Jurnal RESTI (Rekayasa Sistem dan Teknologi Informasi)
Not a member yet
1071 research outputs found
Sort by
CNN-Based Skin Cancer Classification with Combined Max and Global Average Pooling
No full textSkin cancer is one of the most threatening diseases to human health, with an increase in new cases each year. Early detection plays a crucial role in improving recovery rates, however, conventional diagnostic methods such as biopsy are often invasive, time-consuming, and costly. To address this issue, artificial intelligence-based diagnostic systems, particularly Convolutional Neural Networks (CNNs), offer a promising solution for enhancing diagnostic accuracy and efficiency. This study aims to evaluate the performance of a CNN model that combines Max Pooling and Global Average Pooling (GAP) in detecting skin cancer from digital dermoscopic images. The ISIC (International Skin Imaging Collaboration) dataset was used, focusing on two classes: malignant and benign. The combination of Max Pooling and GAP is intended to increase model precision while reducing the risk of overfitting. The experimental results show that the proposed model achieved a precision of 96.35%, indicating strong performance in minimizing false positives. However, the recall was relatively low at 85.99%, suggesting reduced sensitivity in detecting malignant cases. The overall accuracy of the combined model was 91.68%, slightly lower than the Max Pooling-only model (91.79%). Although the combination does not significantly improve accuracy, it effectively enhances precision to 96.35%. This is a critical advantage in a clinical setting, as it directly translates to minimizing false positive diagnoses and preventing patients from undergoing unnecessary invasive procedures like biopsies
Explainable Ensemble Learning for Maternal Health Risk in Low-Resource Settings
No full textMaternal health remains a global challenge, particularly in low-resource settings where accurate and timely risk prediction is essential to reducing maternal mortality. This study proposes an explainable machine learning framework for predicting maternal health risks by integrating ensemble learning methods with SHAP (Shapley Additive exPlanations) for interpretability. This study utilized the publicly available Maternal Health Risk Data Set (MHRDS), comprising physiological features such as systolic and diastolic blood pressure, blood sugar level, body temperature, and age. A total of 18 machine learning models including Random Forest, XGBoost, LightGBM, Neural Networks, and TabNet were evaluated to compare individual classifiers and ensemble approaches comprehensively. The selection of this diverse set of models is grounded in the need to benchmark different algorithmic paradigms, as variations in inductive bias, learning capacity, and robustness to clinical data noise can influence predictive performance and generalizability. This comprehensive comparison enables the identification of optimal model types for integration into ensemble frameworks. Evaluation was performed across three different test scenarios (test sizes of 10%, 20%, and 30%) to assess model consistency under varying data partitions. Stacking, Voting, and Histogram-based Gradient Boosting showed consistently high performance, with Stacking achieving the highest accuracy of 87.2%, followed by Histogram Gradient Boosting (86.9%) and Voting (86.7%) at test size 0.2. SHAP analysis identified blood sugar, systolic blood pressure, and maternal age as the top predictors across all test scenarios. The best-performing models were deployed into a web-based clinical decision support system designed for healthcare practitioners in Indonesia. The proposed approach balances predictive accuracy and model transparency, offering a practical solution for improving maternal care in data-limited environments
Addressing Class Imbalance in Oil Palm Disease and Micronutrient Deficiency Detection Using Meta-Learned Transfer Metric Learning
No full textClass imbalance is a major challenge in oil palm disease and nutrient deficiency detection, where healthy samples dominate while diseased or deficient cases are underrepresented, often leading to biased models with high false-negative rates. To address this issue, this study proposes MetaTMLDA (Meta-Learned Transfer Metric Learning with Distribution Alignment), a hybrid framework that combines Transfer Metric Learning (TML) with MW-FixMatch. TML learns discriminative and domain-invariant features, while MW-FixMatch employs a meta-learned weighting mechanism to adaptively reweight samples, improving sensitivity to minority classes and enhancing robustness against pseudo-label noise. Experiments on four public datasets—Ganoderma Disease Detection, Palm Oil Leaf Disease, and Leaf Nutrient Detection for Boron and Magnesium—demonstrated that the proposed method consistently outperforms TML-DA, MW-FixMatch, SMOTE, Random Undersampling, and Biased SVM. On the smaller datasets (Ganoderma and Palm Oil Leaf Disease), MetaTMLDA achieved accuracy of 0.976, precision 0.951, recall 0.915, Cohen’s Kappa 0.912, and macro F1-score 0.933 for Ganoderma, and accuracy of 0.980, precision 0.972, recall 0.957, Kappa 0.911, and macro F1-score 0.964 for Palm Oil Leaf Disease. On the larger datasets (Boron and Magnesium), the model reached near-perfect accuracy of 0.995, with precision up to 0.967, recall up to 0.973, Kappa above 0.919, and macro F1-scores up to 0.969, highlighting its robustness and balanced predictive performance. These findings confirm that MetaTMLDA effectively addresses both class imbalance and domain shift, providing a scalable solution for precision agriculture through earlier and more reliable detection of oil palm health issues
