ECTI Transactions on Computer and Information Technology (ECTI-CIT)

AI-Based Smart Identification of Medicinal Plants Using Vision Transformer and CatBoost for Biodiversity and Healthcare

Faisal Firdous — 2025-12-06

In most countries, medicinal plants are crucial remedies for disease treatment. Even though the majority are edible, ingesting the incorrect herbal plant can have fatal consequences. It is essential to accurately identify these plants not only for safe usage by individuals but also for various real-time applications like aiding biodiversity conservation, supporting farmers in recognizing local herbs, and also preserving indigenous systems. Numerous automatic methods for identifying medicinal plants have been developed; however, most of them are severely limited, either by the relatively small number of plant species they support or by the fact that they rely on manual visual segmentation of plant leaf surfaces. This means that instead of being easily recognized in their natural environments, which frequently include complicated and chaotic backgrounds, they are snapped against a plain background. Deep learning-based techniques have advanced significantly in recent years. Still, they are trained on data that isn't always fully reflective of the intra-class and inter-class variances among the plant species in consideration. The paper approaches this issue by integrating the hybrid model of a pre-trained vision transformer with a CatBoost classifier tuned with Optuna. The vision transformer model is trained with the Indian medicinal plant dataset with the five most commonly used species. The hybrid model is compared with the deep learning models regarding precision, recall, F1-score, accuracy, and execution time on the same dataset. Our proposed model achieves a training phase accuracy of 93%, which shows the improvement for automating the identification of medicinal plants. In conclusion, our proposed hybrid model reveals enhanced accuracy, improved reliability, and reduced false positives in automating the identification of medicinal plants, contributing effectively to healthcare applications and biodiversity.

A Hybrid GloVe-BERT Fusion Model with Multi-Level Attention-Based CNN-BiLSTM for Sentiment Analysis

Yashaswini Iyer — 2025-12-27

Gauging public sentiment toward climate policy from information-rich news headlines remains challenging for conventional text classification approaches. Conventional sentiment analysis tools miss contextual subtleties in brief headlines, whereas deep learning models capture the public perception more accurately. The proposed work presents hybrid Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) model that uses combination of GloVe and BERT embeddings with an attention layer for sentiment analysis of climate change news headlines. The novelty of this research lies in the use of GloVe and BERT embeddings through a multi-stage fusion strategy and an attention mechanism to enhance text classification performance in a hybrid model. The architecture employs a hierarchical layering approach to fuse static GloVe embeddings with dynamic, contextualized BERT representations through attention modules that enables the network to selectively focus on salient features. To further model complex semantic dependencies, the design incorporates parallel CNN-BiLSTM branches, structured with residual connections and bolstered with additional layers of attention. Evaluated on 1,023 climate-related headlines annotated on a three-point polarity scale, the proposed model achieves an accuracy of 80.47%, outperforming classi- cal baselines (SVM, Naive Bayes, K-NN) and single branch deep networks (CNN:78.63%, BiLSTM:78.36%). The predictive accuracy of the hybrid model is evaluated using a paired t-test to determine whether the difference between models is statistically significant; this is confirmed by rejecting null hypothesis and accepting alternate hypothesis.This study demonstrates that compact, domain-adaptive deep learning models incorporating contextual embeddings and attention mechanisms that can effectively extract sentiment from news headlines, offering scalable, evidence based tools for tracking climate discourse and information policy decisions.

A Language-Adaptive Ensemble Clustering Framework for Emotion Detection in Multilingual Social Media Text

