Bachelor of Science in Data Science and Analytics

Permanent URI for this collection

https://hdl.handle.net/20.500.12311/943

Browse

Recent Submissions

Now showing 1 - 4 of 4
  • No Thumbnail Available
    Item
    Ai-powered offline maize streak virus (MSV) detection system for smallholder farmers: design, implementation, and evaluation
    (Uganda Christian University, 2026-06-05) Liz Treasure Namatovu; Charles Kidega Omoya; Irwin Mwine
    Maize Streak Virus (MSV) remains a highly destructive agricultural pathogen in Sub-Saharan Africa, severely undermining crop yields and threatening smallholder food security. While laboratory diagnostics like PCR and LAMP offer high sensitivity, they remain financially and logistically inaccessible to rural farmers who lack timely extension support. Deep learning and mobile computer vision present a transformative alternative for automated plant disease detection. However, most existing digital solutions assume persistent internet connectivity, cloud infrastructure, or premium hardware, rendering them impractical under the rigid network and power constraints of rural African fields. To bridge this technology adoption gap, this project presents MaizeGuard, an inclusive, dual platform, entirely offline-first AI-powered diagnostic ecosystem engineered for localized MSV classification. The system integrates an optimized MobileNetV2 convolutional neural network architecture trained via transfer learning to classify maize leaves in real time into three categories: Healthy Maize Leaf, MSV Infected Leaf, and Not a Maize Leaf. For smartphone users, the MaizeGuard mobile tier utilizes a React Native and Expo framework integrated with ONNX Runtime for on-device inference and Async Storage for local caching. For users lacking smartphone access, a standalone hardware kit was built using a Raspberry Pi microcomputer, a CSI camera, a 3.5-inch resistive TFT touchscreen, and a custom Python framebuffer application configured via systemd to auto-launch at boot. An optional Node.js/PostgreSQL backend supports asynchronous data synchronization without interfering with core offline diagnoses. The system engineering process was grounded in a systematic review of 24 peer-reviewed studies, which validated lightweight CNN optimization pathways for edge environments and highlighted the scarcity of farmer-ready MSV diagnostic systems. System evaluation encompassed functional validation, multi-platform integration benchmarking, execution latency tracking, thermal profiles, and touchscreen calibration durability under simulated field setups. Experimental results demonstrated that the completed dual-platform ecosystem achieves high classification accuracy, low computational latency, and robust operational resilience under low-resource field conditions. This work contributes a practical, scalable advancement in digital precision agriculture, demonstrating how edge AI design can deliver inclusive diagnostic tools directly to smallholder farmers in Uganda.
  • No Thumbnail Available
    Item
    Predictive and prescriptive analytics to strengthen infectious disease surveillance and response
    (Uganda Christian University, 2026-06-01) Joy Florence Awor; Judith Anita Namaganda; Calvin Diego Rwomothio
    Uganda continues to experience recurring outbreaks of epidemic-prone infectious diseases, driven by persistent gaps in water, sanitation, and hygiene infrastructure, rapid urbanization, population displacement, and a limited capacity for timely detection and coordinated response. Current surveillance systems remain largely reactive and descriptive, focusing on reporting confirmed cases rather than predicting and preventing outbreaks. This study develops a proactive Health Intelligence Platform designed to strengthen Uganda’s epidemic preparedness by integrating predictive and prescriptive analytics with response visualization. Using a curated cholera surveillance dataset as a proof-of-concept, the platform applies machine learning techniques to forecast potential outbreak hotspots and temporal trends. These predictive outputs are operationalized through an interactive dashboard that provides timely alerts, spatial mapping, risk identification, and decisionsupport indicators for public health officials. The system demonstrates how predictive models combined with environmental factors can enhance situational awareness and support faster, data-driven interventions at both district and national levels. The ensemble models trained on 8,702 cholera surveillance records achieved strong short-term forecasting performance: the Random Forest Regressor attained an R2 of 