http://ir.haramaya.edu.et//hru/handle/123456789/10 2026-06-19T09:28:20Z 2026-06-19T09:28:20Z Suleiman Mohamed Abdi Wondwossen Mulugeta (Ph.D) Faizur Rashid (Ph.D) http://ir.haramaya.edu.et//hru/handle/123456789/8485 2026-05-28T06:33:31Z 2024-11-01T00:00:00Z

A DEEP LEARNING APPROACH FOR CLASSIFICATION OF SIX CASES OF CHEST X-RAYS Suleiman Mohamed Abdi; Wondwossen Mulugeta (Ph.D); Faizur Rashid (Ph.D) Artificial Intelligence (AI), particularly deep learning, is transforming healthcare by enabling automated analysis and diagnosis from medical images, addressing critical challenges such as the shortage of radiologists and the demand for accurate diagnostic systems. This study focuses on the classification of six distinct chest X-ray conditions: Normal, Pneumonia, Tuberculosis, Lung Mass, Rib Fracture, and Enlarged Heart. Using a dataset of 10,200 chest X-ray images collected from Hargeisa Group Hospital, two pretrained convolutional neural network (CNN) architectures, DenseNet and GoogleNet, were fine-tuned and evaluated. Comprehensive preprocessing, including noise removal, image enhancement, and augmentation techniques, ensured high-quality and balanced training data. The models demonstrated exceptional performance, achieving classification accuracies of 97% and 96%, respectively, surpassing benchmarks in multi-class medical image classification Despite these promising results, the study encountered limitations. The dataset size, while sufficient for this research, remains relatively small for broader generalizability. Additionally, processing sensitive personal data required compliance with Somaliland’s Data Protection Act, posing challenges in accessing and utilizing X-ray images. These limitations highlight the need for expanded datasets and improved data access protocols for future research.This research establishes a robust framework for automating chest X-ray diagnostics, empowering radiologists with timely and accurate decision support. The findings contribute to advancing AIdriven solutions for healthcare, addressing both global and region-specific challenges 83p.

2024-11-01T00:00:00Z Jemal Abdela Million Meshesha (PhD) Jemal Abate (MSc) http://ir.haramaya.edu.et//hru/handle/123456789/8465 2026-05-25T06:25:56Z 2025-03-01T00:00:00Z

Automatic Question Generation from Afaan Oromo Text UsingDeepLearning Jemal Abdela; Million Meshesha (PhD); Jemal Abate (MSc) In the modern digital era, the availability of electronic content in multiple languages has grownconsiderably. Nevertheless, generating questions from these materials remains a labor-intensiveand time-consuming process, particularly for low-resource languages like Afaan Oromo. Whilesignificant progress has been made in Automatic Question Generation (AQG) for languages suchas English, Chinese, Amharic and Somali. there is a notable lack of technology with NLPonAfaan Oromo QG because Afaan Oromo has the unique linguistic characteristics. This studybridges this gap by developing an AQG system for Afaan Oromo using deep learning models, specifically LSTM, Bi-LSTM, and GRU. To build the model a dataset of 5,000 paragraph-question-answer triples was meticulouslycurated from Afaan Oromo textbooks and educational resources. The dataset underwent preprocessing steps such as tokenization, normalization, and word embedding using Word2Vec. A deep learning model with an attention mechanism was employed to generate questions bydesign science research methodology, Among the models evaluated, the Bi-LSTMdemonstratedthe highest performance, achieving a training accuracy of 95.3% and a validation accuracyof 92.5%. The LSTM model also performed well, with a training accuracy of 92.31%and a validation accuracy of 91.02%, while the GRU model showed marked improvement after hyperparameter tuning, reaching 88.0% training accuracy and 87.0% validation accuracy.The results indicate that the Bi-LSTM model is the most effective for generating both factoidandnon-factoid questions in Afaan Oromo. Future research should explore transfer learning frompre-trained models, expand the dataset through collaborations with educational institutions, andintegrate advanced neural architectures like Transformers to further enhance performance andquestion quality. This study makes a significant contribution to the field of Natural LanguageProcessing (NLP) by pioneering AQG for Afaan Oromo, use as input for QA system, providingafoundation for future research and practical applications in education and language preservation. 138p.

