Sinkron : jurnal dan penelitian teknik informatika

The Mapping Elementary School Digital Transformation Readiness through SERI for Roadmap Development

Sondius Matogu Budiman Silalahi — 2026-04-02

Digital transformation has become a strategic priority in elementary education as schools are increasingly expected to integrate digital technology into teaching, assessment, and institutional management. However, previous studies on school digital readiness have generally focused on isolated aspects such as infrastructure, digital literacy, or leadership, without providing an integrated assessment model that simultaneously evaluates process, technology, and organisational dimensions in elementary school contexts. This study aims to assess the digital transformation readiness of an elementary school using the Smart Education Readiness Index (SERI). A descriptive quantitative case-study approach was employed by adapting the SERI assessment matrix into the elementary school context. The assessment covered three dimensions process, technology, and organisation through twelve indicators. Data were collected through a structured assessment matrix, supporting document review, and expert validation involving two educational technology experts. The results indicate that the school reached a moderate level of digital transformation readiness. The strongest indicators were specific or specialised skills (2.635), digital infrastructure readiness (2.634), digital interconnectivity (2.598), and organisational planning indicators (2.562), while the weakest indicators were assessment (1.708), policy guidance (1.708), general or transversal skills (1.744), and digital storage (1.852). Unlike previous studies that mainly assess digital readiness through separate technological or pedagogical indicators, this study applies a multidimensional institutional assessment framework. This study contributes by proposing a structured and adaptable assessment approach for elementary school digital transformation that supports the development of a more measurable and context-sensitive digital transformation roadmap.

Real time weather forcasting with conditional CNN and TCN-BiLSTM Ensamble at Manokwari

Ilham Tatayo Lie — 2026-04-02

Short-term weather forecasting is fundamentally critical for disaster mitigation in dynamic tropical maritime regions. However, conventional numerical approaches suffer from high computational latency, and spatial deep learning models frequently experience severe performance degradation during nocturnal conditions due to the absence of illumination. This study aims to develop an adaptive, real-time multimodal weather nowcasting system that effectively prevents nocturnal predictive failure through a dynamic conditional ensemble architecture. The proposed framework integrates a Convolutional Neural Network (CNN) to extract optical features from a dataset of 2,515 localized sky images with a Temporal Convolutional Network and Bidirectional Long Short-Term Memory (TCN-BiLSTM) pipeline to process 15,111 corresponding meteorological time-series records from BMKG. To address visual ambiguity, the system strictly employs a day-night gating mechanism, deactivating the CNN at night to rely solely on thermodynamic data. Finally, the optimized model was deployed via TensorFlow.js for decentralized client-side browser inference. Experimental evaluations explicitly demonstrate that the conditional ensemble significantly outperformed all standalone models, achieving a peak accuracy of 92.49% and a Macro F1-score of 0.913 while successfully preserving a robust recall rate for precipitation events. Furthermore, the browser-based deployment completely eliminated server transmission bottlenecks, achieving sub-second warm-start inference latency across heterogeneous consumer devices. Ultimately, the conditional day-night modality gating mechanism effectively mitigates nocturnal visual degradation, proving that implementing this integrated architecture as a client-side web application is highly feasible for delivering instantaneous public weather warnings.

ResASPP-UNet: A Modified U-NET Using ResNeT ASPP for Retinal Blood Vessels Segmentation

Salma Salsabila — 2026-04-17

In medical imaging, segmenting retinal blood vessels is a crucial task. A ResNet encoder, an ASPP module for multiscale feature extraction, and a UNet decoder comprise the modified U-Net architecture that this work suggests for retinal vascular segmentation. The suggested model extracts the green channel and applies CLAHE during data processing to segment retinal blood vessels. Accuracy, sensitivity, specificity, the Dice coefficient, and Intersection over Union (IoU) are used to assess performance. According to the experimental results, the proposed model obtains an Accuracy, sensitivity, specificity, the Dice coefficient, and IoU of 0.9554, 0.7294, 0.9771, 0.7408, and 0.5884 on the DRIVE dataset and an accuracy of 0.9170 on the DRIVE dataset. Meanwhile, 0.9556, 0.7902, 0.9702, 0.7386, and 0.5865 on the STARE dataset, respectively.

Application of Two-Stream Late Fusion on EfficientNetV2 based on Transfer Learning to classify AI-generated paintings

Muhammad Kevin Rinaldi — 2026-04-02

The rapid advancement of generative artificial intelligence (AI) has made synthetic digital paintings increasingly difficult to distinguish from human-made artworks, raising concerns regarding authenticity, copyright protection, and digital forensics. The main objective of this research is to develop a reliable and interpretable framework for distinguishing AI-generated paintings from human-created artworks by integrating visual and noise-based features. To address the limitations of conventional single-stream CNN models, this study proposes a Two-Stream Network with a Late Fusion strategy, combining a visual stream based on EfficientNetV2-S and a noise stream based on Xception with Spatial Rich Models (SRM).The proposed architecture processes semantic visual features and residual noise characteristics independently, followed by weighted decision-level fusion with a ratio of 0.7:0.3. Experiments were conducted using the AI-Artwork public dataset from Kaggle, consisting of 15,000 images with a data split of 64% training, 16% validation, and 20% testing. Model performance was evaluated using accuracy, precision, recall, F1-score, and ROC-AUC, ensuring a comprehensive assessment beyond accuracy alone. The results demonstrate that the proposed method achieves 98% accuracy, 98% precision, a 99% F1-score, and high discriminative capability compared to single-stream baselines. Model interpretability was analyzed using Grad-CAM to examine the contribution of each stream. Despite promising results, this study is limited by evaluation on a single dataset and static fusion weights, which may affect generalization to unseen generative models. Future work includes cross-dataset evaluation, adaptive fusion strategies, and exploration of lightweight architectures. Practically, this approach has potential applications in digital art authentication, forensic analysis, and content moderation systems, as well as supporting emerging policies for AI-generated content regulation and copyright protection.

