BRAC University Institutional Repository
Preserving Knowledge, Advancing Research, Sharing Scholarship
A digital platform for collecting, preserving, and sharing BRAC University’s scholarly, academic, and institutional outputs.
Contribute to the Repository
Faculty members and students are invited to submit their research publications, theses, dissertations, and scholarly works to increase visibility, access, and long-term preservation.
Recent Submissions
Analyzing schizophrenic texts from social media through machine learning and natural language processing
(Institute of Electrical and Electronics Engineers Inc., 2024-01-01) Rodela, Raisa Rahman; Rahman, Mubashira; Efty, Farhan Tanvir; Rahman, Rafeed; Reza, Md Tanzim; Department of Computer Science and Engineering
Schizophrenia is a destructive personality disorder where people may develop harmful symptoms if not diagnosed promptly. The research focuses on identifying language patterns indicative of schizophrenic-prone texts in online communication and intends to contribute to the development of early intervention techniques in mental health using ML and NLP methods. The study has utilized an existing dataset to examine language patterns associated with schizophrenia in social media posts. Various ML, BERT and RNN models have been used to predict textual data suggestive of schizophrenia. The analysis shows excellent results, with the DistilBERT achieving the highest accuracy rates of 97%, GRU acquiring 91% accuracy, and the logistic regression achieving 93% accuracy respectively.
Securing federated learning: a defense mechanism against model poisoning threats
(Institute of Electrical and Electronics Engineers Inc., 2024-01-01) Anan, Fabiha; Kamal, Md. Sifat; Shahed Mamun, Kazi; Ahsan, Nizbath; Reza, Md. Tanzim; Iqbal Hossain, Muhammad; Department of Computer Science and Engineering
Distributed machine learning advancements have the potential to transform future networking systems and communications. An effective framework for machine learning has been made possible by the introduction of Federated Learning (FL) and due to its decentralized nature it has some poisoning issues. Model poisoning attacks are one of them that significantly affect FL's performance. Model poisoning mainly defines the replacement of a functional model with a poisoned model by injecting poison into models in the training period. The model's boundary typically alters in some way as a result of a poisoning attack, which leads to unpredictability in the model outputs. Federated learning provides a mechanism to unleash data to fuel new AI applications by training AI models without access anyone's confidential data. Currently, there are many algorithms that are being used for defending model poisoning in federated learning. Some of them are really efficient but most of them have lots of issues that don't make the federated learning system properly secured. So in this study, we have highlighted the main issues of these algorithms and provided a defense mechanism that is capable of defending model poisoning in federated learning.
Mobile phone use for pregnancy-related healthcare utilization and its association with optimum antenatal care and hospital delivery in Bangladesh
(Public Library of Science, 2023-04-01) Al Kibria, Gulam Muhammed; Hashan, Mohammad Rashidul; Hanif, Abu Abdullah Mohammod; Maniar, Vidhi; Shawon, Md Shajedur Rahman; BRAC James P Grant School of Public Health
Pregnancy-related healthcare utilization is inadequate in Bangladesh, where more than half of pregnant women do not receive optimum number of antenatal care (ANC) visits or do not deliver child in hospitals. Mobile phone use could improve such healthcare utilization; however, limited evidence exists in Bangladesh. We investigated the pattern, trends, and factors associated with mobile phone use for pregnancy-related healthcare and how this can impact at least 4 ANC visits and hospital delivery in the country. We analyzed cross-sectional data from Bangladesh Demographic and Health Survey (BDHS) 2014 (n = 4,465) and 2017–18 (n = 4,903). Only 28.5% and 26.6% women reported using mobile phones for pregnancy-related causes in 2014 and 2017–18, respectively. Majority of the time, women used mobile phones to seek information or to contact service providers. In both survey periods, women with a higher education level, more educated husbands, a higher household wealth index, and residence in certain administrative divisions had greater likelihoods of using mobile phones for pregnancy-related causes. In BDHS 2014, proportions of at least 4 ANC and hospital delivery were, respectively, 43.3% and 57.0% among users, and 26.4% and 31.2% among non-users. In adjusted analysis, the odds of utilizing at least 4 ANC were 1.6 (95% confidence interval (CI): 1.4–1.9) in BDHS 2014 and 1.4 (95% CI: 1.3–1.7) in BDHS 2017–18 among users. Similarly, in BDHS 2017–18, proportions of at least 4 ANC and hospital delivery were, respectively, 59.1% and 63.8% among users, and 42.8% and 45.1% among non-users. The adjusted odds of hospital delivery were also high, 2.0 (95% CI: 1.7–2.4) in BDHS 2014 and 1.5 (95% CI: 1.3–1.8) in BDHS 2017–18. Women with history of using mobile phones for pregnancy-related causes were more likely to utilize at least 4 ANC visits and deliver in health facilities, however, most women were not using mobile phones for that. Copyright: © 2023 Kibria et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Predicting preterm birth among south-asian women using machine learning
(Institute of Electrical and Electronics Engineers Inc., 2024-01-01) Alif, Meheruba Hasin; Main Bhuyan, Farah; Tajrin, Radhika; Department of Computer Science and Engineering
Preterm Birth remains one of the preventable yet large contributors to child and maternal mortality. Despite progress in reducing infant mortality, South Asia continues to report a significant number of preterm birth-related deaths annually. Moreover, those who survive have to bear the long-term impact of physical and neurological disabilities - especially low-income households who are unable to afford healthcare. This study aims to offer a prediction model that can predict preterm birth using social, physical and health records of South-Asian women. Traditional models such as Decision Trees, Random Forest, Support Vector Machines, Logistic Regression and neural network-based deep learning models such as Multilayer Perceptron are used to compare the models' AUC, F1 scores and accuracy points. Random Forest generated the highest accuracy and F1 score (89%), whereas Multilayer Perceptron generated the best AUC score (88%) and had an overall consistent performance. Further analysis revealed the main factors contributing to preterm birth were weight before pregnancy, age and BMI of the mother.
The effects of social determinants on children's health outcomes in Bangladesh slums through an intersectionality lens: An application of multilevel analysis of individual heterogeneity and discriminatory accuracy (MAIHDA)
(Public Library of Science, 2023-03-01) Barua, Proloy; Kibuchi, Eliud; Aktar, Bachera; Chowdhury, Sabrina Fatema; Mithu, Imran Hossain; Quayyum, Zahidul; de Siqueira Filha, Noemia Teixeira; Leyland, Alastair H.; Rashid, Sabina Faiz; Gray, Linsay; BRAC James P Grant School of Public Health
Empirical evidence suggests that the health outcomes of children living in slums are poorer than those living in non-slums and other urban areas. Improving health especially among children under five years old (U5y) living in slums, requires a better understanding of the social determinants of health (SDoH) that drive their health outcomes. Therefore, we aim to investigate how SDoH collectively affects health outcomes of U5y living in Bangladesh slums through an intersectionality lens. We used data from the most recent national Urban Health Survey (UHS) 2013 covering urban populations in Dhaka, Chittagong, Khulna, Rajshahi, Barisal, Sylhet, and Rangpur divisions. We applied multilevel analysis of individual heterogeneity and discriminatory accuracy (MAIHDA) to estimate the Discriminatory Accuracy (DA) of the intersectional effects estimates using Variance Partition Coefficient (VPC) and the Area Under the Receiver Operating Characteristic Curve (AUC-ROC). We also assessed the Proportional Change in Variance (PCV) to calculate intersectional effects. We considered three health outcomes: cough, fever, and acute respiratory infections (ARI) in U5y.We found a low DA for cough (VPC = 0.77%, AUC-ROC = 61.90%), fever (VPC = 0.87%, AUC-ROC = 61.89%) and ARI (VPC = 1.32%, AUC-ROC = 66.36%) of intersectional strata suggesting that SDoH considered do not collectively differentiate U5y with a health outcome from those with and without a health outcome. The PCV for cough (85.90%), fever (78.42%) and ARI (69.77%) indicates the existence of moderate intersectional effects. We also found that SDoH factors such as slum location, mother's employment, age of household head, and household's garbage disposal system are associated with U5y health outcomes. The variables used in this analysis have low ability to distinguish between those with and without health outcomes. However, the existence of moderateintersectional effect estimates indicates that U5y in some social groups have worse health outcomes compared to others. Therefore, policymakers need to consider different social groups when designing intervention policies aimed to improve U5y health outcomes in Bangladesh slums. © 2023 Barua et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.