Journal of ICT Research and Applications

Fine-tuning NER for Triplet Extraction in Medical Knowledge Graph Construction

2025-11-03T11:10:31+07:00

This study presents a new approach for constructing a medical knowledge graph using Named Entity Recognition (NER) to identify entities such as diseases, drugs, or medical procedures, alongside part-of-speech (POS) tagging and dependency parsing to determine words that function as verbs and roots. These extracted words are then used as relations between entities, forming triplets in the format (entity, relation, entity). While the knowledge graph provides a structured representation of medical information, the evaluation primarily reflects the performance of the underlying NLP pipeline (NER, POS tagging, and dependency parsing) used to generate the triplets. Quantitative evaluation was performed using metrics such as precision, recall, and F1-score to assess the accuracy and completeness of entity and relation extraction. The qualitative evaluation involved medical domain experts to assess the relevance and validity of the relationships derived. The results indicate that fine-tuning a pre-trained model for NER and leveraging a pre-trained model for POS tagging and dependency parsing can effectively generate accurate triplets for constructing a medical knowledge graph. This approach demonstrated strong performance, achieving high evaluation scores in both quantitative and qualitative evaluations.

An Intelligent System for Predicting Breast Cancer (ISPBC) using a Novel Feature Selection Technique

2025-11-21T20:02:56+07:00

Breast cancer (BC) is becoming a global epidemic, largely affecting women. Breast cancer cases keep climbing steadily. Thus, early detection technologies or systems that notify patients to this disease are essential. Individuals can start treatment for this life-threatening illness, so that patients may be cured or given longer lives. To achieve this, in this study, an expert intelligence system named Intelligent System for Predicting Breast Cancer (ISPBC) was developed. The proposed system utilizes an innovative feature selection technique known as Enriched Feature Set (EFS) in order to identify the most appropriate and significant features. The proposed EFS employs the advantages of heuristic search techniques and stochastic hill climbing to select the most significant and important features. The Decision Tree and Random Forest techniques are employed for breast cancer diagnosis, distinguishing between malignant and benign types. The suggested model’s performance was evaluated by comparing measures such as accuracy, precision, and recall through the utilization of tenfold cross-validation. To measure the efficacy of the suggested model, ISPBC’s performance was compared to that of base classifiers and models published in the literature. A maximum accuracy of 96.09% was attained by ISPBC according to the results.

Scalable and Efficient Student Behavior Prediction using Parallelized Clustering and AHP-weighted KNN

2025-10-09T09:34:42+07:00

This study proposes a scalable and efficient approach for predicting student behaviour in large-scale educational environments. It introduces a parallelized hybrid model that combines Density-Based Optimized K-Means clustering, Analytic Hierarchy Process (AHP) feature weighting, and Hierarchical K-Nearest Neighbours (KNN), implemented using Apache Spark. The main research question is how to improve scalability, accuracy, and computational efficiency of student behaviour prediction when dealing with large, complex datasets. The model addresses key limitations of traditional methods, such as handling heterogeneous data, treating all features equally, and high computational cost. Two main innovations are presented. First, AHP is used to assign structured importance to features, allowing critical factors like attendance and study time to have greater influence on prediction accuracy. Second, clustering and prediction are parallelized using Spark, enabling efficient real-time processing of large datasets. The approach was evaluated using 18,586 student records and more than 20 million behavioural entries. Results show that Hierarchical KNN consistently outperforms standard KNN as dataset size increases. While traditional KNN shows unstable error rates, peaking at 9.4%, Hierarchical KNN maintains lower and more stable errors between 5.16% and 6.08%. Execution time was also significantly reduced through parallel processing, though gains were limited by communication overhead. Overall, the proposed model offers a robust framework for real-time behaviour analysis, academic risk detection, and targeted educational intervention.

Water Filtration Machine with Monitoring System for Aquades Production and Founding an Optimal Pre- treatment Filter Ratio Before Reverse Osmosis Membrane

2025-12-23T09:53:12+07:00

The increasing demand for distilled water (Aquades) in pharmaceutical and medical applications contrasts sharply with the limited quality of municipal water supplies and the high operating costs of commercial Aquades procurement. At the same time, many small-scale facilities still lack integrated systems capable of meeting the Indonesian Ministry of Health standard (Permenkes RI No. 32/2017). Existing research on reverse osmosis (RO) systems largely focuses on membrane or filtration performance, with limited attention to real-time water-quality monitoring and systematic optimization of pre-treatment filters. This study develops an integrated filtration and monitoring system designed to ensure regulatory compliance while optimizing the composition of pre-treatment materials. The system combines silica sand, activated carbon, and zeolite pre-filters with RO, supported by six analog sensors that continuously monitor pH, turbidity, and Total Dissolved Solids before and after filtration. Validation results show high sensor accuracy, with 99.77% for TDS, 98.10% for pH, and 99.97% for turbidity. Among six tested filter compositions, the 25% silica sand-25% activated carbon-50% zeolite configuration achieves the highest average filtration efficiency of 88.96%. These findings demonstrate that optimized pre-treatment combined with real-time monitoring can significantly improve RO performance and support cost-effective Aquades production for medical use.

Examining Performance of Naïve Bayes and Support Vector Machine for Solid Waste Classification in Automated Sorting Systems

2025-12-23T09:55:15+07:00

The growing volume of global waste poses significant challenges to effective waste management, underscoring the need for innovative classification methods to improve recycling efficiency. This study evaluates the performance of two traditional machine learning models, Naïve Bayes and Support Vector Machines (SVMs), for classifying solid waste materials in an automated sorting system. A dataset of 284 JPEG images, categorized into five classes (cardboard, glass, metal, paper, and plastic), was utilized. Preprocessing involved resizing images to 512x384 pixels, normalizing pixel values, and extracting features using Histograms of Oriented Gradients (HOG) and Color Histograms. Naïve Bayes demonstrated computational efficiency with 98.90% accuracy and an F1-score of 0.908, but struggled with overlapping features, leading to misclassifications, particularly between glass and metal. In contrast, SVM outperformed Naïve Bayes, achieving 99.80% accuracy and an F1-score of 0.965 by effectively handling complex, overlapping features via optimal decision boundaries. The findings highlight SVM’s superior performance for complex datasets, while Naïve Bayes remains a viable option for simpler tasks. This study underscores the potential of traditional machine learning in waste classification. However, it suggests that integrating deep learning models could improve accuracy, scalability, and adaptability in real-world waste-sorting systems.