Diabetes mellitus (DM) remains a major global health challenge due to its increasing prevalence and long-term complications, emphasizing the need for accurate early prediction systems. This study proposes a machine learning-based framework for DM classification using a multi-dataset setting while addressing class imbalance issues. Two independent datasets from Iraq and Germany were employed to evaluate model robustness across different population characteristics. The experimental workflow consisted of data preprocessing, stratified train-test splitting, imbalance handling using Synthetic Minority Over-sampling Technique (SMOTE) and SMOTE-Tomek, 10-fold cross-validation, and hyperparameter optimization via GridSearchCV. Four classification algorithms were compared, namely Logistic Regression (LR), K-Nearest Neighbors (KNN), Random Forest (RF), and Support Vector Machine (SVM). Experimental results demonstrate that data distribution significantly affects classification performance. Under imbalanced conditions, RF achieved the best performance on the Iraqi dataset with an accuracy of 0.98 and an AUC of 1.00, while KNN and RF reached perfect accuracy (1.00) on the German dataset. After applying SMOTE, all models showed more stable performance, particularly in recall, which reached 1.00, indicating effective minority-class detection. In contrast, SMOTE-Tomek produced only marginal additional improvements. The findings suggest that no single classifier is universally optimal for DM prediction. Instead, model effectiveness depends on dataset characteristics and preprocessing strategies. From a practical perspective, the combination of RF and SMOTE shows strong potential for early diabetes screening and clinical decision-support systems. Further validation using larger and more heterogeneous external datasets is recommended.
Malnutrition remains a significant public health challenge in Indonesia, with early detection being crucial for effective intervention. Previous studies utilizing the K-Nearest Neighbor (KNN) algorithm demonstrated promising results in classifying malnourished toddlers based on anthropometric data. However, single-model approaches often suffer from sensitivity to noise and limited generalization. This study proposes a hybrid ensemble model combining KNN and Multi-Layer Perceptron (MLP), integrated with mutual information-based feature selection, to improve classification performance. Using a dataset from Puskesmas Ubung, Bali, comprising 1,319 records with nine anthropometric features and a binary malnutrition label, the model was evaluated under stratified five-fold cross-validation. The proposed KNN–MLP ensemble with top-ranked features achieved 94.3% accuracy, surpassing both standalone KNN and MLP models. Additional metrics, including precision (91.7%), recall (89.4%), F1-score (90.5%), and MAE (0.05), confirmed the model's robustness and reliability. These findings demonstrate that ensemble learning combined with feature selection significantly improves early-stage malnutrition classification, offering a scalable approach for decision-support systems in public health interventions.
Efficient urban waste management is a critical challenge driven by rapid urbanization, with collection routes strongly influencing operational costs and environmental sustainability. This study addresses the optimization of waste collection routes by modeling the problem as a Travelling Salesman Problem (TSP), serving as a foundational step toward more complex routing frameworks. We propose a Lévy-flight-enhanced Grey Wolf Optimizer (LGWO), which extends the standard Grey Wolf Optimizer (GWO) by integrating a lévy flight mechanism designed to strengthen global exploration and mitigate premature convergence to local optima. The performance of LGWO is evaluated against six other metaheuristic algorithms (GWO, ACOR, WOA, PSO, ALO, and ABC) using a real-world dataset of 36 waste collection points in Bandung, Indonesia. Experimental results based on 30 independent trials per algorithm show that LGWO achieves the best overall performance, obtaining the shortest tour (60.85 km) and the lowest mean distance (77.72 km), whereas the Ant Lion Optimizer (ALO) yields the poorest performance with the highest average distance of 89.90 km. These findings indicate that incorporating a lévy flight mechanism into GWO improves solution quality and convergence behavior for TSP-based waste collection routing. This research offers a practical optimization tool for developing more efficient and cost-effective urban waste management strategies. Future work will extend this approach by incorporating dynamic factors such as service times and vehicle capacities, enabling a more realistic treatment of Vehicle Routing Problem (VRP) variants.
