Journal of ICT Research and Applications

Fire Bloom: Energy-Efficient Cluster-Based Routing Protocol for Scalable Wireless Sensor Networks

Tue, 03 Nov 2026 00:00:00 +0700

Wireless sensor networks (WSNs) have been used in smart agriculture, industrial automation, and environmental monitoring applications, which require efficient and reliable routing with energy efficiency. Particularly the selection and operation of the cluster heads (CHs) play significant role in the performance of the network. In this work, we introduce a scalable, reliable and energy efficient cluster-based routing protocol (Fire Bloom), which combines adaptive CH selection and energy-aware routing. The cluster map and the packet flow intensity map even show the communication loads on the base station. The packet delivery ratio (PDR) of Fire Bloom is 97.2%, which is 6–10% better than the competing protocols. Fire Bloom achieved a PDR greater than 91% under a scalability test with 100–500 nodes, which means it can maintain good performance as the density of the network becomes denser. Energy usage is roughly proportional to the traffic volume but lower than previously reported values. Near-linear scalability at runtime is also proposed: increasing the number of nodes from 100 to 1,000 took from 1.2 s to 14.1 s. In general, Fire Bloom lowers overall end-to-end delay, increases load balancing across CHs, boosts energy efficiency, and it is resilient to changes in network conditions and maintains successful packet delivery. These results prove that it is capable of scalable and energy efficient deployments of WSN’s

Simulated Feature Fusion for Parkinson’s Disease Prediction Using Non-aligned Public Handwriting and Voice Datasets

Tue, 03 Nov 2026 00:00:00 +0700

Parkinson’s disease (PD) is a progressive neurodegenerative disorder that affects over 10 million people worldwide, with prevalence increasing due to aging populations. Early symptoms such as tremor, micrographia, bradykinesia, and voice changes often remain undetected until substantial neuronal loss has occurred, leading to irreversible damage and increased treatment costs. This paper presents a comparative study of five machine learning models—Logistic Regression, Random Forest, Support Vector Machine (RBF Kernel), K-Nearest Neighbors, and XGBoost—for Parkinson’s disease prediction using handwriting and voice-based features. Performance was evaluated using accuracy, precision, recall, F1-score, AUC, 5-fold cross-validation, ROC curve analysis, feature importance analysis, and SHAP-based explainability.XGBoost achieved the highest test-set F1-score (0.7407) and AUC (0.8329), while SVM achieved the highest average cross-validation F1-score. Decision-level fusion further improved recall (0.88) while maintaining competitive overall performance. K-Nearest Neighbors produced the lowest overall performance, while XGBoost achieved the best balance between classification accuracy, F1-score, and AUC.

Determinants of Cloud Storage Adoption in Indonesia: Integrating Positive Valence, Security, and Trust

Adele B L Mailangkay, Bagas Aditya Nugraha — Tue, 03 Nov 2026 00:00:00 +0700

Cloud storage adoption in Indonesia remains under-researched, and prevailing technology-adoption models generally assume that user trust is a robust antecedent of adoption intention. Drawing on the positive Valence Framework, this study integrates two positive utility constructs—perceived ubiquity and perceived benefits—with security and trust to explain behavioral intention to adopt cloud storage. Survey data from 202 digitally literate Indonesian internet users were collected through Google Forms and analyzed using partial least squares structural equation modeling (PLS-SEM); common method bias was assessed using Harman’s single-factor test and a full collinearity (VIF) assessment. Perceived benefits exerted the strongest effect on behavioral intention, followed by security and perceived ubiquity, whereas trust had no significant effect. Theoretically, the non-significance of trust indicates that, in an ecosystem dominated by a few globally legitimate providers embedded in users’ device platforms, trust operates as an assumed baseline rather than a differentiating criterion. This reframes trust from a direct antecedent to a boundary-conditioned, hygiene-type factor and refines the applicability of Valence- and TAM/UTAUT-based adoption models to developing digital markets.

AI-based Efficient Hallway Attendance Monitoring using Confidence-driven Face Recognition

Tue, 03 Nov 2026 00:00:00 +0700

Attendance monitoring deployed at entrance areas offers reliable tracking for institutional environments. We propose an efficient face-recognition-based framework combining YOLOv11-based face detection, motion-based tracking, and FaceNet-based recognition. Our key contribution is a confidence-driven face selection method that collects multiple face images per track and selects the highest-quality image for recognition. Experiments on real-world 60-second hallway footage (1,800 frames at 30 FPS) demonstrated up to 77.7% improvement in processing speed (79.36 FPS vs. 44.66 FPS baseline) while maintaining equivalent recognition accuracy. Collection windows of 5 to 7 images provide optimal balance between latency and throughput (73.67–75.86 FPS), demonstrating that strategic recognition frequency reduction enables real-time attendance monitoring suitable for resource-constrained environments

Hybrid and Classical Machine Learning Models for Mixed-crop Recommendation in Precision Agriculture: An Empirical Performance Analysis