RadReader: An Enhanced AlexNet-Based GUI Application for Pneumonia Prediction in Thoracic X-Ray Images
No full textRecent advancements in radiology applications have led to user-friendly interfaces, improving pneumonia diagnosis by accurately differentiating between viral and bacterial pneumonia from thoracic X-rays. This approach enhances diagnostic precision and efficiency while offering intuitive real-time interaction for radiologists. This study aims to achieve two objectives: (i) developing a desktop-based radiology reader application, and (ii) modifying the alexNet architecture for classifying pneumonia based on thoracic X-ray datasets with the output encompassing pneumonia and normal cases. The desktop application assists radiologists in efficient image analysis and is developed using python–Tkinter. Integrate enhanced of AlexNet models which has been modified to better differentiate. The modified alexNet includes changes like adding max pooling in specific blocks and adjusting hidden layer neuron count. The dataset consists of 7442 images, with 4484 positive pneumonia and 2958 normal images obtained from the Mendeley websites. The enhanced alexNet (EAM) model achieves impressive results: 95.36% accuracy, 95.34% precision, 95.28% recall, and 95.31% F1-score for classifying bacterial pneumonia
Enhancing Network Security: Evaluating SDN-Enabled Firewall Solutions and Clustering Analysis Using K-Means through Data-Driven Insights
No full textIn the face of escalating and increasingly complex cyber threats, enhancing network security has become a critical challenge. This study addresses this issue by investigating the optimization of SDN-enabled firewall solutions using a data-driven approach. The research employs K-Means clustering to analyze attack patterns, aiming to identify and understand distinct patterns for improved firewall effectiveness. Through the clustering process, attack data was classified into three clusters: Cluster 0, indicating concentrated attack sources likely tied to high-activity regions or networks; Cluster 1, representing a dispersed distribution of attacks, pointing to diverse origins; and Cluster 2, linked to specific geographic regions or unique attack behaviors. The clustering efficacy was evaluated using the Silhouette Score (0.606) and the Davies-Bouldin Index (0.614), indicating meaningful and reliable clustering outcomes. These findings provide actionable insights into network threat patterns, enabling the refinement and enhancement of SDN-enabled firewalls. The study contributes to the field by demonstrating the potential of clustering techniques in uncovering patterns overlooked by traditional methods and paving the way for further research into alternative clustering algorithms and broader applications in network security
Prediction of Main Transportation Modes using Passive Mobile Positioning Data (Passive MPD)
No full textIndicators of the main mode of transportation used by domestic tourists during tourism trips cannot yet be estimated using Passive MPD which is recorded based on the location of the BTS that captures the cellular activity of domestic tourists. Previous research on identifying transportation modes from Passive MPD has its own shortcomings because it only relies on speed and travel time features. Meanwhile, there is Active MPD which is recorded using active geo-positioning and real-time, where the research involves many features and has a data structure similar to Passive MPD. Therefore, this research aims to conduct a study of the implementation of the method used to identify modes of transportation in Active MPDs to Passive MPDs as an approach to predicting the main modes of transportation. As a result, the transportation mode identification method in the Active MPD can be implemented in the Passive MPD. The best accuracy of 83.56% was obtained by the LightGBM model using all features. However, the Multinomial Logistic Regression model, which only uses 10 selected features, is the most effective and efficient model with an accuracy of 76.43% and a much shorter execution tim
PCA and t-SNE Implementation for KNN Hypertension Classification Visualization
No full textHypertension is a condition that, if allowed to increase, can significantly injure internal organs due to high blood pressure. The objective of this study is to use the K-Nearest Neighbor (KNN) algorithm along with PCA and t-SNE to accurately identify four categories of Hypertension, Normal, Hypertension, Stage 1 Hypertension, and Stage 2 Hypertension. After establishing the scope, a dataset consisting of 7,794 samples was sourced from Labuang Baji Regional General Hospital, Makassar, and contained age, weight, and systolic and diastolic blood pressure parameters. The class distribution is Normal (36.3%), Hypertension (43.12%), Stage 1 Hypertension (8.29%), and Stage 2 Hypertension (12.31%). Experimental results show that the KNN base model achieved 99% accuracy, KNN with PCA reached 100%, and KNN with t-SNE attained 99%. Cross-validation was used to evaluate model generalization, yielding accuracies of 91%, 94%, and 91%, respectively. These findings suggest that KNN, particularly when integrated with t-SNE, is highly effective in visualizing and classifying non-linear data structures. Furthermore, this study demonstrates that incorporating dimensionality reduction techniques enhances the interpretability of classified hypertension data, which is crucial for informed decision-making by mental health committees
