Journal of Computational Science and Data Analytics

Predictive Machine Learning Model for Low Birth Weight in Newborns

Serkalem Negusse Kassaye, Abebe Alemu Balcha, Kula Kakeba, Samuel Alemu — Mon, 30 Dec 2024 00:00:00 +0000

Low birth weight (LBW) occurs when a newborn weighs less than 2500 grams regardless of the gestational age of the infant. LBW is one of the indicators of complex public health problems of an infant, which is 25 times more likely to die than those at expected birth weight. The neonatal mortality rate measures the number of neonates dying before reaching 28 days of age per 1,000 live births within a given year. It affects one out of every seven newborns, about 14.6 percent of babies born worldwide. The prevalence is 7.2% in developed regions and 13.7% in Africa. The neonatal mortality rate in Ethiopia is 29.524 deaths from 1000 live births in 2023. Thus, building accurate LBW prediction models and finding the related risk factors is critical. The early identification and prediction of such disease would reduce premature death rate caused by LBW. The exponentially increasing data availability on the web has a significant role in extracting better insight, and it is vital to develop machine learning models. The current technology facilitates and supports professionals in health care by providing services to societies. Classification techniques help to classify the case according to a certain feature from data and to predict the probabilities of LBW in infants. We have used a public dataset obtained from the Ethiopia Demographic Health Survey to build models. In this work, artificial neural networks and support vector machines are used for training and model building. Moreover, data preparation techniques like scaling and transformation, hyperparameter tuning approaches like GridSearch CV, and feature engineering techniques have been employed and tested. Based on performance evaluation results, the proposed classifiers were capable of predicting of LBW. The neural network with hyperparameter tuning techniques scored an accuracy of over 97.2% with 98.0% of sensitivity, 97.0% of precision, 96.0% of specificity, and an area under the curve (AUC) of 99%.

Human Activity recognition using machine learning

Rediet Desalegn, Hana Mekonen, Melkamu Hunegnaw — Mon, 30 Dec 2024 00:00:00 +0000

Human activity recognition is a broad field of study concerned with identifying the specific movement or action of a person based on sensor data. It can improve people's well-being by automatically assessing and summarizing their daily activities. In this study an automatic detection of nine different human activities from body sensor data was developed. The human activity data was collected from 30 between the age of 12 and 55 to provide data. from those 17 participants were male and 13 participants were female. Participants completed nine activities (sitting, standing, walking, running, upstairs, downstairs, sit-up, jumping and cycling) while wearing an I-phone eight plus in the pocket. Records from 18 individuals used for training, 6 participants data was used for validation and the remaining 6 participants records were used for testing. The human activity data is captured by via phone's integrated acceleration and gyroscope sensors. Hence a total of nine-dimensional data (triaxial accelerometer data (linear acceleration and body acceleration) and triaxial gyroscope data) was acquired using these two sensors. In all case the sample rate used to capture the data was 50Hz. In this study both classical machine learning and deep learning methods were exported. In classical techniques four machine learning classifiers (SVM, KNN, logistic regression and decision tree) have been used to categorize the data. For this to work 18 set of features were carefully computed. Three deep learning models (CNN, LSTM and hybrid CNN-LSTM) have been also trained in the accurate detection of human activity classification.

Improving Data Security in Cloud Storage Systems Using Hybrid Algorithms with Integrity Verification

Erku Kifle Dessie , Asrat Mulatu Beyene — Mon, 30 Dec 2024 00:00:00 +0000

In a cloud system, storage services allow users to store data on the Internet. However, storing data in the public cloud increases the risk of loss, interception, modification, and manipulation by unauthorized users. So, users need to protect their data by applying security mechanisms. In this work, we proposed an efficient hybrid algorithm along with data slicing and integrity verification to improve user-side cloud data security. The proposed hybrid algorithm combines multiple symmetric and asymmetric algorithms to improve both performance and security. It takes the advantages of both to compensate for their weaknesses. The asymmetric algorithm is used for encrypting the symmetric keys, whereas, the symmetric algorithms are used to encrypt and decrypt the data. The proposed framework works by splitting the data into chunks and encrypting each portion separately. File information such as type of algorithm used, hash value, and secret keys are kept at the user's side enabling only the encrypted data to be sent to the cloud. The proposed hybrid algorithm is evaluated and compared against the-state-of-the-art. The results show that the proposed hybrid algorithm outperformed existing ones in terms of throughput, and running time while achieving better degree of data security.

Hate Speech Detection from Transliterated Amharic Social Media Comments Using Machine Learning and Deep Learning Approaches

Zeleke Abebaw, Andreas Rauber , Solomon Atnafu — Mon, 30 Dec 2024 00:00:00 +0000

The rise of transliterated script usage on social media has presented significant challenges to hate speech detection models, as such scripts often bypass models trained exclusively on formal language datasets. Existing Amharic hate speech detection studies predominantly focus on datasets written in formal Amharic scripts using machine learning approaches, leaving transliterated comments underexplored. This research addresses the gap by evaluating the impact of auto-transliterated and manually transliterated datasets, merged with an existing Amharic hate speech dataset, on the performance of machine learning and deep learning classifiers. The study employed a total of 3,000 datasets which is split into ratio of 80:20 for training and testing. The dataset consists of auto-transliterated, manually transliterated, formal Amharic script, and their combinations. The classifiers including Support Vector Machine, single and multichannel Convolutional Neural Networks were assessed. Experimental results show that the multichannel CNN outperformed single-channel CNN models on the existing Amharic dataset, achieving an F1-score of 0.810 compared to 0.783 and 0.769 for single channel and multichannel CNN, respectively. However, combining transliterated datasets with the existing dataset did not improve classifier performance, likely due to the inconsistencies in scrip transliteration and dataset domain dependencies. This study concludes that transliterated datasets should be treated separately for hate speech detection, and combining datasets from different domains and transliteration techniques negatively impacts classifier performance.

A Multi-modal Fusion Technique to Combine Manual and Non-Manual Cues for Amharic Sign Language Recognition: A Systematic Literature Review

Isayas Feyera, Solomon Teferra, Asrat Mulatu Beyen — Mon, 30 Dec 2024 00:00:00 +0000

Amharic Sign Language (ASL) is a vital form of communication for the hearing-impaired community in Ethiopia. Recognizing and understanding ASL is crucial for facilitating communication and accessibility for Amharic-speaking hearing-impaired individuals. ASL relies on manual gestures, including hand shapes and movements, and non-manual cues such as facial expressions and body postures. The objective of this review is to investigate the methodologies employed to combine manual and non-manual cues in ASL recognition systems. In our review, we have considered various inclusion and exclusion criteria to select relevant research papers. After primary data selection, 46 papers which focus on sign language are included in our analysis. We also employed a data extraction form to collect and gather information from these selected papers systematically. Based on the review, combining manual and non-manual cues enhances the accuracy and robustness of sign language recognition systems. These techniques leverage computer vision and machine learning approaches to interpret manual gestures, while also capturing the nuanced information conveyed through facial expressions and body language. Improving hand gesture recognition involves the finding of key points or poses. Despite the advancements in ASL recognition, this review underscores a significant challenge—the lack of available resources, reputable publications, annotated data, and annotating tools specific to Amharic Sign Language are scarce. This shortage hampers the development and evaluation of ASL recognition systems, hindering progress in this field.