Jurnal Politeknik Negeri Batam (PoliBatam)
Not a member yet
3001 research outputs found
Sort by
Analysis of Deep Learning Implementation Using Xception for Rice Leaf Disease Classification
No full textIdentifying rice leaf diseases plays a crucial role in maintaining agricultural productivity and preventing massive losses. In recent years, deep learning models have shown very promising performance in plant disease classification tasks. This study proposes a Rice Leaf Disease Detection System based on the Xception model from Keras Applications, an architecture that is still relatively unexplored for rice plant disease cases. Through preprocessing, data augmentation, and model refinement, the developed system achieved a training accuracy of 93% and a testing accuracy of 89% in classifying rice leaf conditions. In addition, metric evaluation showed precision, recall, and F1-score values of 89%, reflecting the model\u27s ability to make consistent and balanced predictions. The trained model was then integrated into a web-based application to facilitate real-time disease diagnosis through image uploads. The results of this study prove the effectiveness of the Xception architecture in extracting agricultural image features and its potential for application in artificial intelligence-based smart farming systems.Identifying rice leaf diseases plays a crucial role in maintaining agricultural productivity and preventing massive losses. In recent years, deep learning models have shown very promising performance in plant disease classification tasks. This study proposes a Rice Leaf Disease Detection System based on the Xception model from Keras Applications, an architecture that is still relatively unexplored for rice plant disease cases. Through preprocessing, data augmentation, and model refinement, the developed system achieved a training accuracy of 93% and a testing accuracy of 89% in classifying rice leaf conditions. In addition, metric evaluation showed precision, recall, and F1-score values of 89%, reflecting the model\u27s ability to make consistent and balanced predictions. The trained model was then integrated into a web-based application to facilitate real-time disease diagnosis through image uploads. The results of this study prove the effectiveness of the Xception architecture in extracting agricultural image features and its potential for application in artificial intelligence-based smart farming systems
Optimizing Bankruptcy Prediction on Imbalanced Data using XGBoost with Random Oversampling and Chi-Square
No full textIn the midst of modern financial dynamics, the ability to predict corporate bankruptcy holds strategic significance, as it directly affects economic stability and investor confidence. However, the development of a reliable predictive model is often hindered by the complex nature of financial data, particularly the class imbalance between bankrupt and non-bankrupt companies. This imbalance causes models to become biased toward the majority class, thereby reducing their sensitivity in detecting bankruptcy cases which are, in fact, the most critical for financial decision-making. This research aims to construct a more balanced and sensitive bankruptcy prediction model by specifically addressing the issue of data imbalance. The proposed approach integrates the Random Oversampling (ROS) technique to equalize class distribution, Chi-Square feature selection to identify the most informative financial variables, and the Extreme Gradient Boosting (XGBoost) algorithm as the core predictive model. The dataset used is the UCI Taiwanese Bankruptcy Prediction dataset, consisting of 6,819 observations and 96 financial ratio variables. Experimental results show that the Chi-Square method successfully identified 20 influential variables, including Per Share Net Profit Before, Debt Ratio, and ROA(B) Before Interest and Depreciation After Tax. The proposed XGBoost model achieved an overall accuracy of 0.9648 and an F1-score of 0.4286, demonstrating superior performance. These findings confirm that the combination of ROS, Chi-Square, and XGBoost effectively enhances data balance and prediction sensitivity for the bankruptcy class. This research is expected to serve as a foundation for developing financial decision-support systems capable of providing early warnings of potential corporate bankruptcy
Transformer-Based Models for Electronic Health Records and Omics in Healthcare: A Systematic Literature Review
No full textElectronic Health Records (EHRs) have become central to modern healthcare. The emergence of transformer-based models has profoundly influenced how EHRs are used for modelling complex, longitudinal data. Integration with omics technologies improves the precision of disease identification and risk assessment during modelling. While several reviews have examined transformers in healthcare broadly, a systematic synthesis focused on their architectural design, empirical performance and integration of EHRs with omics data remains limited. This study presents a systematic literature review of transformer-based models applied to electronic health records (EHRs) and omics data, and of their integration into healthcare. Following PRISMA guidelines, peer-reviewed studies were retrieved from IEEE Xplore, ACM Digital Library, PubMed, and ScienceDirect, resulting in 14 eligible empirical studies published between 2020 and 2025. The review analyses transformer architectures, submodules, application domains, comparative performance, interpretability mechanisms, and limitations. Findings indicate that architectural design drives task-specific advantages in disease prediction, phenotyping, medication recommendation, and omics analysis. The integration of self-attention with deep learning, temporal modelling, and a pre-trained biomedical transformer improves performance. However, most studies remain centred on EHR, with limited empirical integration of omics data. Persistent challenges include limited generalisability, high computational cost, data quality issues, and insufficient interpretability for clinical deployment. The primary contribution of this review lies in synthesising architectural trends and methodological gaps. By consolidating current evidence, the study provides clear directions for the development of explainable, generalisable, and multimodal transformer-based systems in precision healthcare