Detecting Alzheimer's Based on MRI Medical Images by Using External Attention Transformer
No full textAlzheimer's disease is one of the major challenges in medical care this century, affecting millions of people worldwide. Alzheimer's damages neurons and connections in brain areas responsible for memory, language, reasoning, and social behavior. Early detection of this disease enables more effective treatment and proper care planning. Unfortunately, the traditional method of detecting Alzheimer's has several limitations, such as subjective analysis and delayed diagnosis. One commonly used method is visual inspection, which uses magnetic resonance imaging (MRI). The limitations of visual inspection include subjectivity and its time-consuming nature, especially with large or complex MRI datasets, making accurate interpretation a significant challenge. Therefore, an alternative for detecting Alzheimer’s disease is to use deep learning-based MRI image analysis. One promising approach is to implement the External Attention Transformer (EAT) model. It enhances image classification by using two shared external memories and an attention mechanism that filters out redundant information for improved performance and efficiency. The aim of this research is to evaluate and compare the performance of the baseline Convolutional Neural Network (CNN) model, the Vision Transformer (ViT) model, and the EAT model in detecting Alzheimer's using a dataset of 6400 brain MRI images. The EAT model outperforms the baseline CNN model and ViT model in detecting Alzheimer's, achieving its best results with an accuracy of 0.965 and an F1-score of 0.747 for the test data. Our results could be integrated with clinical analysis to assist in the faster diagnosis of Alzheimer's
Classification Model for Bot-IoT Attack Detection Using Correlation and Analysis of Variance
No full textIndustry 4.0 requires secure networks as the advancements in IoT and AI exacerbate the challenges and vulnerabilities in data security. This research focuses on detecting Bot-IoT activity using the Bot-IoT UNSW Canberra 2018 dataset. The dataset initially showed a significant imbalance, with 2,934,447 entries of attack activity and only 370 entries of normal activity. To address this imbalance, an innovative data aggregation technique was applied, effectively reducing similar patterns and trends. This approach resulted in a balanced dataset consisting of 8 attack activity points and 80 normal activity points. Feature selection using the ANOVA method identified 10 key features from a total of 17: seq, stddev, N_IN_Conn_P_SrcIP, min, state_number, mean, N_IN_Conn_P_DstIP, drate, srate, and max. The classification process utilized Random Forest, k-NN, Naïve Bayes, and Decision Tree algorithms, with 100 iterations and an 80:20 training-testing split. Random Forest showed superior performance, achieving 97.5% accuracy, 97.4% precision, and 97.4% recall, with a total computation time of 11.54 seconds. Pearson correlation analysis revealed a strong positive correlation (+0.937) between N_IN_Conn_P_DstIP and seq, as well as a weak negative correlation (-0.224) between N_IN_Conn_P_SrcIP and state_number. The novelty of this research lies in the application of a data aggregation technique to address class imbalance, significantly improving machine learning model performance and optimizing training time. These findings contribute to the development of robust cybersecurity systems to effectively detect IoT-related threats
Agricultural Cultivation Cost Prediction Using Neural Networks and Feature Importance Analysis
No full textAgriculture is one of the most important sectors integral to human civilization, and technological adaptation is necessary to maintain its quality. This research aims to achieve high productivity in the agricultural sector by using neural networks or Deep Learning methods to predict the cost of agricultural cultivation, as well as identifying significant factors that affect the profitability of potato commodities with Feature Importance analysis. The research process includes the stages of Data Preparation, Data Understanding, Split Data Training, Classification Model Building, Training, and Evaluation. Evaluation techniques such as MAE, MSE, and R² were used to assess the effectiveness of the model. The results showed that the prediction model almost achieved optimal performance, with the Cost of Cultivation C2 factor having the greatest influence in understanding the data and guiding improvements to the significant factors affecting cultivation cost prediction. The main contribution of this research is the application of optimal Deep Learning methods to predict the cost of cultivation as well as identifying the main components that impact the profitability of potato farming in India