Wafa Saadi — 2025-12-27

Social media platforms generate vast streams of emotionally rich textual data, offering valuable opportunities for critical applications, including mental health assessment and the analysis of collective public sentiment. However, detecting emotions in noisy and multilingual content remains challenging, particularly for under-resourced varieties such as dialects. Moreover, supervised learning techniques strongly depend on the availability of manually annotated corpora, whose creation requires substantial human effort and domain expertise. In contrast, unsupervised methods, while avoiding the need for human intervention, often lack sufficient robustness when confronted with the variability and complexity of natural language across diverse linguistic and cultural contexts. We present an ensemble clustering framework that automatically generates emotion labels from Twitter data, without human intervention. Our approach incorporates three emoji-handling strategies in the preprocessing step, enabling diverse semantic representations of emojis. We applied the BERT embeddings combined with PCA for dimensionality reduction within the same experimental framework. An ensemble clustering strategy integrating K-Means, Agglomerative clustering, and Gaussian Mixture Models (GMM) is adopted using multiple ensemble configurations. Experimental evaluation conducted on 10017 English tweets and 4134 Arabic tweets demonstrates that the proposed method achieves a silhouette score of 0.808 on English data using K-Means with Agglomerative and K-Means with GMM ensemble configurations. For Arabic data, silhouette scores of 0.728 and 0.718 are obtained using English and Arabic keywords, respectively. Emoji semantic integration enhances ensemble clustering performance, suggesting its importance for contextual disambiguation. The proposed framework provides a scalable solution for emotion detection in low-resource languages, enabling language-aware applications in multilingual contexts, particularly within linguistically diverse and multilingual populations

Edge-to-Cloud Long Short-Term Memory Model for Ambient Carbon Monoxide Level Prediction

Alauddin Maulana Hirzan — 2026-01-03

Carbon monoxide (CO) is a harmful gas from incomplete fuel combustion, often found in motor vehicle emissions. Prolonged exposure can cause serious health issues or death. While existing Internet-of-Things (IoT) systems monitor CO levels, most lack predictive capability. One prior study used an Artificial Neural Network with limited accuracy (79%). To address this, a new IoT-based CO prediction model is proposed using a Long Short-Term Memory (LSTM) algorithm. The model predicts future CO concentrations based on seasonal patterns, empowering users to anticipate and proactively respond to potential exposure. By leveraging Edge-to-Cloud architecture, this approach enables low-power edge devices to send data to the cloud for accurate forecasting without local model deployment. Based on the evaluation, the model achieved 98.42% accuracy, outperforming previous approaches by 19.42%. It also showed superior performance against other algorithms, with the lowest MAE (0.026305), MSE (0.016004), RMSE (0.126506), and the highest R² (0.997647). Evaluation with AIC and BIC confirmed its reliability, scoring zero after MinMax scaling. The model demonstrates a substantial advancement in predictive CO monitoring, giving users actionable insights to protect health and safety.

A Machine Learning Approach for Multi-Label Classification in Candidate Election Social Media Analysis

Herman Yuliansyah — 2026-01-10

Multi-label text classification in social media comments presents a signi icant challenge in natural language processing. Several previous studies have conducted sentiment analysis on candidate (presidential and guber- natorial) elections using machine learning approaches. However, an opinion can contain more than one category or label simultaneously, such as sentiment, candidate, or certain issues. This study proposes a multi-label classification model to improve accuracy, addressing challenges such as complex language structure, non-standard word usage, and imbalanced data. The proposed model is compared with three popular classica- tion algorithms: Naive Bayes (NB), Support Vector Machine (SVM), and K-Nearest Neighbours (KNN), for handling multi-label text classification tasks. The proposed model comprises a classification pipeline that includes data preprocessing, feature extraction using TF-IDF, and the integration of the GridSearchCV technique to enhance algorithm performance and effectiveness. The evaluation is conducted using multi-label metrics such as Precision, Recall, and F1-Score. The experiment results showed that SVM with GridSearchCV provided the best performance in terms of precision and generalization on the gubernatorial election dataset. SVM + GridSearchCV yielded scores of 97.4% and 99.2% for candidate labels, and 99.2% and 99.0% for sentiment labels. While NB and KNN also showed improvements, their performance was not as significant as SVM. NB outper- formed in computational performance, whereas KNN demonstrated poor performance on high-dimensional data.

A Framework for Geospatial Speed Test Platform using Holistic Approach

Norrarat Wattanamongkhol — 2026-01-10

This study proposes a framework for a geospatial internet speed test platform, specifically designed to assess the broadband network experience in Thailand. The novelty of this research lies in its comprehensive approach to broadband internet quality assessment, uniquely measuring internet performance from ISPs sources to end-users using custom-designed reTerminal hardware devices based on Raspberry Pi 4. Tests were conducted involving four major broadband service providers. Data were collected from 30 speed test devices, which were installed at the source of each service provider network, and 200 internet user devices. The total number of data records collected was 25,000. The results indicated that the Download Percentage Average was 64.30%, while the Upload Percentage Average was 71.84%. The average latency and jitter were 11.82 ms and 16.53 ms, respectively. All speed test parameters at the source of the ISPs network were found to reach almost 100% compared to the speed test network devices of the ISPs. This means that ISPs provide internet quality according to the standards set by the NBTC. These results can assist the NBTC for setting relevant policies or strategies to improve the quality of broadband internet services at user's sites.

