International Journal of Advances in Intelligent Informatics
Not a member yet
235 research outputs found
Sort by
Analyzing risk factors and handling imbalanced data for predicting stroke risk using machine learning
Full text linkStroke is a serious medical condition resulting from disturbances in blood flow to the brain, signaling a chronic health issue that requires an immediate response. Principal risk factors increasing the likelihood of stroke include the presence of pre-existing conditions such as Diabetes Mellitus (DM), hypertension, and high cholesterol levels. Effective preventive measures are crucial to minimize stroke risk, and using predictive methods based on data analysis from the clinical examination dataset over the last three years (2019-2021), known as the general checkup (GCU) dataset, presents an innovative approach. This study aims to predict an individual's stroke risk for the following year. In this context, the study also addresses the preprocessing stage of the GCU dataset, which includes solutions for missing values by substituting them with the statistical mean, label encoding, feature correlation analysis using entropy values, and addressing data imbalance with the Adaptive Synthetic (ADASYN) technique. To evaluate their predictive performance, the research involves comparisons among various machine learning models. The outcome of the experiment shows that the Random Forest model is the best model, with 98.7% accuracy and 63.9% F1-Score. This research highlights the importance of preemptive measures against stroke by utilizing predictive techniques on clinical data, with the Random Forest model proving most effective in forecasting stroke probability
Enhanced diabetes and hypertension prediction using bat-optimized k-means and comparative machine learning models
No full textThis research aims to develop an analytical approach in classification statistics. The proposed approach is the use of machine learning combined with optimization effects. Considering the urgency of research related to exploring the best methods to apply to sports data. This study proposes a novel framework by combining the clustering results of random forest from the k-means method with the bat algorithm optimization to enhance performance prediction in the case of athlete prediction. The proposed method aims to explore data by comparing the quality of classification results in random forest machine learning, extremely randomized trees, and support vector classification methods. We conducted a case study on primary data with 200 respondents from Surabaya State University and the East Java National Sports Committee. The accuracy found in this study indicates that the best approach based on the performance evaluation metric of the proposed approach is the random forest clustering results from the k-means method with bat algorithm optimization, which provides the best accuracy value compared to other machine learning approaches at 81.25%. This research offers a novel machine-learningβoptimization framework that significantly improves athlete performance prediction by integrating k-means clustering, random forest, and bat algorithm optimization. The approach provides higher accuracy than conventional classifiers, enabling more data-driven decision-making for talent identification and sports analytics in Indonesia
Geometry-aware light field angular super-resolution using multiple representations
Full text linkLight Field Angular Super-Resolution (LFASR) is a critical task that enables applications such as depth estimation, refocusing, and 3D scene reconstruction. Acquiring LFASR from Plenoptic cameras has an inherent trade-off between the angular and spatial resolution due to sensor limitations. To address this challenge, many learning-based LFASR methods have been proposed; however, the reconstruction problem of LF with a wide baseline remains a significant challenge. In this study, we proposed an end-to-end learning-based geometry-aware network using multiple representations. A multi-scale residual network with varying receptive fields is employed to effectively extract spatial and angular features, enabling angular resolution enhancement without compromising spatial fidelity. Extensive experiments demonstrate that the proposed method effectively recovers fine details with high angular resolution while preserving the intricate parallax structure of the light field. Quantitative and qualitative evaluations on both synthetic and real-world datasets further confirm that the proposed approach outperforms existing state-of-the-art methods. This research improves the angular resolution of the light field without reducing spatial sharpness, supporting applications such as depth estimation and 3D reconstruction. The method is able to preserve parallax details and structure with better results than current methods
LUNGINFORMER: A Multiclass of lung pneumonia diseases detection based on chest X-ray image using contrast enhancement and hybridization inceptionresnet and transformer