AI-YOLO Based Smart Laboratory Security for Automated Face Recognition and Suspicious Activity Detection
No full textEnsuring laboratory security is a critical consideration within campus environments to effectively prevent theft and suspicious activities. Traditional CCTV systems predominantly rely on manual monitoring, resulting in delayed responses to incidents. This research seeks to develop and implement an Artificial Intelligence (AI)-based laboratory security system, integrating three primary models: YOLOv5 for human object detection, Face Recognition for individual identification, and Media Pipe Pose for real-time analysis of suspicious movements. The system is designed as a Flask-based monitoring website, which displays activity logs, detected individual identities, and automated notifications based on image processing results on a Raspberry Pi connected to CCTV cameras. The research methodology employs an applied experimental approach, encompassing stages such as system design, face dataset collection, data encoding utilizing the Face Recognition Library, and performance evaluation under two lighting conditions (bright and dark) and three distance variations. The test results indicate that the Face Recognition method operates optimally at a distance of 2 meters in bright lighting conditions, achieving an accuracy of up to 92%. However, its performance declines at distances exceeding 3 meters and under low-light conditions. Conversely, MediaPipe Pose exhibits high stability, with an average accuracy of 94% in bright conditions and 89% in dark conditions, and is capable of transmitting real-time notifications for activities such as lifting objects or placing hands into pockets. The AI-based laboratory security system developed has demonstrated effectiveness, adaptability, and responsiveness in the automatic detection of identities and suspicious activities. The integration of YOLO v5, Face Recognition, and MediaPipe Pose models offers an intelligent and efficient security solution that facilitates the implementation of the Smart Campus concept within educational environments
Analysis of the Determinants of Pelni Mobile Adoption Failure in Manokwari: A Hybrid Diffusion of Innovation and Theory of Planned Behaviour Approach
No full textThe adoption of digital services like Pelni Mobile in developing regions faces complex challenges. Despite offering ease of access, its adoption rate in Manokwari Regency remains low. Previous studies have not extensively explored typical barriers such as resistance to change, perceived financial costs, inconvenience, and ease of access. This study analyzes the factors behind Pelni Mobile\u27s adoption failure by integrating the DOI and TPB approaches. Data were collected via online questionnaires from 435 participants and analyzed using SEM-PLS. Findings show that Perceived Financial Cost (P=0.000), Resistance to Change (P=0.000), and Inconvenience (P=0.000) have a significant negative influence on Behavioral Intention to Use. This means perceived costs, resistance to change, and inconvenience can reduce usage interest. Conversely, Perceived Ubiquity (P=0.000) has a significant positive influence on usage intention, and Behavioral Intention to Use positively influences Use Behavior, indicating that ease of access can encourage adoption.The implications highlight the need for strategies to reduce financial barriers, improve accessibility, employ educational approaches to address resistance, and enhance user experience. For developers and policymakers, these results serve as a guide for designing more inclusive digital services tailored to the characteristics of developing communities, particularly in contexts similar to Manokwari. Generalizing the findings to other regions must consider local social, economic, and cultural differences.The adoption of digital services like Pelni Mobile in developing regions faces complex challenges. Despite offering ease of access, its adoption rate in Manokwari Regency remains low. Previous studies have not extensively explored typical barriers such as resistance to change, perceived financial costs, inconvenience, and ease of access. This study analyzes the factors behind Pelni Mobile\u27s adoption failure by integrating the DOI and TPB approaches. Data were collected via online questionnaires from 435 participants and analyzed using SEM-PLS. Findings show that Perceived Financial Cost (P=0.000), Resistance to Change (P=0.000), and Inconvenience (P=0.000) have a significant negative influence on Behavioral Intention to Use. This means perceived costs, resistance to change, and inconvenience can reduce usage interest. Conversely, Perceived Ubiquity (P=0.000) has a significant positive influence on usage intention, and Behavioral Intention to Use positively influences Use Behavior, indicating that ease of access can encourage adoption.The implications highlight the need for strategies to reduce financial barriers, improve accessibility, employ educational approaches to address resistance, and enhance user experience. For developers and policymakers, these results serve as a guide for designing more inclusive digital services tailored to the characteristics of developing communities, particularly in contexts similar to Manokwari. Generalizing the findings to other regions must consider local social, economic, and cultural differences