0.7861 for suspected case prediction and 0.6204 for confirmed case prediction, outperforming XGBoost across all evaluation metrics (MAE and MSE). These accuracy levels represent a meaningful advance over Uganda’s current surveillance infrastructure platforms such as DHIS2, IDSR, and eIDSR which remain descriptive and retrospective, providing no automated outbreak forecasting, no anomaly detection, and no environmental risk integration. The XGBoost-powered 14-day early warning module further generated forward looking case projections consistent with observed transmission trends, correctly identifying the February 2026 peak of 700 confirmed cases as the most severe anomaly in the 15-year record, with a 200% deviation above the rolling average a signal that the existing national system did not surface proactively. Overall, this project showcases a scalable, Uganda-focused Health Intelligence Platform capable of transitioning the country’s surveillance ecosystem from reactive reporting to anticipatory, data-driven epidemic preparedness.
  • No Thumbnail Available
    Item
    A hybrid machine learning-based type 2 diabetes mellitus risk prediction tool for rural Ugandan settings
    (Uganda Christian University, 2026-06-11) Lorraine Paula Arinaitwe; Noela Rugogamu; Aloysious Ssendi
    Type 2 Diabetes Mellitus (T2DM) is becoming a major public health problem in Uganda, especially in rural communities where access to diagnostic equipment and specialised health-care services is limited. This project presents PreDia, a hybrid machine learning-based risk prediction tool developed to support early screening of T2DM in low-resource rural Ugandan settings. The system uses low-cost and clinically accessible features and combines an XG-Boost predictive model with rule-based clinical guidelines from the World Health Organization (WHO) to improve the reliability of risk assessment. To improve transparency and user trust,SHAP (SHapley Additive exPlanations) was integrated to provide understandable explanations for each prediction. The system also includes an offline-first bilingual interface in English and Luganda to support accessibility among Village Health Teams (VHTs). The predictive model was trained and tested using an augmented version of the Pima Indians Diabetes Dataset adapted to reflect selected physiological trends found in Sub-Saharan African populations. Experimental results showed strong predictive performance, achieving an AUC score of 0.9672, an accuracy of 88.96%, a recall of 87.04%, and a precision score of 82.46%. The findings of this study demonstrate that explainable and offline-capable artificial intelligence tools can support community-level diabetes risk screening in resource-constrained healthcare environments.
  • Thumbnail Image
    Item
    Educonnect: Ai-enabled peer-to-peer learning framework for university students
    (Uganda Christian University, 2026-06-01) Ronald Tusiime; Hellen Noeline Nambooze; Lynn Amoit
    EduConnect is a web-based, AI-enabled peer learning platform designed to revolutionize how university students discover study partners, access high quality academic resources, and sustain progress. The project addresses three critical local challenges: fragmented peer-learning coordination, limited personalisation in study support and weak integration between collaboration tools and progress tracking. The implemented client-server architecture utilises a React frontend, Node.js/Express API and SQLite database persistence (sql.js) with optional Python services for advanced NLP intent classification and sequence-to-sequence quiz generation. EduConnect integrates intelligent peer recommendations, multi-discipline resource browsing, adaptive quiz workflows, study groups with direct and group messaging and a chatbot (EduBot) for intent based assistance. A key technical contribution is a trained multinomial logistic regression model for predicting next study topics based on user engagement statistics, trained in Python and exported to a JSON weight matrix for native JavaScript inference. Pilot survey evidence from 50 students indicates strong adoption potential and problem relevance. The average student difficulty in finding study partners was rated at 3.80/5, and overall interest in the platform was 4.36/5, with 78% of respondents willing to join beta testing. Combined with a fully functional, deployed MVP, these findings support the technical viability and educational significance of the EduConnect framework. Keywords: Peer learning, educational technology, recommendation systems, logistic regression, NLP intent classification, student collaboration.