Saudi Cultural Missions Theses & Dissertations
Permanent URI for this communityhttps://drepo.sdl.edu.sa/handle/20.500.14154/10
Browse
69 results
Search Results
Item Restricted Weakly Supervised Learning for Medical Image Segmentation(Saudi Digital Library, 2026) Alshewaier, Hateef; Sun, XianfangMedical image segmentation plays a critical role in clinical diagnosis and treatment planning; however, the performance of deep learning models typically depends on large quantities of pixel-wise annotated data, which are costly and time-consuming to obtain. This thesis investigates annotation-efficient approaches for medical image segmentation by developing weakly supervised learning frameworks that reduce annotation requirements while maintaining high segmentation accuracy. The first contribution proposes an enhanced scribble-based segmentation framework that extends a modified U-Net architecture with increased feature-learning capacity, mix-based data augmentation, and global and local consistency regularisation. These components improve learning from sparse scribble annotations and promote anatomically coherent segmentation masks. Experimental results demonstrate improved segmentation performance compared with existing scribble-supervised methods. The second contribution presents an ensemble learning framework for weakly supervised medical image segmentation based on single bounding-box annotations. The proposed framework integrates U-Net, ResUNet, and DeepLabV3 using bagging and boosting strategies to exploit complementary architectural characteristics. Despite relying solely on coarse single-bounding-box supervision, the ensemble models improve robustness, reduce model-specific bias, and yield more stable segmentation predictions. Experimental results demonstrate that the ensemble methods consistently outperform the individual models, with boosting achieving the highest segmentation performance. The third contribution introduces a novel dual bounding-box supervision strategy that utilises an inner bounding box to represent high-confidence object regions and an outer bounding box to provide contextual anatomical information. A region-aware weighted loss function encourages accurate object localisation while reducing ambiguity in boundary regions. The proposed framework significantly outperforms conventional single-box weakly supervised approaches and narrows the gap between weak and full supervision. Overall, the findings demonstrate that high-quality medical image segmentation can be achieved using sparse scribble annotations, ensemble learning techniques, and coarse bounding-box supervision without requiring exhaustive pixel-wise labels. The proposed methods reduce annotation effort while maintaining competitive segmentation accuracy, providing a practical and scalable pathway for deploying deep learning-based segmentation systems in clinical environments.7 0Item Restricted PRIVACY-PRESERVING INTRUSION DETECTION FOR THE INTERNET OF MEDICAL THINGS USING ENSEMBLE AND FEDERATED LEARNING(Saudi Digital Library, 2026) Alsolami, Theyab; Ilyas, MohammadThe rapid proliferation of the Internet of Medical Things (IoMT) has transformed healthcare by enabling continuous monitoring, intelligent diagnostics, and data-driven clinical decision-making. However, this increased connectivity has significantly expanded the attack surface of healthcare systems, exposing sensitive patient data and critical medical devices to cyber threats such as intrusion and data exfiltration attacks. Ensuring both strong security and strict privacy preservation in IoMT environments remains a fundamental and unresolved challenge. This dissertation investigates the design and evaluation of robust and privacypreserving intrusion detection systems (IDS) for IoMT networks using advanced machine learning techniques. The research first examines the effectiveness of ensemble learning–based IDS models in centralized settings, evaluating Stacking, Bagging, and Boosting approaches with Random Forest and Support Vector Machine base learners on the WUSTL-EHMS-2020 dataset. Experimental results demonstrate that ensemble learning significantly enhances detection performance, with the Stacking model achieving an accuracy of 98.88%, followed by Bagging at 97.83%, while Boosting exhibits comparatively lower performance. Building on these findings, the dissertation extends intrusion detection to decentralized and privacy-sensitive IoMT environments through a federated learning (FL) framework integrated with Differential Privacy (DP) and secure aggregation mechanisms. Multiple experimental configurations are systematically analyzed, including raw imbalanced data, centralized SMOTE, and clientside (per-client) SMOTE under varying privacy budgets (ϵ = 3.0, 10.0, and non-private baselines). Results show that privacy-preserving federated models frequently match or exceed non-private baselines. In particular, raw imbalanced and per-client SMOTE configurations achieve high detection accuracy (approximately 94.6%) even under strict privacy constraints (ϵ = 3.0), demonstrating effective learning with minimal utility loss. Furthermore, client-side data balancing consistently outperforms centralized balancing, providing improved training stability while maintaining full data decentralization and patient confidentiality. Overall, this dissertation presents a comprehensive, scalable, and privacy compliant intrusion detection framework for IoMT systems. By integrating ensemble learning, federated learning, class imbalance mitigation, and differential privacy, the proposed approach successfully balances detection accuracy, privacy preservation, and computational efficiency. The findings provide both theoretical insights and practical guidelines for deploying secure and regulation-compliant IDS solutions in real-world healthcare IoMT environments.6 0Item Restricted Development and Evaluation of Machine Learning–Driven Workflow Models to Minimize Waste in Intravenous Compounding.