Behavioural Predictors of Forward Head Posture Risk: A Correlation, Machine Learning, and Clustering Analysis
No full textForward Head Posture (FHP) has become increasingly common among university students due to prolonged digital device use and inadequate ergonomic behaviour. This study aims to identify the behavioural factors that most strongly predict neck tension, which is used as an indicator of FHP risk, among laptop users at Universitas Ciputra. A total of 141 survey responses were collected, capturing digital lifestyle patterns that include screen exposure, posture habits, ergonomic awareness, physical activity, and screen-related symptoms. The analysis followed a complete methodological sequence that involved data preprocessing, correlation testing, supervised machine-learning modelling, and K-Means clustering. The results show that headache after screen use, frequency of head-down posture, ergonomic knowledge, and weekly exercise emerged as the most influential behavioural predictors of neck tension, with head-down posture demonstrating the strongest association (r = 0.437). Correlation testing supported three of the four hypotheses, while the Random Forest model achieved the highest predictive performance (71.01% cross-validated accuracy). The clustering analysis revealed two distinct behavioural subgroups with different ergonomic risk profiles. These findings highlight specific behavioural targets that can support ergonomic-awareness efforts and help reduce the likelihood of FHP development in academic environments
Smart Glove Design to Improve Accessibility Communication for the Deaf
No full textDeaf people rely on hand gestures as their primary means of communication; however, communication barriers often arise when surrounding individuals do not understand sign language. This study presents the design and evaluation of an Internet of Things (IoT)-based smart glove to improve communication accessibility for deaf individuals. The proposed system utilizes multiple MPU6050 motion sensors integrated with an Arduino Nano to detect finger and hand movements. Gesture recognition is implemented using a rule-based approach with predefined threshold values, enabling real-time detection without the need for training data. System performance was evaluated through response time and recognition accuracy measurements, as well as qualitative observations related to system stability and usability. Experimental results show response times ranging from 146–147 ms, indicating a fast and stable system. Recognition accuracy varies between 70% and 85%, depending on gesture complexity and finger movement patterns. Although the accuracy is moderate compared to machine learning-based approaches, the proposed system offers advantages in computational efficiency, simplicity, and ease of implementation. These findings demonstrate the potential of the smart glove as a practical assistive communication device, while also highlighting opportunities for further development through improved gesture modeling and user-centered evaluation.Deaf people rely on hand gestures as their primary means of communication; however, communication barriers often arise when surrounding individuals do not understand sign language. This study presents the design and evaluation of an Internet of Things (IoT)-based smart glove to improve communication accessibility for deaf individuals. The proposed system utilizes multiple MPU6050 motion sensors integrated with an Arduino Nano to detect finger and hand movements. Gesture recognition is implemented using a rule-based approach with predefined threshold values, enabling real-time detection without the need for training data. System performance was evaluated through response time and recognition accuracy measurements, as well as qualitative observations related to system stability and usability. Experimental results show response times ranging from 146–147 ms, indicating a fast and stable system. Recognition accuracy varies between 70% and 85%, depending on gesture complexity and finger movement patterns. Although the accuracy is moderate compared to machine learning-based approaches, the proposed system offers advantages in computational efficiency, simplicity, and ease of implementation. These findings demonstrate the potential of the smart glove as a practical assistive communication device, while also highlighting opportunities for further development through improved gesture modeling and user-centered evaluation
Comparison of K-Nearest Neighbor, Naïve Bayes, and C4.5 Algorithms for Predicting Academic Stress Risk in Students Based on Psychological Survey Data
No full textAcademic stress is a psychological problem experienced by many students and can have an impact on learning achievement, mental health, and quality of life. This study aims to compare the performance of the K-Nearest Neighbor (KNN), Naïve Bayes, and C4.5 (Decision Tree) algorithms in predicting the level of academic stress risk in students based on psychological survey data. Data were obtained from 700 active students at Ngudi Waluyo University through a questionnaire covering physiological, psychological, and behavioral aspects, with a total of 15 indicators using a Likert scale. The data then underwent pre-processing, labeling, standardization, and division into training and test data with a ratio of 80:20. The evaluation was conducted using the Accuracy, Precision, Recall, F1-Score, Confusion Matrix, and AUC-ROC metrics. The results showed that the Naïve Bayes algorithm performed best with an accuracy of 93.26%, precision of 93.35%, recall of 92.26%, and F1-score of 92.80%. The KNN algorithm was in second place with an accuracy of 91.43%, while the C4.5 algorithm had the lowest performance with an accuracy of 80.60%. Based on these results, Naïve Bayes is recommended as the most optimal algorithm for predicting academic stress risk in students using psychological survey data. This study is expected to assist educational institutions in identifying students at risk of stress early on and supporting the development of more effective prevention strategies
