Nahla Barakat

The University of Queensland's institutional repository, UQ eSpace, aims to create global visibility and accessibility of UQ’s scholarly research.

Card fraud has become a widespread issue as technology advances and credit cards become more popular. Over the last decade, criminals have found diverse ways to steal card information or create duplicates without the cardholder’s knowledge by deploying adaptive approaches against the system. Therefore, detecting suspicious behaviour of a card is critical for preventing fraudulent transactions. However, detecting fraud behaviour is challenging task, given that the transactions’ data are imbalanced, and the adaptive fraudsters’ approaches against the system. In this paper, a combination of data and model centric approaches are proposed for card fraud detection. Starting by data analysis and preprocessing, this was followed by feature engineering stage, then two main family of models were used. The first is an Auto Encoder for anomaly detection, and the second is time series prediction, utilizing Long Short-Term Memory (LSTM) and Convolution Neural Network. The variations of LSTM used include CNN-LSTM, LSTM-Attention mode, and CNN-LSTM with an attention layer. The latter achieved the State-of-the-Art F-score of 93% on the IBM data set. Best F-scores of 94% is also achieved by stacked LSTM on Sparkov dataset.

Social media recently became convenient platforms for different groups with common concerns to share their experiences, including Adverse Drug Reactions (ADRs). In this paper, we propose a two stage intelligent algorithm which we call “Simi_to_Prob”, that utilizes social media forums; for ranking ADRs, and evaluating the ADRs prevalence considering different age and gender groups as its first stage. In the second stage, ADRs are predicted utilizing a different data set from the Food and Drug Administration (FDA). In particular, Natural Language Processing (NLP) is used on social media to extract ranked lists of ADRs, which are then validated using novel intrinsic evaluation methods. In the second stage, feature engineering is used to extend the input feature space, then a two stage supervised machine learning method is used to predict future ADRs incidences. Our results show correct ranked list of ADRs for three antihypertensive drugs, where high Spearman’s rank correlation coefficients (rs) of of 0.7458, 0.6678 and 0.5929 were obtained between SIDER database for drug ADRs, and our obtained lists from social media. Furthermore, Relatedness between ADRs and age and gender groups achieved high area under the ROC curve (AUC) reaching 0.959. The second stage results showed high AUCs of 0.96 and 0.99 for the prediction of future ADRs probabilities. The proposed algorithm shows that mining social media can provide reliable source of information, and additional features that can be used to boost supervised machine learning methods’ performance in different domains including Pharmacovigilance research.

The aim of this study is to develop an intelligent diagnostic system utilizing machine learning for data cleansing, then build an intelligent model and obtain new cutoff values for APRI (aspartate aminotransferase-to-platelet ratio) and FIB-4 (fibrosis score) for the prediction and staging of fibrosis in children with chronic hepatitis C (CHC). Random forest (RF) was utilized in this study for data cleansing; then, prediction and staging of fibrosis, APRI and FIB-4 scores and their areas under the ROC curve (AUC) have been obtained on the cleaned dataset. A cohort of 166 Egyptian children with CHC was studied. RF, APRI, and FIB-4 achieved high AUCs; where APRI had AUCs of 0.78, 0.816, and 0.77; FIB-4 had AUCs of 0.74, 0.828, and 0.78; and RF had AUCs of 0.903, 0.894, and 0.822, for the prediction of any type of fibrosis, advanced fibrosis, and differentiating between mild and advanced fibrosis, respectively. Machine learning is a valuable addition to non-invasive methods of liver fibrosis prediction and staging in pediatrics. Furthermore, the obtained cutoff values for APRI and FIB-4 showed good performance and are consistent with some previously obtained cutoff values. There was some agreement between the predictions of RF, APRI and FIB-4 for the prediction and staging of fibrosis.

Recently, smart home technologies and Internet of Things (IoT) have made it easy to control, monitor and secure properties, specially; while being away. In this context, several security and alarm systems have been proposed over the last decade. However, most of these systems lack any intelligent components, which make the alarm system generalize beyond the strictly, user predefined actions; therefore minimize false and nuisance alarms. In this paper, we propose an intelligent alarm system, which utilizes machine learning techniques to intelligently detect threats based on the environment data collected by different sensors and Arduino microcontroller; and then notify the owner by email message on real time basis. As the results section shows, the proposed system achieved high performance, and low false positive rate (false alarms).

Over the last decade, "Gamification" has been widely used in many domains for better user engagement and for increasing motivation in different ways. One of the important domains where gamification has been utilized is health and wellbeing; where it has been used in an expanded set of objectives. This ranges from maintaining healthy life style, to helping patients increase their adherence to their treatment and medications. In general, the concept of "Gamification" refers to the use of game design elements in non-game context. In this paper, Gamification has been utilized to help in managing the stress of young cancer survivors in an innovative way, with two main objectives. The first is to help improve young cancer survivors' moral and decrease their worries. The second is to monitor their health conditions by letting them indirectly answer medical questions while playing the game. This is achieved by mapping the text of medical questions to suitable wording, then embedding them in the game story and flow. The answers will be then persistently saved in a file, and given to the medical specialists, who can get clear indication on the survivor health condition on regular basis.

Classification performance of different algorithms can often be improved by excluding irrelevant input features. This was the main motivation behind the significant number of studies proposing different families of feature selection techniques. The objective is to find a subset of features that can describe the input space, at least, as good as the original set of features. In this paper, we propose a hybrid method for feature ranking for support vector machines (SVMs); utilizing SVMs support vectors (SVs). The method first finds the subset of features that least contribute to interclass separation. These features are then re-ranked using correlation based feature selection algorithm, as a final step. Results on four benchmark medical data sets show that the proposed method, though simple, can be a promising feature reduction method for SVMs and other families of classifiers as well.