Retrieval-Augmented LLM-Based Empathetic Chatbot for Early Postpartum Depression Screening in Aceh

Cut Amalia Saffiera — 2026-04-02

Postpartum Depression (PPD) remains a significant maternal mental health concern, particularly in low-resource settings where access to professional psychological services is limited. Although digital mental health tools have emerged to address this gap, most existing chatbot-based systems rely on rule-based interactions, offer limited personalization, and lack integration of structured clinical screening mechanisms. This study addresses the lack of culturally adapted, LLM-based empathic chatbots for postpartum mental health screening in low-resource Indonesian settings. We design and implement an AI-driven conversational chatbot that integrates a Retrieval-Augmented Generation (RAG) architecture with a Large Language Model (LLM) to enable context-aware, knowledge-grounded response generation. The system incorporates a Patient Health Questionnaire-9 (PHQ-9)–based screening module to support early identification of depressive symptoms and adaptive conversational support. An early-stage usability evaluation was conducted through a seven-day user interaction study involving 30 postpartum mothers in Aceh, with 12 participants completing the System Usability Scale (SUS). The system achieved an average SUS score of 85.63, indicating excellent perceived usability. While the evaluation focuses on usability rather than clinical effectiveness, the findings suggest that the proposed system demonstrates feasibility as a culturally adapted, scalable digital support tool for early postpartum mental health screening. Further studies with larger samples and long-term evaluation are required to assess clinical impact and sustained user engagement.

Multimodal Detection Models for Poultry Fraud Monitoring on Jetson Nano

Rachmad Atmoko — 2026-04-02

This study defines an indoor commercial poultry-house scenario with no Global Positioning System (GPS) signal, variable bird density, illumination shifts, occlusion, and normal versus fraud episodes characterized as abnormal poultry population behavior (an unauthorized deviation between observed bird count and expected inventory baseline). We evaluate an unmanned aerial vehicle (UAV) to an edge-computing pipeline on Jetson Nano by comparing three models: You Only Look Once version 11 (YOLOv11) with red-green-blue (RGB) input, YOLOv11 with RGB and thermal late fusion, and a convolutional neural network (CNN) backbone with a support vector machine (SVM) classifier. The dataset contains 12,000 frames with synchronized RGB-thermal augmentation to preserve modality alignment. Evaluation covers mean Average Precision (mAP), precision, recall, F1-score, counting errors via mean absolute error (MAE) and root mean square error (RMSE), and edge metrics including frames per second (FPS), latency, and memory. YOLOv11 RGB+thermal records mAP@0.5 of 0.94 (Table 4a), MAE of 1.4, and RMSE of 2.0 (Table 4b), compared with YOLOv11 RGB at 0.91, 1.8, and 2.5 and CNN-SVM at 0.85, 2.6, and 3.4 (Table 4a-4b). For edge throughput, CNN-SVM reaches 28 FPS, while YOLOv11 RGB reaches 18 FPS and YOLOv11 RGB+thermal reaches 14 FPS (Table 8). As a scenario study, these metric-supported results indicate that YOLOv11 RGB+thermal is accuracy-first, CNN-SVM is speed-first, and YOLOv11 RGB is a balanced option for real-time poultry fraud monitoring.

From Methodologies to Metrics: A Review of Aspect-Based Sentiment Analysis Approaches

Marwa Esmaeel — 2026-04-02

Abstract: ABSA (Aspect-Based Sentiment Analysis) has been developed as a fine-grained sentiment analysis tool, which finds the sentiment towards a particular aspect, enabling more accurate sentiment mining in a variety of domains. Over the past decade ABSA research has transcended lexicon-driven and traditional machine learning methodology using deep learning and transformer-based pre-trained language models to generative large language models. Nevertheless, underlying issues remain: implicit aspect extraction, low cross-domain and cross-lingual robustness, dataset imbalance, and interpretability concerns of complex neural networks. In addition, the rapid scaling of ABSA subtasks has led to some fragmentation in methodological advances in earlier investigations. By methodically reviewing the development of methodological paradigms, benchmark datasets, and evaluation approaches, this review has offered a systematic and rigorous assessment of the literature on ABSA. Unlike previous reviews, the study adopts a holistic, task-aware view and makes a direct connection between ABSA subtasks and the accompanying modeling methodologies. The review explores new research directions such as explainable ABSA, meta-based learning frameworks, multilingual and low-resource modeling, and large language model integration, thus providing a structure toward the road to developing more resilient, interpretable, and generalizable ABSA systems.