Ontology Generation and Instance Extraction of Medicine Information from Thai Semi-structured Data

Taneth Ruangrajitpakorn — 2026-01-10

An ontology is a widely used knowledge base for representing domain knowledge. Developing a knowledge-representing ontology is difficult, as it requires both domain and engineering expertise. Yet, such ontologies are essential for enabling intelligent systems to comprehend real-world knowledge through structured concept networks. In the Thai context, ontology research remains limited due to the scarcity of structured resources, standardized schemas, and annotated corpora for automatic knowledge extraction. This study addresses this gap by proposing a pattern-based methodology for ontology generation and instance extraction from Thai semi-structured medicine data, providing an alternative to resource-intensive deep-learning methods. The proposed approach identifies patterns of collocated Thai text and builds a collocation tree of word sequences, in which shared sequences represent ontological properties and variable sequences represent instance values. The method was applied to two complementary Thai medicine datasets, namely I-Med (a hospital dispensing-record database) and Pobpad (a public health-information website), to generate and integrate ontology components. These templates were transformed into ontological properties and converted into RDF/OWL format to produce a standard ontology usable for querying and reasoning. The generated ontology achieved high performance (Precision = 0.97, Recall = 0.90, F1 = 0.91) and received favorable assessments from domain experts. The results indicate that the proposed approach can effectively extract structured knowledge from Thai semi-structured text and produce a reliable ontology suitable for medical knowledge representation, providing a data-driven foundation for future Thai intelligent systems.

Ripeness Evaluation Using Near-Infrared (NIR) Spectroscopy and NLP for Interpretability

Panudech Tipauksorn — 2026-01-10

This study develops a non-destructive avocado ripeness classification system using a low-cost, portable near-infrared (NIR) scanner and machine learning. Traditional ripeness assessment methods are often destructive and subjective, limiting their efficiency in agricultural practices. To address this issue, we developed a custom NIR scanner capable of capturing spectral information across 18 discrete wavelength bands for avocado ripeness classification. The research focuses on Buccaneer avocados sourced from the Royal Project, with samples collected from both the Royal Project Gardens and Sorting Plant. A total of 120 kg of avocados were systematically sampled and categorized by agricultural experts into three ripeness stages: raw, ripe, and aged. This study applies Multiplicative Scatter Correction (MSC) to preprocess NIR spectra, enhancing feature separation before machine learning model training. This study evaluates three classification models: Random Forest, XGBoost, and the Gaussian Mixture Model (GMM). Random Forest achieved the highest classification accuracy (78%) with an AUC score of 0.93, followed by XGBoost (74% accuracy, AUC 0.91). GMM performed worse, with 42% accuracy and an AUC of 0.58. Additionally, Natural Language Processing (NLP) was applied to convert model predictions into human-readable ripeness descriptions, assisting farmers in decision-making. This study demonstrates that low-cost NIR technology combined with AI-driven analysis enables efficient, non-destructive classification of avocado ripeness.

Enhanced Case-Based Reasoning Framework with Weighted Jaccard Similarity for Malnutrition Diagnosis in Toddlers

Joko Handoyo — 2026-01-17

Malnutrition is a serious condition caused by nutrient deficiency that poses a high risk to toddler growth and development, potentially leading to long-term health problems or even death if left untreated. Early detection of malnutrition symptoms is crucial to enable prompt and appropriate medical interventions. This study aims to develop an expert system capable of diagnosing malnutrition diseases quickly, accurately, and efficiently, particularly as a knowledge-based decision support tool in toddler healthcare. The method used is Case Based Reasoning (CBR), which applies experiences from previous cases to solve new ones. The system processes data consisting of 22 symptoms and 8 types of malnutrition diseases, supported by a database of 22 real cases. Each symptom is associated with the likelihood of a disease based on its similarity to previous cases. Performance evaluation results show an accuracy of 80% and a sensitivity of 85.7%, indicating that the system is fairly reliable in recognizing positive cases (REUSE) and providing appropriate diagnoses. In conclusion, the CBR- based expert system can serve as an effective diagnostic aid for medical personnel in quickly identifying malnutrition in toddlers, thereby supporting more efficient and targeted decision-making.