2025-03-01T00:00:00Z Jemal Abdela Million Meshesha (PhD) Jemal Abate (MSc) http://ir.haramaya.edu.et//hru/handle/123456789/8464 2026-05-25T06:21:18Z 2025-03-01T00:00:00Z

Automatic Question Generation from Afaan Oromo Text UsingDeepLearning Jemal Abdela; Million Meshesha (PhD); Jemal Abate (MSc) In the modern digital era, the availability of electronic content in multiple languages has grownconsiderably. Nevertheless, generating questions from these materials remains a labor-intensiveand time-consuming process, particularly for low-resource languages like Afaan Oromo. Whilesignificant progress has been made in Automatic Question Generation (AQG) for languages suchas English, Chinese, Amharic and Somali. there is a notable lack of technology with NLPonAfaan Oromo QG because Afaan Oromo has the unique linguistic characteristics. This studybridges this gap by developing an AQG system for Afaan Oromo using deep learning models, specifically LSTM, Bi-LSTM, and GRU. To build the model a dataset of 5,000 paragraph-question-answer triples was meticulouslycurated from Afaan Oromo textbooks and educational resources. The dataset underwent preprocessing steps such as tokenization, normalization, and word embedding using Word2Vec. A deep learning model with an attention mechanism was employed to generate questions bydesign science research methodology, Among the models evaluated, the Bi-LSTMdemonstratedthe highest performance, achieving a training accuracy of 95.3% and a validation accuracyof 92.5%. The LSTM model also performed well, with a training accuracy of 92.31%and a validation accuracy of 91.02%, while the GRU model showed marked improvement after hyperparameter tuning, reaching 88.0% training accuracy and 87.0% validation accuracy The results indicate that the Bi-LSTM model is the most effective for generating both factoidandnon-factoid questions in Afaan Oromo. Future research should explore transfer learning frompre-trained models, expand the dataset through collaborations with educational institutions, andintegrate advanced neural architectures like Transformers to further enhance performance andquestion quality. This study makes a significant contribution to the field of Natural LanguageProcessing (NLP) by pioneering AQG for Afaan Oromo, use as input for QA system, providingafoundation for future research and practical applications in education and language preservation. 138p.

2025-03-01T00:00:00Z Feyisa Abebe Abebe Belay Adege (PhD) Mr. Tadesse Kebede http://ir.haramaya.edu.et//hru/handle/123456789/8419 2026-05-20T07:28:23Z 2025-04-01T00:00:00Z

MACHINE LEARNING-BASED PREDICTION OF UNDER-FIVEMORTALITY USING HEALTH,SOCIO-DEMOGRAPHIC,ANDCLIMATE DATA Feyisa Abebe; Abebe Belay Adege (PhD); Mr. Tadesse Kebede Health is a state of full well-being and a cornerstone of international development, withconsiderable investments made over the last three decades to reduce morbidity and mortality. Under-five mortality, which is generally defined as the death of children under the age of five, is still a critical public health challenge in developing countries. The study here proposesamachine learning-based approaches for predicting under-five mortality using health, socio-demographic, and climate data from Eastern Hararghe, Ethiopia. The data used in this study were collected from the Hararghe Health DemographicSurveillance System and Ethiopian National Meteorology Agency. The followingeight supervised machine learning algorithms were considered: Naïve Bayes (NB), Support VectorMachine (SVM), K-Nearest Neighbors (KNN), Decision Tree (DT), RandomForest (RF), eXtreme Gradient Boosting (XGBoost), Attentive Tabular Network (TabNet), andConvolutional Neural Network (CNN). The proposed framework covers data preprocessing, exploratory data analysis, model training, prediction, performance evaluation, andidentification of the key determinants of under-five mortality. It was observed that an80:20split produced an optimum performance in the models. Preprocessing techniques werethenapplied to enhance data quality before training the machine learning models. There weretwoexperimental setups: one with a data-balancing technique and the other without.Results indicated that balanced datasets always outperformed. Amongst all the modelsdeveloped, the XGBoost recorded the highest accuracy, having testing accuracy scoreof97.9%, precision of 98%, F1-score of 98%, and recall of 98%. The determinants of under-fivemortality identified in this study were antenatal care, child gender, wealth index, total numberof alive children, preceding child alive, physical healthy ,birth place and weight of baby. Intheend, the XGBoost algorithm emerged as the best among other models, proving to be themost reliable predictive model for under-five mortality. This study has shown the potential ofmachine learning providing helps in tackling critical public health challenges by leveragingdiverse datasets to enhance decision-making and interventions. Household-level climatedatawere not utilized in this thesis, which would be taken into account by future researchers. 129p.

2025-04-01T00:00:00Z