(Saudi Digital Library, 2026) Ghawaa, Yazeed Mohammed; Lin, Alex CIntroduction: Intravenous (IV) medication waste represents a significant financial burden for healthcare systems. Leveraging advanced technological and analytical tools may improve the prediction and reduction of IV medication wastage. Artificial intelligence (AI) and machine learning (ML) offer strong capabilities to enhance pharmaceutical care, particularly in minimizing medication waste. Objectives: Aim 1: Assess the extent of intravenous (IV) dose waste within a sample hospital by quantifying both the volume of discarded doses and the associated drug acquisition costs. Further analysis will classify the wasted doses by preparation type: commercially pre-mixed, batch-prepared, and extemporaneously compounded. Aim 2: Develop multiple machine learning-driven workflow system (MLWS) models by exploring diverse combinations of predictive variables through advanced machine learning techniques, with the objective of minimizing waste from extemporaneously compounded IV doses. Aim 3: Evaluate the effectiveness of the developed machine learning-driven workflow system (MLWS) models in reducing waste from extemporaneously compounded IV doses and the estimated associated drug acquisition costs. This ML driven study has strong potential to reduce IV medication waste, reduce hospital workload and financial burden, and advance UN Sustainable Development Goals related to minimizing waste and promoting sustainable healthcare. Methods: A retrospective analysis and Machine Learning (ML) were conducted using one year of data collected from the inpatient pharmacy at The Christ Hospital between May 1, 2024, and April 30, 2025. The IV medication data was used for three primary purposes. First, we quantified prepared IV doses, the amount of IV dose waste, and the associated drug acquisition costs. These results were summarized using frequencies and percentages. Second, multiple Machine Learning–driven workflow system (MLWS) models were developed to predict and reduce IV dose waste. The dataset was categorized into two classes— cancelled and noncancelled doses—and then split into training (80%) and testing (20%) sets. A range of hyperparameters was evaluated to identify the best-performing model among three Machine Learning algorithms: Random Forest (RF), Extreme Gradient Boosting (XGBoost) and Logistic Regression (LR), then Soft Voting ensemble (SVE) was utilized to further enhance the performance. Model performance was compared using area under the ROC curve (ROC-AUC), accuracy, precision, recall, and F1-score. Finally, the optimized models were integrated using a Soft Voting ensemble to predict IV doses likely to be wasted. This approach identifies IV doses that would be prepared but not ultimately administered, allowing for timely intervention to halt preparation. Results: The results for Aim 1 indicated the monthly volume of IV doses prepared over a 12-month period from May 2024 to April 2025. On average, 15,666 doses were prepared each month, with a standard deviation range of 1,373. When The mean of number of wasted IV doses in using previously prepared (unused IV doses) was 1,754, marking 11.2% of all doses prepared. In comparison to not using previously prepared (unused IV doses), the mean number of wasted IV doses was 1,800, highlighting 11.5% of all doses prepared. The results of Aim 3 in Table 5 summarize the monthly predictions generated by the optimized machine learning (ML) models developed in Aim 2 to estimate the number of IV doses that might be prepared but ultimately not administered. These results include: (1) the total number of IV doses prepared, (2) the number of doses the models predicted would be cancelled, and (3) the number of cancelled doses the models failed to predict. On average, the ML models projected that approximately 1,200 IV doses per month would be prepared but not administered (SD: 136). In contrast, unpredicted cancellations missed by the models—averaged 598 per month (SD: 113). The number of doses predicted to be cancelled reflects the models’ ability to prevent IV waste, whereas the unpredicted cancellations represent the true wasted doses—those the models failed to anticipate. Overall, these findings show that a substantial proportion of anticipated administrations did not occur, highlighting important limitations in the models’ predictive accuracy and their effectiveness in reducing IV waste. A Wilcoxon signed rank test was performed to assess the impact of the ML models on the monthly total number of IV doses prepared and the number of doses prepared but not administered, comparing real world outcomes with model generated predictions. The analysis revealed statistically significant differences across all evaluated metrics. Specifically, significant differences were observed in the total IV doses prepared (p < 0.001) and the total wasted IV doses (p < 0.001). Additionally, the