Multiclass SVM with Kernel Optimization for Schizophrenia Subtype Classification Using Clinical Symptom Records

Reisa Maulidya Rohman — 2026-04-02

Schizophrenia is a mental disorder that affects about 0.3% of the world population. It is characterized by a wide range of symptoms that form several subtypes. Overlapping symptoms and subjective clinical assessments may reduce consistency and make subtype classification challenging. Machine learning algorithms that use patients’ medical records offer a potentially objective approach for subtype classification. This study aims to classify four schizophrenia subtypes: paranoid, catatonic, undifferentiated, and residual, based on subtype labels recorded in the hospital using a multiclass SVM approach with kernel optimization. The dataset consists of 218 medical records of schizophrenia patients with 25 binary symptom variables used as input features. SVM was trained using two multiclass approaches, namely OAO and OAA. Evaluation was performed using five-fold stratified cross-validation. Performance was calculated using accuracy, macro-precision, macro-recall, and macro F1-score. Optimal performance was achieved using the OAA approach with an RBF kernel at C = 10 and gamma = 0.1. This configuration achieved an accuracy, macro-precision, macro-recall, and macro F1-score of 0.89, 0.90, 0.86, and 0.87, respectively. These results show that the multiclass approach, kernel functions, and parameter configuration influence classification performance. The proposed model may serve as a screening or decision-support tool to assist subtype identification based on clinical symptom records.

Cross-Architecture Performance Evaluation of Transfer Learning Models for Multi-Class Vehicle Damage Severity Classification

Mochammad Fatih Ulumuddin — 2026-04-02

Automated vehicle damage classification supports objectivity and scalability in insurance claim processing and digital inspection systems; however, prior studies often report performance improvements without controlled experimental settings or statistical validation, limiting methodological reliability. This study establishes a statistically controlled cross-architecture evaluation framework to determine whether pretrained convolutional neural networks significantly outperform a custom baseline model in multi-class vehicle damage classification. A dataset of 891 labeled vehicle images categorized into heavy, medium, light, and normal damage was partitioned using stratified sampling (70% training, 15% validation, 15% testing). Four architectures Baseline (CustomCNN), VGG16, ResNet50, and MobileNetV2 were trained under identical preprocessing and optimization settings with two training durations (30 and 50 epochs). Five-fold cross-validation and paired t-test analysis were applied to assess statistical significance. At 30 epochs, MobileNetV2 achieved the highest accuracy (75.76%), while at 50 epochs VGG16 obtained the best performance (78.03%). Extending training duration did not produce statistically significant improvement (p > 0.05). Pretrained architectures significantly outperformed the baseline model, whereas ResNet50 did not demonstrate superior performance. The novelty of this study lies in its statistically validated comparative framework. Although limited by moderate dataset size and single-source imagery, the findings provide practical guidance for selecting efficient convolutional neural networks in vehicle damage classification systems.

Analysis of the Effectiveness of a Music Learning Information System for Early Childhood (Golden Age) using the Technology Adoption Model (TAM)

Susan Juli Safitri — 2026-04-08

Advances in information technology have brought significant changes to the education sector, including in the learning process for early childhood. The integration of information systems into learning activities is no longer limited to administrative functions but also serves as a learning medium capable of enhancing the effectiveness of the learning process. This study aims to analyze the effectiveness of implementing a music learning information system and to examine the relationship between the quality of the information system and student learning outcomes. The research analysis framework is based on the Technology Acceptance Model developed by Fred D. Davis to explain how perceptions of ease of use and information quality influence users’ adoption of technology. This study employs a quantitative approach with a correlational design. The study population consists of students in the New Primary music program at Yamaha Forte Music Bandung, totaling approximately 800 students. The research sample was determined using purposive sampling, calculated as rxy = (nΣXY − (ΣX)(ΣY)) / √[(nΣX² − (ΣX)²)(nΣY² − (ΣY)²)], with a total of 63 students who actively supported the use of the digital learning system. Data collection was conducted via a Likert-scale questionnaire measuring three primary dimensions of the information system: usability, information quality, and learning impact. Validity was assessed using Pearson’s Product-Moment correlation, while reliability was evaluated using Cronbach’s Alpha > 0.70. The research results indicate that the effectiveness of the music learning information system falls into the “very effective” category, with an average score of 81.3%. The usability dimension received the highest score of 87.5%, followed by information quality at 84.2%, and learning impact at 81.3%. The results of the correlation analysis indicate a positive relationship between the quality of the information system and student learning behavior, such as increased practice discipline, learning motivation, and self-confidence in music.

Improving Multi-Class Public Complaint Classification with LSTM, Word2Vec, and Random Oversampling

Azza Nimasari — 2026-04-02

Digital transformation in the public sector encourages local governments to enhance service quality through online complaint management systems. However, the high volume of incoming complaints and significant data imbalance across 31 Organisasi Perangkat Daerah (OPD) pose challenges for efficient manual classification, often resulting in delays and misclassification. This study proposes an automated text classification model that integrates Long Short-Term Memory (LSTM), Word2Vec, and Random Oversampling (ROS), optimized using the Adam algorithm. The novelty of this research lies in the integration of sequential modeling and imbalance handling to address an extreme multi-class classification problem involving 31 OPD categories within a highly imbalanced dataset. The research stages include text preprocessing, word embedding construction using Word2Vec, data balancing through ROS, and model training using LSTM. Experimental results show that the proposed model achieves an accuracy of 0.72, with macro-average precision, recall, and F1-score of 0.67, 0.67, and 0.66, respectively. Considering the complexity of classifying 31 classes and the presence of severe data imbalance, the macro F1-score of 0.66 indicates that the model is reasonably effective in capturing classification patterns, although performance is not yet evenly distributed across all classes. Overall, the combination of LSTM, Word2Vec, and ROS demonstrates potential as a baseline approach for automating public complaint classification in complex multi-class scenarios. The proposed model can improve the accuracy and speed of complaint distribution to the appropriate OPD, thereby enhancing the efficiency and responsiveness of public service delivery compared to conventional manual methods.