No full textLung pneumonia is categorically a serious disease on Earth. In December 2019, COVID-19 was first identified in Wuhan, China. COVID-19 caused severe lung pneumonia. The majority of lung pneumonia diseases are diagnosed using traditional medical tools and specialized medical personnel. This process is both time-consuming and expensive. To address the problem, many researchers have employed deep learning algorithms to develop an automated detection system for pneumonia. Deep learning faces the issue of low-quality X-ray images and biased X-ray image information. The X-ray image is the primary material for creating a transfer learning model. The problem in the dataset led to inaccurate classification results. Many previous works with a deep learning approach have faced inaccurate results. To address the situation mentioned, we propose a novel framework that utilizes two essential mechanisms: advanced image contrast enhancement based on Contrast Limited Adaptive Histogram Equalization (CLAHE) and a hybrid deep learning model combining InceptionResNet and Transformer. Our novel framework is named LUNGINFORMER. The experiment report demonstrated LUNGINFORMER achieved an accuracy of 0.98, a recall of 0.97, an F1-score of 0.98, and a precision of 0.96. According to the AUC test, LUNGINFORMER achieved a tremendous performance with a score of 1.00 for each class. We believed that our performance model was influenced by contrast enhancement and a hybrid deep learning model
Privacy-Preserving U-Net Variants with pseudo-labeling for radiolucent lesion segmentation in dental CBCT
Full text linkAccurate segmentation of radiolucent lesions in dental Cone-Beam Computed Tomography (CBCT) is vital for enhancing diagnostic reliability and reducing the burden on clinicians. This study proposes a privacy-preserving segmentation framework leveraging multiple U-Net variantsβU-Net, DoubleU-Net, U2-Net, and Spatial Attention U-Net (SA-UNet)βto address challenges posed by limited labeled data and patient confidentiality concerns. To safeguard sensitive information, Differential Privacy Stochastic Gradient Descent (DP-SGD) is integrated using TensorFlow-Privacy, achieving a privacy budget of Ξ΅ β 1.5 with minimal performance degradation. Among the evaluated architectures, U2-Net demonstrates superior segmentation performance with a Dice coefficient of 0.833 and an Intersection over Union (IoU) of 0.881, showing less than 2% reduction under privacy constraints. To mitigate data annotation scarcity, a pseudo-labeling approach is implemented within an MLOps pipeline, enabling semi-supervised learning from unlabeled CBCT images. Over three iterative refinements, the pseudo-labeling strategy reduces validation loss by 14.4% and improves Dice score by 2.6%, demonstrating its effectiveness. Additionally, comparative evaluations reveal that SA-UNet offers competitive accuracy with faster inference time (22 ms per slice), making it suitable for low-resource deployments. The proposed approach presents a scalable and privacy-compliant framework for radiolucent lesion segmentation, supporting clinical decision-making in real-world dental imaging scenarios
Solar module defects classification using deep convolutional neural network
Full text linkSolar modules are essential components of a solar power plant, that are designed to withstand scorching heat, storms, strong winds, and other natural influences. However, continuous usage can cause defects in solar modules, preventing them from producing electrical energy optimally. This paper proposes the development of a deep learning-based system for identifying and classifying solar module surface defects in solar power plants. Module surface condition are classified into five categories: clean, dirt, burn, crack, and snail track. The dataset used consists of 8,370 images, including primary image data acquired directly from the mini solar power plant at the Renewable Energy Laboratory of PLN Institute of Technology, and secondary image data obtained from public repositories. The limitation in the number of images in each category was overcome using data augmentation techniques. The proposed classification model combines Deep Convolutional Neural Networks (DCNN) with transfer learning models (DenseNet201, MobileNetV2, and EfficientNetB0) to perform supervised image classification. Training and testing results on the three models demonstrated that the combination of DCNN + DenseNet201 provided the best performance, with a classification accuracy of 97.85%, compared to 97.25% accuracy for DCNN + EfficientNetB0 and 94.98% for DCNN + MobileNetV2. This research shows that DCNN-based image classification reliably diagnoses solar module defects and supports using RGB images for surface defect classification. Applying the developed system to solar power plant maintenance management can help in accelerating the process of identifying panel defects, determining defect types, and performing panel maintenance or repairs, while ensuring optimal power production
Multi-objective optimization algorithm for improving the efficiency of speeded up robust features of image stitching
No full textImage stitching to generate panoramic or composite images. This research proposes improved parameters of the fundamental matrix in the standard SURF method using multi-objective optimization techniques. This paper compares three metaheuristic algorithms (MOWOA, MOGWO, MOGA) and evaluates their performance using the hypervolume indicator (HV). The optimal points were chosen from non-dominated solutions using the MCDM and the minimization weighted sum methods (WSM). There were two objective functions: 1) minimum of image subtraction and 2) minimum of histogram. The MOWOA is superior to the other. This approach significantly reduces stitching errors and improves performance by 24.48% over standard SURF. The proposed multi-objective optimization of fundamental matrix parameters significantly enhances SURF-based image stitching by reducing alignment and blending errors, resulting in smoother and more coherent panoramic or composite images, by leveraging superior metaheuristic performance, particularly from MOWOA, which outperforms other algorithms. This approach increases stitching robustness and accuracy, making it highly valuable for real-world applications such as mapping, surveillance, and visual reconstructio