This research develops a culinary recommendation system in Lombok by integrating the Latent Dirichlet Allocation (LDA) and Content-Based Filtering (CBF) methods. This integration aims to overcome the limitations of pure CBF, which relies solely on basic restaurant attributes and is less capable of capturing the semantic context of tourist reviews. Data was obtained through web scraping of Google Maps using the Apify.com platform, covering 825 restaurants and 20,114 reviews. The research stages included data collection, text preprocessing, topic modeling using LDA, feature engineering, similarity calculation using cosine similarity, and system evaluation. Evaluation was performed using Precision@K, Recall@K, F1-Score, and Mean Average Precision (MAP). The results show that the hybrid CBF+LDA model provides a significant improvement compared to pure CBF, with Precision@3 of 0.9333, Recall@3 of 0.1312, F1-Score of 0.2300, and MAP of 0.9628. These findings indicate that the integration of LDA topics enriches the semantic representation of reviews, thereby improving the relevance of recommendations. This research contributes to the development of artificial intelligence-based tourism recommendation systems and provides practical implications for promoting local cuisine, enhancing tourist experiences, and utilizing digital reviews as a basis for decision-making in the regional tourism sector.
Facial expressions are an important component of nonverbal communication that enable humans to understand each other's emotional states intuitively. Although humans can easily recognize expressions such as smiles or frowns, replicating this ability in computational systems remains a complex challenge. Therefore, an automated system capable of accurately and efficiently identifying facial expressions is needed. This research aims to compare the accuracy of CNN and SVM methods in facial expression recognition using the JAFFE dataset, which is limited to one demographic (Japanese women) with 284 images (80% for training, 10% for validation, and 10% for testing). CNN extracts features through convolution and pooling processes, while SVM is used as a classification algorithm based on statistical learning. The recognition process is divided into three main stages: data preprocessing, feature extraction, and facial expression classification. The system recognizes seven emotional categories: anger, disgust, fear, happiness, neutral, sadness, and surprise. Results show that CNN outperforms SVM with an accuracy of 86%, while SVM achieves 81%. The limitations of the dataset may affect generalizability, and further research can use larger, more diverse datasets
Gold is one of the most favored investment instruments due to its stability and its ability to preserve value against inflation. However, its price movements are volatile and influenced by various global economic factors, currency exchange rates, and geopolitical conditions, making gold price forecasting a significant challenge. This study aims to develop a gold price forecasting system using the Long Short-Term Memory (LSTM) algorithm, a variant of the Recurrent Neural Network (RNN) that excels in processing time-series data. The dataset consists of historical daily gold buying and selling prices from 2015 to 2025, collected from Yahoo Finance, Logam Mulia, and the official website of Bank Indonesia. The modeling process follows the CRISP-DM methodology, which includes business understanding, data preparation and exploration, modeling, and evaluation stages. Time Series Cross Validation (TSCV) is used to validate the model. LSTM performance is compared with other models such as GRU, CNN-1D, and Simple RNN to identify the best-performing architecture. Evaluation results indicate that LSTM achieved the highest performance with an R² score of 0.99 for selling prices and 0.98 for buying prices on the final test dataset. The system is deployed online, making it accessible in real-time. This research is expected to assist investors, financial analysts, and the general public in making smarter investment decisions based on valid historical data and advanced forecasting technology.