percentage of wasted IV doses showed a statistically significant difference (p = 0.0024). Conclusion: The results suggest that machine learning models can serve as an effective tool for predicting IV medication waste in the pharmacy setting.14 0Item Restricted Enhancing Speech Clarity through Audio-Visual Source Separation: A Multimodal Learning Approach Using Lip Movements and Facial Cues(Saudi Digital Library, 2025) Alqahtani, Fai; Wang, WenwuThis dissertation investigates the problem of audio-visual speech separation (AVSS) under structured and speech-like background interference, with a particular focus on culturally relevant noise such as the Adhan. Existing approaches often rely on lips-only or frontal-view visual cues, which limits robustness in real-world conditions involving head movement, occlusion, or expressive dynamics. To address these challenges, a deep learning framework was developed that integrates both acoustic and visual modalities. The audio branch employs HTS-AT to process mel-spectrogram features, while the visual branch leverages TimeSformer to encode multi-view, multi-region facial inputs, including lips, cheeks, and eyebrows, from five synchronised camera angles. Cross-modal attention is used for fusion, allowing the model to align visual motion with speech acoustics. Experimental results demonstrate that the proposed model outperforms a lips-only baseline across all objective and perceptual metrics. Specifically, the system achieved SDR = 11.20 dB, SI-SNR =12.00 dB, STOI = 0.91, and PESQ = 4.05, compared with baseline scores of SDR = 8.45 dB, SI-SNR = 9.68 dB, STOI = 0.88, and PESQ = 4.12. Grad-CAM visualisation further revealed that the proposed model attends to a broader range of articulatory and expressive cues, whereas the baseline concentrated narrowly on the lips. These findings highlight the effectiveness of multi-view, multi-region integration and attention based fusion in improving intelligibility and perceptual quality under structured noise. The work demonstrates the potential of such models for practical deployment in environments where traditional audio-only or lips-only systems fail to ensure speech clarity.7 0Item Restricted Efficient Intrusion Detection for IoMT: Integrating Machine Learning, Feature Selection, and Fuzzy Logic(Saudi Digital Library, 2026) Balhareth, Ghaida; Ilyas, MohammadThe internet of medical things (IoMT) has transformed healthcare by enabling real-time patient monitoring, remote diagnoses, and effective data exchange among connected medical devices and clinical systems. The increasing reliance on interconnected medical equipment has also intensified cybersecurity risks, as resource-constrained devices and wireless communication channels are vulnerable to attacks such as man-in-the-middle, spoofing, data injection, and ransomware. Intrusion Detection Systems (IDSs) play a critical role in mitigating these threats; however, traditional IDS approaches often struggle with high-dimensional IoMT data, class imbalance, and uncertainty in traffic patterns, which can increase false alarms and reduce reliability in safety-critical environments. This dissertation investigates efficient and deployable IDS designs for IoMT networks by integrating machine learning, feature selection, and fuzzy logic to improve detection reliability while reducing model complexity. First, the dissertation provide an extensive examination of IDS approaches proposed for IoMT, classifying them into machine learning, deep learning, fuzzy logic , and hybrid categories, while analyzing IoMT architectures and security vulnerabilities across layers. Next, it develops an efficient IDS model based on machine learning classifiers combined with feature selection techniques to enhanced detection accuracy and reduce computational cost in edge and gateway settings. Building on this direction, the dissertation proposed a multi-level feature selection pipeline that combines complementary ranking methods and consensus selection to identify consistently informative features, followed by a fuzzy inference system that supports uncertainty-aware intrusion classification using interpretable rule-based reasoning. The suggested IDS systems exhibit robust detection capabilities with reduced false-alarm rates, utilizing small feature sets appropriate for gateway and edge deployment throughout benchmark tests. The dissertation presents a cohesive security system that prioritizes efficiency, interpretability, and practical implementation for the protection of IoMT communications and the safeguarding of sensitive healthcare information. Future works will expand these IDS designs to include other IoMT datasets and real network traffic, while further investigating robustness in the context of concept drift and increasing adversarial strategies, all while maintaining low complexity and transparency.13 0Item Restricted Sex Differences in ICU Mortality and Prediction of Prolonged ICU Stay: A Study Using the MIMIC-III Critical Care Database(Saudi Digital Library, 2026) Asiri, Mohammed; Perez Concha, OscarBackground: Sex differences in critical illness outcomes remain contested after illness severity adjustment. Separately, early prediction of prolonged ICU stay has direct clinical utility for resource planning and discharge decision-making. Objectives: This study aimed to determine whether biological sex is independently associated with in-hospital mortality after adjusting for SOFA-based