A Statistical Benchmarking of Imbalance-Aware Ensemble Models for Cervical Cancer Prediction

Sumarna Sumarna — 2026-04-02

Cervical cancer remains one of the leading causes of cancer-related mortality among women worldwide, particularly in developing countries. Early prediction through machine learning has the potential to support clinical decision-making; however, cervical cancer datasets often suffer from severe class imbalance, which reduces the ability of conventional models to correctly detect minority cases. This study aims to improve minority class detection in cervical cancer prediction by evaluating several imbalance-aware ensemble learning approaches. The proposed study compares five models, namely Random Forest (RF), SMOTE combined with Random Forest (SMOTE+RF), Balanced Random Forest (BRF), EasyEnsemble, and RUSBoost. The models were evaluated using 5-fold cross-validation with performance metrics including accuracy, recall, F1-score, and Area Under the Curve (AUC). Statistical validation was conducted using the Friedman test, followed by the Wilcoxon signed-rank test and Kendall’s W effect size analysis to assess the significance and magnitude of performance differences. Unlike prior studies that primarily focus on performance improvement, this study introduces a statistically rigorous comparative evaluation to assess both significance and practical effect of imbalance-aware ensemble methods. Experimental results show that imbalance-aware ensemble methods significantly improve minority detection compared to the baseline RF model. In particular, BRF achieved the highest AUC of 0.9469 with improved recall stability, while RUSBoost produced the highest F1-score of 0.7451. Although the Friedman test indicated no statistically significant difference among models (p = 0.2037), the Kendall’s W value of 0.297 suggests a small-to-moderate practical effect. These findings indicate that imbalance-aware ensemble learning can enhance the robustness of cervical cancer prediction models, particularly for minority class detection. The results highlight the importance of incorporating imbalance-handling strategies in medical prediction systems and suggest potential directions for future research in improving diagnostic decision-support models.

Novel Adaptive TOPSIS for User-Centered Mobile University Libraries: Enhancing Usability and Engagement

Sony Veri Shandy — 2026-04-15

The rapid development of mobile technology has prompted universities to enhance digital library services. However, many existing mobile library applications still suffer from usability issues, reducing user satisfaction and effectiveness in information access. This study aims to design and evaluate a mobile university digital library interface by integrating an Adaptive TOPSIS method with a User-Centered Design (UCD) approach. The research followed several stages: user needs analysis, criteria weighting using Adaptive TOPSIS, interface design, prototyping, and iterative usability testing with 30 university students. Adaptive TOPSIS was employed to prioritize design alternatives based on user preferences, enabling more objective and adaptive decision-making during the design process. The prototype included book searching, digital catalog access, and online borrowing features. Usability testing measured effectiveness, efficiency, and satisfaction. The system achieved a task success rate of 92%, average task completion time of 18.4 seconds, and error rate of 5.6%. The System Usability Scale (SUS) score averaged 84.2, indicating excellent usability. These results demonstrate that combining Adaptive TOPSIS with UCD significantly enhances both decision-making in design and overall user experience. This study underscores the value of integrating multi-criteria decision-making methods with user-centered UI/UX design to improve accessibility, efficiency, and satisfaction in mobile university digital libraries.

An Enterprise Architecture Blueprint for HL7-Based RIS–PACS Integration Using TOGAF ADM: A Mobile Diagnostic Access Framework

Wandi Purnama — 2026-04-02

The shift to digital radiology is frequently hindered by data fragmentation and restricted mobility for healthcare professionals. This research addresses these deficiencies by developing an Enterprise Architecture (EA) blueprint for hospital radiology departments using the TOGAF ADM framework. The design emphasizes HL7 standards for administrative data transfer and DICOM protocols for imaging, featuring a web-based Mobile Viewer using the WADO protocol for remote high-resolution access. Architectural validation via gap analysis demonstrates that the proposed integration can theoretically eliminate manual data entry between HIS and RIS-PACS through an HL7 broker mechanism, potentially shortening clinical report turnaround times. However, this study is limited to architectural validation and does not include live clinical deployment. The proposed architecture offers a scalable roadmap for "film-less" environments, though further longitudinal studies are required to assess long-term clinical user satisfaction.

Integration of Chatbot and Complaint Website Using Agile Scrum with Load Testing and UAT

Galih Adi Winasis — 2026-04-02

This study investigates an integrated public complaint service that combines a non-AI, rule-based WhatsApp chatbot, a web-based administrative dashboard, and a RESTful API to improve early response, status traceability, and ticket-based two-way communication. The system was developed using an Agile Scrum approach, implementing the chatbot in Node.js, the backend services and dashboard in Laravel, and PostgreSQL as the centralized database, while real-time dashboard updates were delivered via WebSocket. Evaluation was conducted through User Acceptance Testing (UAT) for core functional flows and RESTful API load testing using Apache JMeter under gradual-load conditions (Typical Busy, Peak, Stress) and an extreme surge condition (Spike/Burst). The UAT results indicate that all core scenarios passed, covering ticket-based complaint submission, duplicate prevention via a one active ticket per WhatsApp number rule, administrator validation and routing, and real-time conversation synchronization within the ticket context. Under gradual-load conditions, all evaluated endpoints maintained a 0% error rate with sub-second average latency in the range of a few hundred milliseconds, indicating stable baseline behavior as workload increased progressively. Under Spike/Burst, the system remained error-free but latency increased, with average response times of 6,593 ms for create complaint, 18,010 ms for status tracking, 18,321 ms for chat message, and 14,308 ms for mixed load, with throughputs of 7.06 req/s, 2.62 req/s, 2.05 req/s, and 5.90 req/s, respectively. Overall, the results demonstrate end-to-end functional feasibility, stable baseline performance under gradual load, and a resilience boundary under extreme surge, motivating targeted optimization of synchronous processing, history retrieval, and payload serialization to improve Spike/Burst time responsiveness.