Acute Lymphoblastic Leukemia (ALL) is among the most common pediatric blood cancers and progresses rapidly, necessitating early and accurate detection. Manual diagnosis via microscopic analysis of blood samples is time-consuming and highly dependent on specialist expertise. This study proposes a hybrid model that combines a Convolutional Neural Network (CNN) with a Support Vector Machine (SVM) to automatically detect ALL from blood-cell images. The CNN performs deep feature extraction from images, while the SVM serves as the classifier to determine ALL status. The dataset comprises microscopic images labeled as ALL or normal and is processed through preprocessing steps such as augmentation and normalization. The adopted CNN produces optimized feature representations. Experimental results show that the hybrid CNN–SVM model with an RBF kernel achieves the best performance, with an accuracy of 96.4%, precision of 95.8%, recall of 96.1%, and an F1-score of 96.0%, surpassing pure CNN-based baselines. Training converged at the 41st epoch, with a training accuracy of 97.2%, validation accuracy of 95.9%, training loss of 0.09, and validation loss of 0.11, indicating stable learning without overfitting. The model’s ROC curve lies well above the chance diagonal, with an Area Under the Curve (AUC) of 0.914, means there is a 91.4% chance the model assigns a higher score to a truly positive (leukemia) image than to a negative (normal) image.These findings suggest that the CNN–SVM hybrid approach enhances leukemia detection performance compared with conventional CNN-only methods and holds promise as a fast, accurate, and efficient image-based decision-support tool for early leukemia diagnosis in digital hematology.
Recognizing infant cries is essential for healthcare, yet conventional representations such as spectrograms and MFCC often fail to capture temporal dynamics, limiting classification performance. This study introduces Polarized Amplitude Time Spiral Encoding (PATSE), a novel transformation that encodes amplitude and time into spiral-based polar representations, enabling richer visual features for deep learning. To address data scarcity and imbalance, audio augmentation techniques time stretching, time shifting, pitch scaling, and polarity inversion were applied, expanding the dataset from 457 to 6855 samples. A Convolutional Neural Network (CNN) trained on PATSE images achieved notable improvements, with overall accuracy increasing from 80% before augmentation to 93% after augmentation. The model attained high performance on the dominant Hungry class (F1-score = 0.96) while also enhancing recognition of minority classes such as belly pain, burping, discomfort, and tired. These results confirm the effectiveness of PATSE in improving generalization and reducing bias, offering a distinctive advantage over linear representations. The proposed framework provides a foundation for intelligent infant cry monitoring and early detection systems in healthcare.
Gastro-esophageal reflux disease (GERD) is a widespread condition that often leads to severe complications, including esophagitis, which significantly affects patient health and quality of life. While endoscopy is the gold standard for diagnosing esophagitis, its reliance on specialized equipment and trained professionals can limit accessibility and efficiency. This study introduces an innovative approach to diagnosing esophagitis by leveraging Convolutional Neural Networks (CNN) for automated classification of endoscopic images. By utilizing the Kvasir dataset, which includes a comprehensive collection of gastrointestinal endoscopy images, the model is trained to distinguish between esophagitis and normal-Z-line conditions with remarkable accuracy. The CNN model achieved outstanding results, with an accuracy of 96.04%, precision of 98.94%, recall of 93.00%, and an F1-score of 95.88%, demonstrating its potential to outperform traditional diagnostic methods. These findings underscore the ability of CNN to not only enhance diagnostic precision but also to reduce human error, making the process faster, more reliable, and more accessible. This research contributes to the growing body of work in medical image analysis, suggesting that CNN-based models can transform clinical practices by supporting timely, accurate diagnoses while alleviating the burden on medical professionals. The integration of deep learning in this domain holds the promise of advancing healthcare accessibility and efficiency globally