A two-layered collaborative approach for network intrusion detection system using blended shallow learning gaussian naΓ―ve bayes and support vector machine models
Full text linkThe majority of network intrusion detection systems use a signature matching technique. To detect abnormalities and unfamiliar attacks using machine learning methods is a more reliable approach. However, due to significant variations in attack trends, applying a single classifier is impractical for the effective detection of all types and forms of attacks, particularly rare attacks such as User2Root (U2R) and Remote2Local (R2L). Consequently, a hybrid strategy is expected to provide more promising performance. The proposed Two-Layered Collaborative Approach (TLCA) particularly addresses the problem as mentioned earlier. Principal Component Analysis optimizes variables to handle the variation resulting from every kind of attack. The proposed method investigates several types of attacks and discovered that the behaviors of U2R and R2L attacks are similar to those of regular usersβ characteristics. To identify DoS and Probe attacks, TLCA uses a Shallow Learning classifier, such as Gaussian NaΓ―ve Bayes, as Layer 1. Similarly, the Support Vector Machine at Layer 2 discriminates between U2R and R2L and typical occurrences. We have divided KDDTrain+ into Set 1 and Set 2 by observing that it involves two 2-dimensional PCA analyses. Cross-sectional Correlated Feature Selection (CCFS) is employed to choose key attributes. PCA and KPCA are applied to datasets to reduce dimensionality. The results obtained using the proposed method on the NSL-KDD dataset are compared with those of available benchmark methods. According to the experimental data, TLCA outperforms all single machine learning classifiers and surpasses many current cutting-edge IDS approaches. The proposed method achieves detection rates of 92.4%, 92.3%, 95.6%, and 100% for DoS, Probe, R2L, and U2R, respectively. The proposed TLCA also demonstrates a better ability to identify unusual attacks. It also yields improved detection rate results for known attacks, at 94.1%, and for unknown attacks, at 91.1%, when using the KDDTest+ dataset for testing
Performance analysis on convergence of particle swarm optimization and incremental conductance MPPT method for NTR 5E3E PV module
Full text linkParticle swarm optimization (PSO), a technique in Artificial Intelligence, is one of the MPPT methods used to optimize the output of a Photovoltaic (PV) system. The PSO is well known for its convergence in Maximum Power Point Tracking (MPPT). However, no comprehensive study has been conducted on the performance of the PSO and incremental-conductance (INC) MPPT combination for the NTR 5E3E PV module. This study aims to provide a detailed performance analysis of the convergence of PSO and INC combination compared to PSO MPPT during maximum power (MP) tracking on NTR 5E3E PV module. This research work studies the relationships among PV parameters and other parameters affected during the implementation of PSO-INC MPPT. The study found that, in terms of efficient power and time consumption during the Maximum Power (MP) tracking process, the PSO-INC MPPT combination provides the highest average peak power at the shortest time compared to standalone PSO. The efficiency of PSO-INC Average Power is near 98.9% to 99.93%, compared to PSO MPPT, which is between 95.7% and 99.3%. The PSO and INC MPPT were tested on a boost converter without altering the specific electrical component characteristics to ensure accurate output during testing. Furthermore, a boost converter is sufficient to meet the overall requirements for the research work and simulation testing. The characteristics of the PSO and INC MPPT are observed using MATLAB/Simulink. This research assesses the robustness of the PSO-INC combination, advancing hybrid MPPT technology by demonstrating its performance
Integrating hedge algebras and optimization techniques to reduce forecasting errors in fuzzy time series model
Full text linkAccurate forecasting in fuzzy time series (FTS) models is essential for applica-tions such as financial markets, traffic fatalities, and academic enrollments. How-ever, a persistent challenge in FTS forecasting is the determination of optimal interval lengths in the universe of discourse (UD), which significantly impacts prediction accuracy. This study introduces a novel hybrid approach that inte-grates Hedge Algebra (HA) with Particle Swarm Optimization (PSO) and Simu-lated Annealing (SA) to enhance forecasting accuracy. HA enables adaptive, non-uniform interval partitioning based on linguistic semantics, while PSO and SA jointly refine these intervals to reduce forecasting errors. Unlike convention-al FTS models with fixed partitioning, our approach leverages HAβs mathemati-cal structure alongside PSOβs global search and SAβs local refinement to en-hance adaptability and robustness. The model is evaluated on diverse datasets, including enrollment data, traffic fatalities, and gasoline prices, demonstrating superior forecasting accuracy over existing FTS models, as measured by Mean Squared Error (MSE) and Root Mean Squared Error (RMSE)