Multi-Variable Agrometeorological Parameter Combination for Drought Early Warning System Using Hybrid SSA–AMBP Algorithm

Syahrull Rezaa — 2026-04-02

Drought is one of the hydrometeorological disasters that has a significant impact on the agricultural sector, water availability, and food security, thus requiring an accurate and adaptive early warning system. This study aims to develop and evaluate a drought Early Warning System (EWS) model based on a combination of multi-variable agrometeorological parameters using a hybrid approach of Singular Spectrum Analysis (SSA) and Adaptive Model-Based Prediction (AMBP). The agrometeorological data used includes rainfall, air temperature, humidity, solar radiation, wind speed, and other supporting variables processed in the form of monthly time series over a period of ten years. The SSA method is used to perform signal denoising and extract dominant components from the data, while AMBP is applied as an adaptive predictive model to generate SPI-6 drought index forecasts. Model performance is evaluated using RMSE and the coefficient of determination (R²) in the model training and evaluation phases. The results show that the hybrid SSA–AMBP model has the best performance compared to single methods, with an RMSE value of 0.149 and R² of 0.983 in the model training phase, and an RMSE of 0.176 and R² of 0.941 in the model evaluation phase. In addition, the 2026 prediction results show a seasonal pattern with indications of a moderate dry period from October to November. These findings indicate that the developed model demonstrates relatively high predictive accuracy and stability based on RMSE and R² evaluation metrics within the SPI-6 dataset used in this study, and it has the potential to serve as a conceptual basis for decision-support in drought risk mitigation, water resource management, and sustainable agricultural planning.

Web3-Based Cyber Incident Reporting System With Smart Contracts and Non-Fungible Token Rewards

Danang Juniar Permana — 2026-04-02

The rising frequency of cyber threats increases the need for incident reporting that is transparent, efficient, and privacy-preserving. This study designs and implements a hybrid Web2-Web3 cyber incident reporting prototype that anchors report references on a blockchain while storing full incident details off-chain, and explores non-fungible token (NFT) recognition incentives for reporters. Using an SDLC-based iterative prototyping approach, we built a React single-page application integrated with a Laravel REST API and MySQL for off-chain storage, and deployed Solidity smart contract modules on the Arbitrum Sepolia testnet to record report identifiers and UUID pointers (dataPointer) and to mint NFTs after administrative validation. We conducted black-box functional testing across core scenarios (submission, storage, pointer anchoring, validation, and minting) and a user acceptance study with 25 participants (15 cybersecurity students and 10 IT practitioners) using a 5-point Likert questionnaire. All tested scenarios executed as expected in the test environment, and on-chain events were traceable to corresponding backend records via transaction receipts and logged identifiers. The acceptance evaluation yielded an overall mean score of 3.4/5 (about 68%), indicating moderate acceptance and supporting the work as a prototype feasibility study rather than organizational-level generalization. The prototype demonstrates a practical workflow for hybrid incident reporting with transaction-level traceability and recognition incentives; future work should strengthen cryptographic binding (e.g., content hashing) and validate the approach with CSIRT stakeholders in operational settings.

Bidirectional Long Short-Term Memory for Early Detection of Running Injuries in Imbalanced Data

David David — 2026-04-02

Running-related injuries are a common sports health issue that can impair athletic performance and potentially terminate an athlete’s career. Early injury detection is therefore critical, as injuries are cumulative in nature and influenced by training load patterns over time. Consequently, data-driven predictive approaches based on time-series analysis are required to support athlete monitoring systems with a safety-oriented focus. This study aims to develop an efficient, accurate, and safety-first injury prediction model for running athletes. The study utilizes daily running activity time-series data obtained from Kaggle. The proposed model is based on a Bi-Directional Long Short-Term Memory (Bi-LSTM) architecture to capture bidirectional temporal dependencies, combined with Focal Loss to address extreme class imbalance. In addition, domain-specific feature engineering is applied through the Acute:Chronic Workload Ratio (ACWR). Model performance is evaluated against tabular-data-based models, namely XGBoost and Balanced Bagging, across multiple experimental configurations. Experimental results indicate that the lightweight Bi-LSTM configuration achieves a Recall of 90.7%, outperforming the benchmark models while maintaining a competitive AUC. These findings demonstrate that sequential modeling is more effective in detecting rare injury events. Overall, this study confirms that Bi-LSTM-based sequential modeling is well suited for early detection of running injuries and suggests its potential applicability in athlete monitoring systems that prioritize safety.

Implementation of Semantic Search in an Academic Repository Using Sentence-BERT and FAISS

Ihsan Lubis — 2026-04-02

Academic repositories serve as centralized platforms for storing and managing scientific documents, including research papers, reports, and administrative records. Yet, traditional keyword-based search systems often struggle to deliver relevant results. These systems typically fail to capture the contextual meaning of user queries, which leads to mismatches when the query terms differ from those found in the documents. To overcome this limitation, this study introduces a semantic search approach for academic repositories by combining Sentence-BERT as the text embedding model with FAISS as the vector-based similarity search engine. In the proposed system, documents stored in a MySQL database are first preprocessed to remove HTML tags, then converted into semantic vector representations using Sentence-BERT. These vectors are indexed with FAISS, enabling fast and efficient similarity searches compared to conventional keyword matching. The system architecture integrates FastAPI as the backend service for indexing, searching, and evaluation, while CodeIgniter 4 functions as the frontend framework for document management by administrators and end users. Evaluation was carried out using three test sets, each containing ten queries. Performance was measured using Recall@K, normalized Discounted Cumulative Gain (nDCG), Mean Reciprocal Rank (MRR), Mean Average Precision (MAP), and search latency. Experimental results show that the system achieved an average Recall@K of 0.61, a MAP of 0.39, and a No-Hit rate of 0.033, meaning all queries successfully retrieved results. Although the nDCG value declined in the third test set, the system consistently returned relevant documents.