Spelling correction is an automatic correction feature useful in detecting spelling errors and providing word suggestions if necessary. Spelling correction is one of the crucial preprocessing phases in text mining. The Damerau-Levenshtein Distance method is one of the spelling correction methods that has high accuracy. This method has four types of operations: insertion, deletion, substitution, and transposition. The basic approach in detecting spelling errors in the Indonesian language is to use a dictionary search. Despite its accuracy, the Damerau-Levenshtein Distance method has a slow computation time. Furthermore, when the dictionary contains several suggested words that have the same distance from the target word, it will be difficult to prioritize the most appropriate suggestions. To overcome this problem, we introduce a caching mechanism to store previously calculated corrections, thereby speeding up the computation process. In addition, we use the cosine similarity method to rank words in Damerau-Levenshtein Distance results. The results of our approach have a significant improvement in accuracy, increasing from 72.13% to 83.60% by integrating caching and cosine similarity for ranking, which shows a significant improvement in both efficiency and effectiveness
This study proposes a deep learning model using the ConvNeXt Tiny architecture to detect autism spectrum disorder (ASD) from facial images, addressing the need for an early, efficient, and accessible diagnostic tool. The model integrates facial image preprocessing techniques like Contrast Limited Adaptive Histogram Equalization (CLAHE) and data augmentation, with facial segmentation performed by MTCNN. The ConvNeXt Tiny model is trained using transfer learning and evaluated through metrics such as accuracy, precision, recall, and F1-score, and compared with traditional CNN models like ResNet50 and EfficientNet-B0. The results demonstrate that the proposed model outperforms ResNet50 and EfficientNet in all evaluation metrics, achieving a classification accuracy of 84%. It also demonstrates a balanced performance across both classes (autistic and non-autistic), with high precision and recall for both, leading to a high F1-score. Furthermore, the model's computational efficiency makes it suitable for web and mobile applications, enabling scalable and real-time screening for ASD in children. The study's contributions include the development of a novel, lightweight ASD classification system, a comparative analysis of ConvNeXt with other CNN models, and the creation of a prototype for early ASD detection. This approach not only provides a promising alternative to conventional diagnostic methods but also sets the groundwork for further research and practical implementation in clinical settings.
This research evaluates the performance of the drone detection system based on YOLOv5 in a variety of environmental conditions. The four main variables under test were drone height, camera type, light intensity, and camera-to-object distance. Thirty-six different scenarios were used with three different camera types (1080p, 2K, and Canon 600D). The height of the drones varied from 1 to 14 meters, and the variations in illumination ranged from 0 to 46 lux. Results showed consistent YOLOv5 performance with an average accuracy of 60%, precision of 62%, recall of 58%, F1-score of 60%, and IoU of 75%. ANOVA revealed that light intensity, camera distance, and drone height all had a significant impact on detection accuracy (p < 0.05), but camera type was not statistically significant. The best results were obtained under the following conditions: high light levels (>40 lux), camera distances <10 m, and drone altitudes between 6 and 9 m. These findings demonstrate the importance of environmental setup in improving the performance of object detection systems based on deep learning. This research helps design a more reliable and adaptable drone detection system for real-world applications. This work provides practical guidelines for implementing deep learning-based aerial surveillance and highlights optimal operational parameters for YOLOv5 systems.
The use of computer technology in the healthcare sector is growing, especially in supporting recommendation systems and early detection of various health conditions. Artificial intelligence, particularly deep learning, has made significant contributions in analysing complex data such as medical imaging. One of the leading deep learning methods is Convolutional Neural Network (CNN), which is able to extract visual features hierarchically and accurately. Baby blues is a psychological disorder often experienced by mothers after childbirth and can have a serious impact on the mother's mental health and relationship with the baby. Early detection of baby blues is crucial to provide appropriate interventions and prevent worse outcomes. This research aims to implement CNN algorithm to detect baby blues through facial image analysis. Using a dataset of postpartum mothers, a CNN model was developed to recognise visual patterns related to baby blues symptoms. The results showed that the CNN model was able to identify baby blues conditions with an accuracy of 53% on the dataset used. This research proves the effectiveness of CNN in detecting visual patterns related to babyblues disorder, and is expected to be a solution in supporting early diagnosis and appropriate treatment for postpartum mothers.