illness severity, comorbidity, and ICU case-mix, and to develop predictive models for prolonged ICU stay of five days or more using first-24-hour clinical features. Methods: The study used the MIMIC-III critical care database. Sequential multivariable logistic regression with multiple imputation by chained equations was applied to 31,000 adult first ICU admissions. For prediction of prolonged ICU stay, logistic regression, LASSO, Random Forest, and XGBoost models were evaluated on 26,729 patients using a stratified 70/30 train-test split. Results: Female sex was not independently associated with in-hospital mortality in the fully adjusted model (OR 1.11, 95% CI 0.99–1.25, p = 0.064). For prolonged ICU stay prediction, Random Forest achieved the highest AUROC (0.829) and the lowest Brier score (0.126), while XGBoost achieved the highest AUPRC (0.591). Mean Glasgow Coma Scale was the dominant predictor across models. Conclusion: Biological sex was not independently associated with in-hospital ICU mortality after adjustment for illness severity, comorbidity, and ICU case-mix. First-24-hour neurological status and oxygenation parameters were the strongest early predictors of prolonged ICU stay.8 0Item Restricted INVESTIGATING THE MECHANISMS AND DYNAMICS OF POST-TRAUMATIC BRAIN INJURY PATHOLOGIES: FROM ASTROCYTE REACTIVITY TO SEIZURE-LIKE ACTIVITY(Saudi Digital Library, 2026) Mufti, Shatha; Shi, RiyiTraumatic brain injury (TBI) is a leading cause of death and long-term disability worldwide, frequently resulting in complex secondary injuries and chronic neurological conditions such as post-traumatic epilepsy (PTE). Despite the high prevalence of TBI, identifying effective treatments for the subsequent sequalae has been challenging due to the unclear mechanisms of how TBI leads to conditions like PTE. My work addresses this gap by utilizing an in vitro TBI-on-a-chip model to simulate concussive impacts on primary cortical networks and conduct targeted mechanistic investigations into the biochemical and electrophysiological mechanisms driving TBI-induced pathologies like PTE. In the first study, a combination of electrophysiological microelectrode array (MEA) recordings and immunocytochemistry are used to investigate the role of the toxic reactive aldehyde acrolein in promoting injury-induced pathologies. The results identify acrolein as a primary driver of post-TBI astrocyte reactivity and neuronal network hyperexcitability and show that sequestering acrolein pharmacologically with hydralazine effectively mitigates these pathological changes. Furthermore, a novel machine learning (ML) framework using convolutional neural networks and Grad-CAM was employed to analyze astrocyte structural remodeling and pinpoint specific morphological features associated with injury and treatment conditions. In the second study, MEA recordings capturing changes in firing and bursting dynamics after impact are analyzed to examine the mechanisms underlying the transition of injured networks into seizure-like activity (SLA) using custom algorithms to quantify cross-correlogram shapes and identify bursts in network firing. The findings show that injury not only increases the synchronization of neuronal firing, which is a hallmark of SLA, but also reorganizes firing hierarchies and alters leader-follower relationships among neurons. Examining multiple dimensions of network activity simultaneously provides a more comprehensive understanding of how TBI alters neuronal communication and network dynamics and promotes SLA. In the third study, further investigations into altered neuronal interactions that lead to SLA are performed by applying a multilayered ML pipeline, which includes LSTM autoencoders, UMAP clustering, and deep Granger causality to MEA recordings of neuronal networks treated with bicuculline, a standardized experimental model of SLA. The results reveal that distinct neuronal subpopulations with unique activity profiles emerge during SLA, even in states of global network synchronization. Collectively, the findings in this thesis demonstrate that post-traumatic pathologies, especially PTE, are driven by identifiable molecular targets and structured network reorganization, suggesting a variety of options for therapeutic targeting. Furthermore, this work highlights the TBI-on-a-chip system as a versatile platform for investigating the pathophysiology of both TBI and PTE, enabling mechanistic studies into injury progression, the identification of novel therapeutic targets and biomarkers, the screening of new drugs, and the development of promising diagnostic tools.14 0Item Restricted Feasibility of a Multi-Dimensional AI System for Gifted Student Identification in Saudi Education(Saudi Digital Library, 2026) Alahdal, Ashwaq; Abuelmaatti, AishaIdentifying gifted students poses a considerable challenge in educational research, especially in contexts that require extensive data collection. This study introduces a data-driven method for identifying gifted students, employing machine learning models that integrate both simulated and actual datasets. A simulated dataset was developed to reflect the traits of gifted Saudi students, based on genuine academic patterns and educational research, covering academic, creative, and cultural dimensions. By utilizing a randomized classifier, gifted students were classified based on indicators from various disciplines. The model attained 96% predictive accuracy on the dataset