Comparative Academic Performance Prediction in Primary Schools Using Linear Regression and Random Forest

Agustinus Sembiring — 2026-04-02

Predicting academic performance is an important aspect of data-driven decision making in education, particularly in primary schools where early identification of learning difficulties is crucial. This study compares the performance of Linear Regression and Random Forest Regression models for predicting students’ academic performance using an Educational Data Mining approach. The experiment uses the Students Performance Dataset from Kaggle, consisting of 1000 student records with eight predictor variables, including demographic and learning-related attributes. The target variable is the average score derived from mathematics, reading, and writing results. Model development and evaluation are conducted using Python in Google Colaboratory. Performance is assessed using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and the coefficient of determination (R²), while Random Forest is further optimized using GridSearchCV with 5-fold cross-validation. The results show that Linear Regression achieves the best performance (R² = 0.162, RMSE = 13.40, MAE = 10.49), outperforming both the default Random Forest (R² ≈ 0.000) and the tuned Random Forest (R² ≈ 0.112). Although the explained variance is relatively low, this finding indicates that simple demographic features provide limited predictive power for academic performance. A case study using a local dataset from a private primary school involving 132 sixth-grade students further confirms that Linear Regression performs more effectively than Random Forest for small and simple educational datasets. These results highlight the importance of aligning model selection with dataset characteristics in educational data mining.

Artificial Intelligence Usage Intention for Sustainable Development: A Neo ESG Perspective Using Hybrid Methods

Vo Dinh Cao Nguyen — 2026-04-02

This study finds that the rapid development of artificial intelligence, together with the growing pressure to implement environmental, social, and governance principles, has driven firms to search for new models of sustainable governance. However, prior research has lacked empirical evidence on the role of artificial intelligence usage intention within a dynamic environmental, social, and governance framework and its interplay with social and environmental dimensions. To address this gap, the study reconceptualizes environmental, social, and governance by representing governance through artificial intelligence, the social dimension through diversity, equity, and inclusion, and the environmental dimension through exploitative green innovation and exploratory green innovation. Based on survey data from 357 firms, a hybrid methodological approach employing partial least squares structural equation modeling, artificial neural networks, and fuzzy set qualitative comparative analysis is applied. The results reveal that diversity, equity, and inclusion has the strongest effect on sustainable development (β = 0.533; t = 13.061; p < 0.001), followed by artificial intelligence, while exploitative green innovation plays a supportive role and exploratory green innovation shows no significant impact. Artificial neural networks validate these findings with stable predictive accuracy, while fuzzy set qualitative comparative analysis identifies multiple alternative pathways to sustainability (equifinality). The study contributes by positioning artificial intelligence as a new governance mechanism within environmental, social, and governance and highlighting the central role of diversity, equity, and inclusion, while also offering strategic guidance for integrating technological and social factors to foster sustainable development.

Improving Optic Disc and Optic Cup Segmentation with Flip-Gamma Augmentation and SegFormer

Fitri Salamah — 2026-04-10

The Cup-to-Disc Ratio (CDR) is widely used as a diagnostic indicator for glaucoma, although variations and irregularities can influence its accuracy in the Optic Disc (OD) and Optic Cup (OC). To overcome this challenge, automated image segmentation is used. However, image segmentation is challenged by image blurriness, noise, and uneven illumination, which can affect segmentation quality and increase the risk of misdiagnosis. To address these challenges, this study applies a combined Flip-Gamma Augmentation and SegFormer approach for OD and OC segmentation. Flip-Gamma augmentation increases image diversity and improves image quality by adjusting brightness and contrast. Meanwhile, the SegFormer uses a Transformer-based backbone and efficiently extracts multi-scale features to enhance segmentation performance. Experimental results on the Drishti-GS dataset show that applying Flip-Gamma (δ = 0.8, 0.9, 1.1, 1.2) is associated with improved segmentation performance across all classes, with sensitivity (90-99%), DSC (90-99%), IoU (82-99%), and ROC (94-99%), indicating consistent segmentation of OD, OC and background regions. Furthermore, a one-sided Mann-Whitney U test indicates differences in performance compared to other augmentation methods. These findings suggest that the proposed augmentation strategy is beneficial for segmentation on the Drishti-GS dataset. However, further validation on larger and more diverse datasets is required to assess generalizability.

Decision Support System (DSS) for Rodenticide Selection using the TOPSIS Method

Ayu Tri Nur Fitasari — 2026-04-02

Selecting an appropriate rodenticide is a critical decision in pest control operations, as each product differs in effectiveness, application cost, safety level, environmental impact, and resistance potential. In practice, rodenticide selection is often based on technician experience or habitual product use, which may result in subjective and less optimal decisions. This study aims to develop a decision support system for rodenticide selection using the TOPSIS method within a multi-criteria decision-making (MCDM) framework. The evaluation is conducted based on six criteria: effectiveness, application cost, safety derived from LD50 values, secondary poisoning risk, resistance potential, and application convenience. To improve the robustness of the decision-making model, this study incorporates an adaptive TOPSIS approach through scenario-based weighting and compares the results with the Simple Additive Weighting (SAW) method. The findings show that alternatives with a balanced performance in terms of safety and operational cost consistently achieve higher rankings, with Warfarin Bait and Zinc Phosphide appearing as top-performing options across different evaluation scenarios. In addition, the proposed model is implemented in a web-based system using a prototype development approach, enabling automated calculations and transparent ranking results. This study provides a structured and practical decision support model that integrates technical, economic, and environmental considerations to support more objective decision-making in pest control management.