The Indonesia Smart College Card (Kartu Indonesia Pintar-Kuliah / KIP-K) is a national initiative aimed at expanding access to higher education for students from socioeconomically disadvantaged backgrounds. This study, conducted at Politeknik Negeri Cilacap, investigates the prediction of KIP-K acceptance based on the type of high school attended by applicants. A comparative analysis was carried out using three supervised machine learning algorithms: Extreme Gradient Boosting (XGBoost), Random Forest, and Support Vector Machine with Radial Basis Function (SVM-RBF). The dataset, sourced from institutional admission records between 2022 and 2024, comprises information on school types (public, private, vocational, madrasah, and others), demographic attributes, and the KIP-K acceptance status. The data were split into training and testing sets using a 50:50 stratified sampling technique to preserve class distribution. Model performance was evaluated using standard classification metrics, including accuracy, precision, recall, and F1-score. Additionally, confusion matrices, ROC curves, and feature importance visualizations were used to enhance model interpretability. The experimental results demonstrate that the XGBoost algorithm consistently outperformed the other models across all performance metrics. Specifically, XGBoost exhibited the highest discriminatory power with an AUC of 0.93, followed by Random Forest (0.90) and SVM-RBF (0.85). These findings affirm the suitability of tree-based ensemble methods for classification tasks in educational domains and emphasize the predictive relevance of school type in determining KIP-K eligibility. The study presents a data-driven decision support framework that can contribute to more objective, transparent, and equitable scholarship allocation practices, particularly within the context of vocational higher education institutions in Indonesia
This study focuses on optimizing the performance of a three-modality biometric recognition system (fingerprints, facial and voice recognition) with global decision fusion, designed for access control to secure areas. When the biometric database contains a large volume of information, the verification module's processing time increases considerably due to the complexity of template comparisons. To address this issue, we implemented an optimization strategy based on parallel programming, specifically targeting the intensive processing loops within the verification module. Using Microsoft's Task Parallel Library, we parallelized all critical loops associated with the three biometric modalities. By effectively exploiting for and foreach statements, our parallelized implementation enables optimal distribution of tasks across available processor cores. We validated our approach by conducting repeated experiments on data sets of varying sizes (50 to 600 individuals), with a rigorous analysis of temporal performance. The results show a significant reduction in execution times: for 600 entries, the processing time goes from 1.68 ms in sequential mode to 0.77 ms in parallel mode. These performances were evaluated over several iterations to ensure the statistical reliability of the results, in particular by calculating averages and standard deviations and including error bars in the comparative graphs. The practical implications of this work are significant: the module can be deployed in corporate security systems, airports or banks, while respecting ethical considerations and privacy constraints. Finally, this work paves the way for future extensions, including the integration of other biometric modalities, deployment on distributed clusters or the adoption of more advanced parallelization frameworks.
This study aims to design and implement an intelligent traffic light system for one-way open and close road conditions, commonly encountered during road repair projects. These situations often cause congestion due to alternating vehicle flow in a single lane. To address this issue, the system utilizes a Reinforcement Learning (RL) algorithm to dynamically adjust the traffic light timing based on real-time traffic conditions. The research was conducted in three main stages: (1) designing the network topology and IoT devices using Raspberry Pi, ESP modules, and Access Points (APs), (2) implementing the intelligent traffic light system, and (3) conducting a functional evaluation. A key performance metric evaluated was the response time of the system. Experimental results showed that the traffic light system achieved an average response time of 0.51 seconds, indicating that it is responsive and suitable for real-time operation. The successful integration of RL and MQTT-based communication also demonstrates the feasibility of deploying this system in dynamic traffic environments. Further research is recommended for field testing with additional sensor integration and advanced RL models to enhance system accuracy and efficiency