A Multi-Criteria Decision Approach to Livability Assessment Using Hybrid FUCOM–VIKOR
No full textPersistent disparities in regional livability across Central Java pose challenges for effective and equitable poverty alleviation policies. Without objective prioritization, government interventions risk being inefficient and misdirected. This study aims to assess the livability level of 35 regencies and cities in Central Java and to identify regions that should be prioritized for policy intervention. Secondary data for 2024 were obtained from the Central Statistics Agency (BPS) of Central Java Province. A hybrid multi-criteria decision-making approach combining the Full Consistency Method (FUCOM) and VIKOR was employed. FUCOM was used to generate consistent and objective weights for six indicators (Human Development Index, Life Expectancy, Number of Poor Residents, Open Unemployment Rate, Access to Proper Sanitation, and GRDP per capita), while VIKOR was applied to produce compromise-based rankings of regional livability. The ranking results were visualized using a bar chart to enhance interpretability and facilitate regional comparison. The results indicate that Salatiga City, Magelang City, and Surakarta City exhibit the highest livability levels, whereas Brebes Regency, Banjarnegara Regency, and Pemalang Regency consistently rank lowest, indicating an urgent need for targeted government intervention. Model validation using Normalized Discounted Cumulative Gain (NDCG = 0.9835) and Spearman Rank Correlation (ρ = 0.883) demonstrates strong consistency with reference data. These findings suggest that the FUCOM–VIKOR hybrid approach provides a robust and practical decision-support tool for evidence-based regional development planning and poverty alleviation prioritization.Persistent disparities in regional livability across Central Java pose challenges for effective and equitable poverty alleviation policies. Without objective prioritization, government interventions risk being inefficient and misdirected. This study aims to assess the livability level of 35 regencies and cities in Central Java and to identify regions that should be prioritized for policy intervention. Secondary data for 2024 were obtained from the Central Statistics Agency (BPS) of Central Java Province. A hybrid multi-criteria decision-making approach combining the Full Consistency Method (FUCOM) and VIKOR was employed. FUCOM was used to generate consistent and objective weights for six indicators (Human Development Index, Life Expectancy, Number of Poor Residents, Open Unemployment Rate, Access to Proper Sanitation, and GRDP per capita), while VIKOR was applied to produce compromise-based rankings of regional livability. The ranking results were visualized using a bar chart to enhance interpretability and facilitate regional comparison. The results indicate that Salatiga City, Magelang City, and Surakarta City exhibit the highest livability levels, whereas Brebes Regency, Banjarnegara Regency, and Pemalang Regency consistently rank lowest, indicating an urgent need for targeted government intervention. Model validation using Normalized Discounted Cumulative Gain (NDCG = 0.9835) and Spearman Rank Correlation (ρ = 0.883) demonstrates strong consistency with reference data. These findings suggest that the FUCOM–VIKOR hybrid approach provides a robust and practical decision-support tool for evidence-based regional development planning and poverty alleviation prioritization
Analysis of the Best Social Media Platforms for Promotion Using Machine Learning and RFE Feature Selection: A Comparative Study of Gradient Boosting, XGBoost, CNN, and SVR
No full textThis study aims to identify the most effective social media platforms for digital marketing. The use of social media for promotion continues to grow, yet many businesses still struggle to determine which platforms have the greatest impact. Therefore, this study compares the performance of various machine learning platforms to predict the best platform. The algorithms used are Gradient Boosting Regressor, XGBoost Regressor, Convolutional Neural Network (CNN), and Support Vector Regression (SVR) to estimate digital conversion potential based on user reviews, ad reach, and content trend patterns. A Knowledge Discovery in Databases (KDD) workflow is used to identify the most important key factors. This process includes data preprocessing, TF-IDF feature extraction, sentiment analysis, feature engineering, and feature elimination (RFE). The results showed that the CNN algorithm excelled in prediction, with the highest R² score of 0.74 and the lowest RMSE of 14.78. CNN predictions showed YouTube topping the list in terms of conversion potential, followed by Facebook and TikTok. These results highlight the higher promotional effectiveness of video-based platforms and the importance of machine learning in digital marketing decision-making. However, this study is limited by its reliance on static user review and ad reach data, which may not fully capture the dynamic changes of social media platforms