Image Processing-Based Machine Learning for Multi-Class Apple Bruise Classification

Daranat Tansui — 2026-01-17

This research develops an automated apple bruise detection system that combines digital image processing and machine learning to enhance quality control. The study employs preprocessing techniques (resizing, grayscale conversion, thresholding) and extracts key features including pixel count, mean intensity, maximum and minimum pixel values, and estimated bruise area. A dataset of 513 apple samples was created and divided into training (70%), validation (15%), and test (15%) sets. Among the five evaluated classifiers, Decision Tree and Gradient Boosting demonstrate identical peak performance (99.35% accuracy, 0.9941 F1-score), with Decision Tree offering superior computational efficiency (200 times faster training). Random Forest achieves 98.70% accuracy, outperforming conventional methods (SVM, KNN, Logistic Regression). Notably, the Decision Tree accurately classifies severe bruises (Class 2), which is crucial for quality assurance. The system's effectiveness is validated through comprehensive metrics, including confusion matrices and ROC analysis. These results highlight the practical viability of implementing Decision Tree-based solutions in commercial fruit grading systems, offering an optimal balance between accuracy (99.35%) and operational efficiency (0.0012 seconds of training time). The findings enhance automated post-harvest inspection capabilities while addressing critical industry needs for rapid and reliable bruise detection.

MelodyCraft: A Prompt-Based Modular AI Framework for Synchronized Lyrics and Instrumental Music Generation

Ayushi Chauhan — 2026-01-24

Current artificial intelligence (AI) music generation systems are usually not synchronized with lyrics and melodies, emotional correlativity, or objective assessment criteria. To overcome these weaknesses, this study proposes MelodyCraft, a prompt-based modular AI architecture that combines MusicGen-small to produce instrumental music and a Mixtral transformer to create genre- and emotion-specific lyrics. The fine-tuning of QLoRA-based optimization was performed using 28,000 prompt-lyric pairs with a validation loss of 1.85 and a BLEU score of 0.65, which indicates high lyrical coherence and style fidelity. A spectral analysis of 10 human-composed and 10 AI-generated music of pop, rock, and jazz types showed no statistically significant (p > 0.05) spectral centroid, bandwidth, or roll-o differences among them, which is why it was deemed to have almost human acoustic realism. In addition, a human listening experiment involving 16 subjects found average Mean Opinion Scores (MOS) of 3.9-4.0 in melody quality, lyrical coherency, and emotional connection, creating perceptual parity with human compositions. MelodyCraft is a full-stack multimodal music-generating system that generates semantically, rhythmically, and emotionally consistent music and can serve as a guide for future creative systems assisted by AI.

A Hybrid of Modified Capsule and Transformer Model for Sepsis Diagnosis

Tuan Anh Vu — 2026-01-24

Sepsis is a critical and urgent medical condition that imposes a global health burden due to high mortality and the risk of long-term disability without prompt treatment. In this study, we propose a novel hybrid modified capsule and a transformer encoder (CaT) using a selected subset of biomarkers for the diagnosis of sepsis. The biomarkers are identified by a dual selection strategy that combines the differential expression analysis of immune-related genes with the Boruta algorithm using a random forest model. The modified capsule network consists of 4 parallel capsule layers, each implemented as a feedforward unit comprising a linear transformation followed by ReLU activation. On the validation set using Leave-One- Dataset-Out Cross-Validation, the CaT model shows better performance compared to other machine learning and deep learning models, with an accuracy of 96.8%, sensitivity of 98.0%, specificity of 87.9%, Mathews correlation coefficient of 85.6%, and area under curve of 98.0%. These findings highlight the robustness, generalization, and effectiveness of the proposed CaT model, demonstrating its potential as a reliable tool for the prediction of sepsis in clinical practice.