This study analyzes factors influencing students’ final project completion status in a higher education context using six classification models: C4.5, Random Forest (RF), C4.5 with SMOTE, RF with SMOTE, Cost-Sensitive Random Forest (RF-CS), and Cost-Sensitive C4.5 (C4.5-CS). The dataset consists of 1,017 student records categorized into Ideal and Tidak Ideal, with a severe class imbalance where the minority class represents only 16.49% of the data.The results indicate that baseline models achieved high overall accuracy but showed limited effectiveness in identifying the minority Tidak Ideal class. SMOTE-based models improved minority-class recall but introduced a higher number of false positives, highlighting a trade-off between recall and precision. In contrast, cost-sensitive learning produced the most substantial improvement in minority-class detection. Among all evaluated models, Cost-Sensitive Random Forest demonstrated the most balanced performance by significantly reducing false-negative errors while maintaining reasonable overall accuracy.These findings confirm that algorithm-level cost-sensitive approaches are more effective than oversampling techniques for handling severe class imbalance in educational datasets. The proposed model provides a reliable basis for early identification of students at risk of delayed final project completion and supports data-driven academic decision-making
This study aims to compare the performance between a PID cascade controller and a self-tuning Neural Network adaptive control (PID NN) in controlling an ROV on an octarotor platform with 6 degrees of freedom (DOF), namely Surge, Sway, Heave, Roll, Pitch, and Yaw. The conventional PID control system is used as a proven baseline, while the Neural Network-based adaptive control is applied to adjust parameters in real-time, expected to address the nonlinearity and external disturbances that are difficult to handle by static PID. This study involves an analysis of overshoot, rise time, settling time, and final position for each channel, as well as a comparison of the performance of the two control methods. The results show that PID NN provides faster rise times and lower overshoot in most channels, such as Surge overshoot of 8.3%, rise time of 6.2 seconds, and Sway overshoot of 2.1%, settling time of 90 seconds, compared to PID, which has higher overshoot and longer stabilization times. However, for the Yaw and Heave channels, although PID NN showed larger overshoot and longer settling time, PID was faster in achieving stability. Overall, although PID NN demonstrated superiority in terms of rapid stabilization for Roll and Pitch, further adjustments are needed to optimize Yaw and Heave to achieve faster stabilization without compromising system stability and overall control performance. This study opens opportunities for further development in the field of adaptive control for high-complexity multirotor systems.
Virtual YouTuber (VTuber) has a feature that allows fan and viewer interaction through live streaming chat that contains textual data based on emotions and opinions. The previous study examined sentiment analysis in various domains. However, live-streaming chat has short, informal, and unstructured text characteristics, making it challenging to analyze its sentiment. Decision Trees (DT) have advantages in interpretability and training speed, while Support Vector Machines (SVM) can handle high-dimensional data and avoid overfitting. Still, few studies examine DT and SVM in live streaming chat. This study aims to propose a sentiment analysis model in VTuber live streaming chat by comparing the performance of DT and SVM. VTuber Lives streaming chat was collected and preprocessed through cleaning, expansion-contraction, case folding, tokenization, stopword removal, and lemmatization. VADER and AFINN Lexicon labeled positive, neutral, or negative sentiments. Later, TF-IDF is used for feature extraction, and K-Fold cross-validation is used to evaluate the sentiment analysis model based on DT and SVM. A confusion matrix measures the model’s performance by knowing the accuracy, precision, recall, and F1 score values. The results of 10-fold cross-validation show that the proposed model with a combination of DT+AFINN with hyperparameter optimization achieves an accuracy of 96.26%. The combination of DT+AFINN shows its superiority in sentiment analysis of VTuber live chat data compared to DT+VADER, SVM+AFINN, and SVM+VADER.
Indonesia is an agricultural country whose population lives from the agricultural sector, the Government has encouraged agriculture to realize the use of modern technology for the agricultural sector which is useful for increasing farmer productivity and making farmers' work easier with quality harvests, as well as more efficient food security in Indonesia. One way to save time and energy for farmers to find out nutrients and pH is by using a nutrient and pH monitoring system for hydroponic plant water. The purpose of this research is to create an IoT-based automatic system, used for systematic real-time monitoring of water nutrients in hydroponic plants using a TDS sensor to measure the mass weight of nutrients and a pH meter to measure the acidity of water, a DHT22 sensor to measure temperature and humidity, and a DSB18B20 sensor to measure water temperature, the data obtained from these sensors will be displayed on the LCD. The final result of the tool created is a nutrient and pH monitoring system for water in pakcoy plants, the results of sensor readings will be sent in real-time to the dashboard and WhatsApp. This research method uses the R and D (Research and Development) method with the ADDIE (analyze, design, develop, implement, evaluate) model. The IoT expert validation value was 0.91 (very appropriate) and the farmer effectiveness validation value was 0.85 (very effective). The outputs generated from this research can be used to make it easier for hydroponic farmers to monitor plant nutrient balance automatically and systematically.
