible uses in neuroscience and clinical domains.
This paper presents a novel three-parameter probability distribution, the half Cauchy Rayleigh inverted Weibull distribution, constructed via a specific transformation of the half Cauchy distribution. The proposed model’s statistical properties, namely its probability density function, cumulative distribution function, moments, variance, skewness, kurtosis, order statistics, and hazard rate function, are provided. The performance of the maximum likelihood estimation method is evaluated using a Monte Carlo simulation. To estimate the parameters, four different sampling set designs are used. Comparisons with several well-known distributions are used to assess the flexibility and modeling power of the proposed distribution. Based on common goodness-of-fit metrics, the empirical findings indicate that the proposed model provides a better fit to the real data. The results demonstrate the novel distribution’s potential as a reliable and adaptable tool for modeling industrial data. The study concludes that the suggested distribution contributes significantly to the development of statistical models, especially in situations where a flexible yet economical distribution is needed.
Bilinear neural network method and Physic-Informed neural network method have been applied in many aspects of sciences. This paper investigates a class of exact and approximate solutions for a fourth-order nonlinear (3+1)-dimensional Ablowitz–Kaup–Newell–Segur (3AKNS) equation by combining the Hirota bilinear method with a Physics-Informed Neural Network (PINN) framework. First, the governing equation is reduced via a traveling wave transformation, and the Hirota bilinear form is constructed using a logarithmic transformation. Based on this formulation, various analytical solutions, including kink and lump structures, are derived for different parameter cases. To enhance computational efficiency and extend the solution capability, a PINN-based approach is developed to approximate the bilinear function. A softmax-weighted residual loss is introduced to improve convergence and stability during training. Extensive numerical experiments are conducted to analyze the performance of the proposed method in terms of loss convergence, accuracy, and residual decay. The graphical results demonstrate that the proposed framework successfully captures a wide range of nonlinear wave phenomena, including localized lump solutions and smooth kink-type waves. Moreover, the convergence analysis confirms the robustness and competitive accuracy of the method across all considered cases. The integration of analytical and deep learning techniques provides a flexible and efficient tool for studying high-dimensional nonlinear evolution equations. Furthermore, the practical effectiveness of the proposed structure is validated through a fluid flow prediction problem, where the model successfully captures nonlinear spatio-temporal flow patterns and achieves competitive prediction performance.
This paper represents a novel, model-free method for R-peak detection in electrocardiogram (ECG) signals, leveraging an adaptive outlier scoring technique called the mass-ratio-variance based outlier factor (MOF). Unlike traditional pretrained models, or fixed-threshold methods, this approach dynamically adjusts to incoming signal patterns, enabling robust, and accurate detection of R peaks across diverse, and previously unseen ECG data. The core of the method employs a rapid, overlapping window-based strategy that computes MOF scores to distinguish R-peaks as pronounced outliers within the ECG waveform. The algorithm incorporates Z-score normalization, iterative candidate identification, dynamic refractory period enforcement, and physiological post-processing filters to minimize false positives, and maintain beat-to-beat consistency. Experimental evaluation on the widely used MIT-BIH Arrythmia Database demonstrates that the proposed MOW-ECG algorithm achieves excellent performance with a sensitivity of 96.71%, and a positive predictive value of 99.39%, outperforming several state-of-the-art methods in accuracy, and robustness. The adaptive, and model-free nature of the algorithm make it especially suitable for real-time applications, and continuous monitoring scenarios where pretraining is impractical or infeasible. This work contributes a computationally efficient, and interpretable solution for automated cardiac even detection, highlighting its potential integration into clinical ECG systems for enhanced patient care, and diagnostic reliability.