Ahmed Mohammed Gimba, Pradeep Kumar Mishra — Tue, 03 Nov 2026 00:00:00 +0700

Precise mixed-crop recommendation is essential for increasing agricultural productivity, improving land utilization, and supporting sustainable farming. Effective crop combinations depend on several environmental and soil factors, making accurate recommendations a challenging task. This study compared classical and hybrid machine learning approaches for mixed-crop recommendation using an indigenous dataset from Andhra Pradesh, India. The dataset consisted of seven key features: nitrogen, phosphorus, potassium, temperature, humidity, pH, and rainfall. The study evaluated Support Vector Machine (SVM), K-Nearest Neighbor (K-NN), Decision Tree (DT), Logistic Regression (LR), Random Forest (RF), and Extra Trees (ET), alongside a proposed hybrid Random Forest-Extra Trees (RF-ET) model. Performance was measured using accuracy, precision, recall, F1-score, k-fold cross-validation, and confusion matrix analysis. The results show that the proposed hybrid RF-ET model achieved the best performance, obtaining 95.91% accuracy, 95.98% precision, 95.91% recall, and 95.94% F1-score. The model also demonstrated strong robustness through cross-validation. These findings indicate that hybrid machine learning can significantly improve the accuracy and reliability of mixed-crop recommendation systems, providing farmers with better decision-support tools for sustainable and efficient agricultural production.

Smart Leaf Identification and Analysis using Artificial Intelligence

Balamurali Ravichandran, Abhinav Tiwari, Joydeep Roy — Tue, 03 Nov 2026 00:00:00 +0700

This paper describes a leaf classification framework based on machine learning methodologies. The dataset included 36 leaf species with 340 instances and 16 features. Decision Tree, Logistic Regression, Support Vector Machine, XGBoost, AdaBoost, and Random Forest were applied initially on the raw dataset. It was found that Decision Tree, XGBoost and Random Forest performed well, with accuracy scores of 53%, 60% and 85% respectively. After feature reduction based on Information Gain and correlation matrix analysis, four highly correlated features were removed. This led to a marginal improvement in accuracy to 60% for Decision Tree and to 65% percentage for XG Boost, while the accuracy of Random Forest dropped to 74%. We also applied SMOTE (Synthetic Minority Oversampling Technique) to balance the raw dataset. This enhanced the accuracy of Random Forest and XGBoost to 91% and 81% respectively. With Computer Vision, we achieved an accuracy of 66% for SVM, 61% for Random Forest, 43% for K-NN and 37% for the Decision Tree classifier on the image dataset. The pre-trained VGG16 model achieved an impressive 90.47% accuracy, outperforming the custom CNN model and conventional models like Linear Discriminant Analysis (LDA) and KNN with various K-values from the literature

Adverse Drug Reaction Detection with Drug Prediction Using BERT–LSTM–SVM

Spandana Shivanadhuni, R. Vijaya Prakash — Tue, 03 Nov 2026 00:00:00 +0700

Social media can be a useful resource for assessing patient-reported drug experiences. Adverse drug reactions (ADRs) are an important public health issue, but finding social media posts in which adverse reactions are described is a difficult task, since not all posts are related to ADR. The proposed work in this research is a hybrid model of BERT-LSTM-SVM for detection of ADRs and drug classification. The input is preprocessed as follows: the text is cleaned, tokenized, normalized, and noise is removed. Bidirectional Encoder Representations from Transformers (BERT) is used to produce contextual representations of the text, and a Long Short-Term Memory network (LSTM) is used to model sequential dependencies to aid binary classification of ADR and non-ADR text. Then, a Support Vector Machine (SVM) classifier is used to classify the associated medicine based on the extracted textual features. In our validation of the method, the accuracy, sensitivity, specificity and precision were 0.99 with a recall. The proposed framework combines contextual language representations, sequential modeling, and supervised classification approaches to facilitate the automatic detection of ADRs and medicine identification.

Multi Objective Optimization of Distribution in Supply Chains using Green AI via an Enhanced Hippopotamus Algorithm

Tue, 03 Nov 2026 00:00:00 +0700

Modern supply chains require innovative, sustainable, and computationally efficient solutions for physical distribution. This paper proposes the Enhanced Hippo Optimization (EHO) algorithm, a novel, nature-inspired metaheuristic designed to solve the Multi-Objective Vehicle Routing Problem (MO-CVRP). The (EHO) algorithm mathematically models three key hippo herd behaviors: exploration (herd positioning), defensive coordination, and strategic withdrawal, to achieve a dynamic balance between global exploration and local exploitation. By integrating real-world geospatial data from the Google Maps Directions API and customer demand modeling via a Zipf distribution (α = 0.9), (EHO) simultaneously reduces total distance, delivery time, operating costs, and CO₂ emissions. The algorithm was evaluated under twenty scenarios across a network of nineteen cities in Egypt, outperforming six advanced algorithms Particle Swarm Optimization (PSO), Genetic Algorithm (GA), Ant Colony Optimization (ACO), RUN, Artemisinin Optimizer (AO), and Fata Morgana in eighteen cases. The framework achieved an average reduction of 26.9% in distance and emissions, with total computational runtime ranging from 12.77 to 25.1 seconds depending on the scenario. Visual analyses, including optimized route maps and radar performance charts, confirmed the robustness, scalability, and practical effectiveness of the EHO system in transforming logistics into a driver of economic and environmental sustainability.