examined and 98% on the global Cagle dataset. The findings revealed that academic and creative variables were the most significant predictors of giftedness. This research provides a practical framework for educational systems to identify gifted students in contexts where detailed data are limited, thereby enhancing equity and effectiveness in programs for gifted students. Keywords: Artificial intelligence in education, gifted students, machine learning, Random Forest classifier, classification, educational data analysis.10 0Item Restricted Real-Time IoT Data Cleaning and Anomaly Detection Using Context-Aware Frameworks and Large Language Models(Saudi Digital Library, 2026) Alotaibi, Obaid Haylan B; Eric, Pardede; Sarath, TomyThe Internet of Things (IoT) has delivered significant benefits to various domains such as healthcare, business, and industry by generating vast amounts of data in real time. However, IoT-generated data often suffers from low quality due to issues that can significantly affect data analysis results and lead to inaccurate decision making. Enhancing the quality of real-time data streams has become a challenging task because the characteristics of IoT data make anomaly detection particularly chal lenging, which is crucial for informed decisions. Traditional IoT data cleaning tech niques primarily rely on batch processing methods, which introduce latency and fail to effectively handle real-time streaming IoT data. Many studies have proposed different techniques to overcome these challenges, such as cleaning data in real time; however, no comprehensive data cleaning framework has been proposed. This thesis proposes a comprehensive streaming data cleaning framework aimed at improving the quality of real-time data streams. Central to this framework is a real-time anomaly detection model for structured IoT data streams. The model de tects multiple types of anomalies and classifies them as either significant events or errors. Additionally, the proposed method incorporates context-awareness to further enhance detection reliability. Building upon this detection capability, the framework includes an automated repair system that addresses detected anom alies via multiple repair techniques: delete, replace, or keep, using statistical mea surements and machine learning based on anomaly classification. To enhance user decision-making, the framework integrates large language mod els for data stream cleaning, providing context-aware recommendations and sen sitivity assessments. Large language models operate locally, assisting users to dynamically refine contexts and sensitivity levels based on real-time interaction streams across diverse applications. Overall, this thesis highlights the proposed framework’s effectiveness in guiding users by providing a clear picture, thereby enhancing decision-making accuracy in real-time environments and enabling confident, real-time responses to genuine anomalies.31 0Item Restricted A DATA ANALYTICS FRAMEWORK TO SUPPORT DECISION MAKING IN RAILWAY INFRASTRUCTURE ASSET MANAGEMENT(Saudi Digital Library, 2026) Alotaibi, Abdulaziz; Cardenas, IsidroThe process of management of the assets of the railway infrastructure is becoming increasingly dependent on the big amounts of the condition-monitoring information produced by the recent inspection technologies. Although this type of data can give a detailed picture of the track condition, it also brings issues of interpretation, prioritisation and decision making. The current asset-management methods usually are based on evaluating thresholds and disjointed analysis tools, which restrict their strengths in promoting proactive and data-driven maintenance practices. The study creates and assesses a combined visual analytics system in order to aid decision making in the management of railway infrastructure assets. The framework integrates data pre-processing, analytical intelligence, machine-learning, and interactive visual analytics to convert raw track geometry data into actionable decision-support products. The research design was a mixed-methods research design comprising of two large-scale case studies, one of them on the basis of the UK and Saudi Arabian railway networks, and the other one on the basis of expert validation. The data of track geometry measured by Network Measurement Trains and Track Geometry Inspection Vehicles was analysed to prove the relevance of the framework to the different operational and environmental conditions. The case study of the UK is a fully developed, regulation-based data environment whereas the Saudi Arabian case study is a developing network that is functioning in the harsh desert conditions. Findings indicate that the suggested framework improves the interpretability of complex condition data using integrated 2D, 3D, and GIS-based visual analytics. The unsupervised and supervised methods were combined to form machine-learning techniques which enhanced the performance of fault detection and classification and led to quantifiable reductions in false positive alerts compared with the baseline threshold-based methods. A comparative analysis shows that the framework can be adjusted to differences in data maturity, regulatory environment, and operational issues. The study brings on board a transferable and validated visual analytics model that provides the balance between advanced data analytics and feasible decision support in the management of railway infrastructure assets.13 0