http://dlib.iit.ac.lk/xmlui/handle/123456789/1 Tue, 05 May 2026 05:53:01 GMT 2026-05-05T05:53:01Z http://dlib.iit.ac.lk/xmlui/handle/123456789/3268 A transformer neural network for EHR data imputation Hewa Dewage, Emith Dinsara Electronic Health Records (EHR) are a very valuable data mine for clinical data analytics. The prevalence of missing data in EHR poses a significant challenge to healthcare analytics and patient care. This affects the reliability of data-driven decisions. It is evident that healthcare professionals can obtain an immense advantage and make educated and efficient decisions in clinical context. To deal with missing data in EHR, apart from other complexities such as sparsity etc. prevalent in EHR, imputation methodologies are introduced. Most traditional imputation methods are not suitable for EHR data imputation due to the suboptimal accuracy and generalizability. This study intends to address the need for a solid and robust EHR imputation methodology, proposing a transformer-based neural network model to effectively handle various missing data mechanisms, such as Missing Completely At Random (MCAR), Missing At Random (MAR), Missing Not At Random (MNAR). The research adapts the self-attention mechanism in transformers, which enables the model to dynamically weigh features based on relevance, enhancing adaptability to diverse data patterns. A structured approach was applied which included data preprocessing, model training and iterative validation with custom training loops. The implementation of the transformer neural network was able to outperform the existing EHR data imputation methodologies by a close yet considerable margin and a demonstration platform was built to upload, impute and download datasets. Wed, 01 Jan 2025 00:00:00 GMT http://dlib.iit.ac.lk/xmlui/handle/123456789/3268 2025-01-01T00:00:00Z http://dlib.iit.ac.lk/xmlui/handle/123456789/3267 Leveraging Change Point Detection for Enhanced Credit Card Fraud Detection Kankanam Arachchi Haggallage Don, Vihangi Credit card fraud represents a trend within the financial sector that creates tremendous monetary loss and lost consumer confidence. Traditional fraud detection methods, founded on static models and predefined rules, have difficulty keeping pace with the ever-changing fraudsters' strategies and therefore generate excessive numbers of false positives and let through fraudulent transactions. This project mitigates these limitations by applying Change Point Detection (CPD) techniques to identify sudden changes in behavioral patterns of transactions that define fraudulent activities. By integrating CPD with machine learning algorithms such as Isolation Forest, Local Outlier Factor (LOF), and One-Class SVM, the system enhances the accuracy and timeliness of detecting fraud and lowering false alarms. The work draws on a mixed methodology that combines CPD with anomaly detection models to dynamically monitor streams of transactions. CPD identifies abrupt changes in transactional patterns, for instance, unusual spikes in the value or volume of transactions, whereas the machine learning classifiers mark such anomalies as possibly fraudulent. The system is evaluated using a publicly accessible 284,807 credit card transaction dataset and has an accuracy of 99.02% when employing One Class SVM and operates well against false positives compared to the with CPD applied and without CPD applied. Key performance indicators such as precision, recall, and F1-score are quantified to validate the effectiveness of the proposed solution. This research unveils the unrevealed capabilities of CPD as a credit card fraud detection method through its findings of an effective solution that suits financial institutions. The system operates through a modular design structure allowing simple integration of standard fraud detection software frameworks implemented Python Frameworks. Future development should center on time-sensitive processing as well as better feature engineering with cloud deployment to enhance both performance and scalability dimensions. This research completes an essential void in fraud identification through its data-oriented adaptive system that updates its detection of new fraud methods in real-time. Wed, 01 Jan 2025 00:00:00 GMT http://dlib.iit.ac.lk/xmlui/handle/123456789/3267 2025-01-01T00:00:00Z http://dlib.iit.ac.lk/xmlui/handle/123456789/3266 Food-Drug Interaction and Allergy Prediction System Using Machine Learning Prabuddha, Sahan Problem: Individuals with allergies often struggle to identify potential allergic reactions when consuming packaged food or taking multiple medications, especially in cases where ingredient lists are complex or not clearly understood. Existing systems typically focus on either drug-drug interactions or single-allergen predictions, lacking a comprehensive, personalized approach. This research addresses this gap by developing a system that predicts potential allergic reactions caused by interactions between food and medicine ingredients, considering individual-specific health data such as gender and blood type. Methodology: The system utilizes Optical Character Recognition (OCR) to extract food label text and combines it with user-entered medication details. Two machine learning models are developed—one for predicting allergy severity using Gradient Boosting, and another for predicting allergy descriptions using XGBoost and Random Forest. The dataset is preprocessed using normalization and SMOTE for balancing. Personalized features such as gender and blood group are integrated to improve model accuracy. The system is built as a web-based platform using React.js for the front end and a Python backend for model execution and prediction. Initial Result: The allergy severity prediction model achieved an accuracy of 96.27%, while the allergy description prediction model achieved 84.01% (XGBoost) and 77.52% (Random Forest). OCR accuracy depends on label clarity and formatting but generally performs well under clean image conditions. Although results show strong initial performance, improvements are planned for label variability handling, ingredient name normalization, and large-scale testing with real world data. Wed, 01 Jan 2025 00:00:00 GMT http://dlib.iit.ac.lk/xmlui/handle/123456789/3266 2025-01-01T00:00:00Z http://dlib.iit.ac.lk/xmlui/handle/123456789/3265 PerfectFit-An Intelligent Fashion Recommendations System Combining Voice input, Skin Tone and Body Shape Analysis, and Explainable AI for Transparent and Personalized Style Guidance. Mendis, Gishya The PerfectFit solution addresses the challenge of accurately classifying women's body shapes from the uploaded images to provide personalized fashion recommendations. Traditional methods relying on manual measurement or surface classifiers are likely to fail on the richness of real images, with the variations of pose, lighting, and background. Existing fashion recommendation systems mostly ignore useful aspects such as skin tone and occasion that limit their personalization and user satisfaction. The basic problem therefore is to design an automatic, strong, and interpretable system that is capable of identifying body shapes well and merging multiple attributes of users to offer personalized advice on dressing. To solve this problem, PerfectFit employs a deep learning approach founded on the MobileNetV2 convolutional neural network model. MobileNetV2 was chosen based on its trade-off between accuracy and computational expense that qualifies it for use in real-time scenarios. The model relies on transfer learning using the initialization of pretrained ImageNet weights followed by fine tuning on a judiciously selected dataset of body shape images appropriately labeled. The architecture includes bottleneck residual blocks and depthwise separable convolutions to reduce model size without sacrificing performance. Additional layers such as Global Average Pooling and a Dense softmax output layer were added to classify five distinct body shape categories. Complementary modules for skin tone detection and explainable AI, powered by OpenAI’s GPT 3.5, were integrated to enhance recommendation personalization and transparency. Wed, 01 Jan 2025 00:00:00 GMT http://dlib.iit.ac.lk/xmlui/handle/123456789/3265 2025-01-01T00:00:00Z