Orchestrating the Circular Smart Palm Ecosystem: A Design Science Research Using TOGAF-Based Enterprise Architecture

Ibnu Hamdani — 2026-04-02

The global palm oil industry is currently confronting a sustainability trilemma involving production efficiency, regulatory compliance (e.g., EUDR), and environmental circularity. Existing supply chain models are characterized by fragmented information systems, where Operational Technology (OT) remains disconnected from Enterprise IT, resulting in traceability gaps and underutilized biomass waste. This study addresses these challenges by designing a comprehensive Enterprise Architecture (EA) for a Circular Smart Palm ecosystem. Adopting a Design Science Research (DSR) methodology, the research applies the TOGAF Architecture Development Method (ADM), supported by ArchiMate 3.1 modeling, to develop an integrated architectural blueprint. The proposed architecture consolidates blockchain-enabled traceability, digital export compliance, IoT-driven bioenergy valorization, and value-added kernel processing into a unified enterprise framework. The resulting architecture is qualitatively validated through architectural consistency analysis and gap assessment between the As-Is and To-Be states. From a theoretical perspective, this study contributes to Enterprise Architecture literature by extending EA as an integrative mechanism for circular bioeconomy implementation in sustainability-driven agribusiness ecosystems. The findings demonstrate that enterprise-level architectural integration can transform sustainability compliance from an operational constraint into a strategic enabler for traceability, circular value creation, and digital governance.

Finite-Key Analysis of BB84 and B92 QKD with Discrete Phase Randomization and Koashi Bound

Brenendra Putra Oktaviansyah — 2026-04-02

Quantum Key Distribution (QKD) enables theoretically secure key exchange based on fundamental quantum principles such as the no-cloning theorem and Heisenberg’s uncertainty principle. However, practical implementations remain vulnerable to side-channel attacks caused by device imperfections, while many existing studies primarily analyze asymptotic security or isolated attack scenarios rather than realistic finite-key conditions. Unlike prior studies that focus on asymptotic or single-attack analyses, this work presents a comprehensive finite-key security evaluation of BB84 and B92 protocols under hybrid side-channel attacks using Discrete Phase Randomization (DPR) as a lightweight mitigation strategy and the Koashi bound for improved phase-error estimation in B92. Numerical simulations are performed using realistic system parameters with a finite-key size of 100 billion pulses across ten representative attack scenarios. The results show that applying DPR (M = 32) significantly suppresses phase-sensitive attack-induced errors, reducing the quantum bit error rate (QBER) from 11–50% to approximately 1.5–3.02%, thereby restoring practical secure key generation. B92 with the Koashi bound achieves secure transmission distance improvements from 181.6 km to 190.8 km without attacks and reaches 187.0 km under hybrid attacks with DPR, slightly exceeding BB84 in certain conditions. Peak secret key rates reach 0.1363 bit/pulse for BB84 and 0.0741 bit/pulse for B92. These findings demonstrate that non-orthogonal protocols can remain competitive under realistic finite-key constraints using practical mitigation techniques, although literature based induced QBER assumptions remain a limitation.

A Comparative Methodological Study of Automated Machine Learning for Multiclass Stunting Prediction Using Anthropometric Data

Joharini Joharini — 2026-04-02

Stunting remains a major public health challenge among children under five years old and requires reliable early screening to support timely nutritional interventions, particularly in resource-limited healthcare settings. However, many existing stunting prediction studies rely on complex socio-economic variables and manually selected machine learning models, which limits reproducibility and practical deployment. This study proposes an automated machine learning (AutoML)–based framework for multiclass stunting prediction using routinely collected anthropometric data. The prediction task is formulated as a multiclass classification problem encompassing normal growth, stunted, severely stunted, and above-normal nutritional status. The proposed framework integrates standardized preprocessing, systematic model comparison, stratified 10-fold cross-validation, and controlled hyperparameter optimization, evaluated under SMOTE and non-SMOTE preprocessing scenarios. Experimental results demonstrate that reliable multiclass prediction can be achieved without socio-economic variables. Under SMOTE preprocessing, the optimized k-Nearest Neighbors model improves minority-class sensitivity, increasing accuracy from 0.9806 to 0.9820 with an MCC of 0.9688, while under non-SMOTE conditions, Random Forest achieves robust performance with an accuracy of 0.9985 and an MCC of 0.9975 without resampling. Confusion matrix, ROC, and learning curve analyses confirm strong discriminative capability and stable generalization for both models. Overall, the findings indicate that the proposed AutoML-based framework provides a practical, scalable, and reproducible solution for early multiclass stunting screening using anthropometric data alone.

Contextual Smart School Architecture Integrating SERI and TIER for Digital Transformation

Agustinus Sembiring — 2026-04-02

The digital transformation of elementary education has become an inevitable demand in the era of the Fourth Industrial Revolution. Nevertheless, schools in non-metropolitan regions continue to face persistent challenges, including limited infrastructure, low technology penetration, and insufficient policy support. This study aims to design a contextual smart school architecture by integrating the Smart Education Readiness Index (SERI) and the Transformation Impact and Essential Readiness (TIER) framework. A descriptive–qualitative approach, supported by quantitative survey data from 40 educators and education personnel, was employed to assess institutional readiness and formulate strategic priorities. The SERI assessment revealed an average digital readiness score of 3.12 (scale 0–4), with four dominant dimensions: Teaching and Learning Process (3.45), Assessment (3.28), Innovative Analysis (3.21), and IR 4.0 Policy (3.30). These dimensions were further validated through a Prioritisation Matrix weighted at 60% for cost factors, 20% for key performance indicators, and 20% for contextual proximity. The findings emphasize that effective digital transformation must leverage local strengths, be aligned with global frameworks, and be implemented through structured strategies. The key contribution of this research lies in the proposal of a modular, integrated, and sustainable smart school architecture model that can be replicated nationally to bridge global standards with local realities. This study provides both theoretical insights and practical implications for policymakers and educational leaders seeking to advance equitable digital transformation in non-metropolitan schools.

Improving Brain Tumor Classification Performance Using EfficientNetB0 Integrated with CBAM Attention Mechanism

Abd Salam At Taqwa — 2026-04-02

Accurate classification of brain tumors using magnetic resonance imaging (MRI) requires robust automated methods to support clinical diagnosis, particularly when tumor types present subtle visual distinctions. In this study, the Convolutional Block Attention Module (CBAM) is incorporated into the EfficientNetB0 architecture to improve feature representation for multi-class brain tumor classification. The performance of the proposed model is evaluated against the baseline EfficientNetB0 under identical training and testing conditions. EfficientNetB0 with CBAM achieves a training accuracy of 99.76% and a validation accuracy of 99.45%, with corresponding training and validation losses of 0.0085 and 0.0241. On an independent test dataset, the model attains a test accuracy of 99.25% and a loss of 0.0207. In comparison, the baseline EfficientNetB0 model attains a training accuracy of 52.68%, validation accuracy of 46.20%, and test accuracy of 43.32%, accompanied by significantly higher loss values. At the class level, the proposed model demonstrates macro-average precision, recall, and F1-score of 0.99, whereas the baseline model yields macro-average values of approximately 0.54 for precision and recall, and 0.50 for F1-score. Although CBAM integration increases computational time per evaluation step from 395 ms to 601 ms, the marked improvement in classification accuracy and error reduction underscores the value of attention mechanisms. These results demonstrate that attention-based feature refinement significantly enhances deep learning performance for medical image classification, particularly in multi-class brain tumor diagnosis.

Developing an Integrated Capital Assistance and Community Training System Using Agile Scrum

Ahmad Muzaki Zuhdi — 2026-04-02

Local governments increasingly require Cross-Agency Integration platforms to deliver transparent, auditable public services, yet capital assistance and community training programs are often managed through fragmented applications and manual workflows, leading to duplicated data, slow verification, and limited status traceability. This study develops an integrated capital assistance and community training system for local government using Agile Scrum, and evaluates its functional acceptance, usability, and security readiness to support Public Service Digitalization. Requirements were elicited through observation and interviews across three service-managing municipal agencies, while system governance and evaluation also involved the Communication and Informatics Office. The system was implemented as a web application with iterative sprints and backlog prioritization. Evaluation employed a User Acceptance Test (Likert 1–5, 10 items), System Usability Scale, and penetration testing using OWASP ZAP focusing on session management and HTTP security headers. Fifteen agency users participated in the evaluation. The system achieved 93% functional acceptance and a System Usability Scale score of 82.3, indicating excellent perceived usability. Security scanning found no high-risk issues, while medium- and low-risk findings were dominated by missing headers (Content Security Policy and X-Frame-Options) and incomplete cookie flags, which can be mitigated through standard hardening. The proposed platform improves cross-agency coordination and citizen-facing transparency while meeting usability expectations. Agile Scrum enabled rapid alignment with stakeholders and incremental quality improvements. Future work includes analytics, financial-system integration, and continuous security monitoring.

Field Evaluation of an IoT-VFD Smart Ventilation System for Energy-Efficient Rice Seed Storage

Hendra Yufit Riskiawan — 2026-04-02

Stable storage conditions are required in Rice Seed Storage to preserve seed quality and suppress fungal contamination, yet many warehouse ventilation systems still rely on inefficient on-off operation with limited responsiveness to changing temperature and humidity conditions. This study addresses the lack of integrated IoT-VFD control with field-validated energy and microclimate performance in seed warehouses. It proposes an IoT-based Ventilation Control architecture that combines ESP32, MQTT communication, and a Variable Frequency Drive to regulate a three-phase exhaust fan in both offline and online operating modes. The novelty of this work lies in integrating variable-speed control, real-time supervision, and field-based performance validation within a single seed warehouse deployment. The prototype was implemented in a 900 m³ warehouse at Politeknik Negeri Jember and evaluated through a 7-day field trial with continuous monitoring of temperature, humidity, and motor speed. The controlled system brought warehouse conditions closer to the intended storage setpoints and produced statistically significant improvements in both temperature and humidity (p < 0.001). Control performance was stable, with high target-hit accuracy and low RMSE, while energy testing showed lower electricity consumption than conventional non-VFD operation. Over an equivalent 2-hour operating period, energy use was reduced by 30.4%. The system also maintained 99.64% MQTT uptime, and no mold incidence was observed during controlled operation. These findings indicate that the proposed IoT-VFD architecture is a practical approach for improving microclimate stability, reducing energy use, and supporting fungus-preventive seed warehouse management.