رانية روني عزيز
Heart disease (HD) is the leading cause of death worldwide. The early and accurate diagnosis of HD is the key to reduce mortality rate. To diagnose HD, vital signs such as blood sugar, heart rate, electrocardiogram (ECG), blood pressure, cholesterol level, etc. are needed. These biomarkers are generally obtained from sensors, such as pulse sensor, blood pressure sensor, etc., attached to a person's body. Integrating internet of things (IoT), artificial intelligence (AI), and cloud computing technologies automates the HD diagnosis process. IoT devices are used to collect and transmit patient physiological data to an AI-based HD diagnosis model deployed in a cloud computing platform. Many studies developed IoT-based HD diagnostic systems using deep learning (DL) models, a branch of AI. However, most focused on improving diagnostic accuracy by increasing DL model complexity, which leads to more computation time and need for larger training data. In this thesis, different DL models are proposed, developed, and implemented for HD diagnosis. Cleveland, Statlog, and Comprehensive benchmark datasets from UCI are utilized to train and test the models. Traditional approaches including support vector machine (SVM) and artificial neural network (ANN), were used to diagnose HD. Each achieved a satisfactory accuracy level (89.7% and 93.44% respectively). Two simplified DL models based on one-dimensional convolutional neural network (1D-CNN) and bi-directional long short-term memory (Bi-LSTM) were proposed to improve diagnosis accuracy. Both reached 94.96% accuracy. This accuracy could not further enhance due to the small and imbalanced dataset. To overcome this issue, DL-based data generative model employing generative adversarial network (GAN) was developed. The GAN model generates additional data to enlarge the small and imbalanced dataset. Then, the 1D-CNN and Bi-LSTM models are retrained using the enlarger dataset. This boosted 1D-CNN accuracy to 99.1% and Bi-LSTM accuracy to 99.3%. Moreover, principal component analysis (PCA) is used to investigate the time complexity of the proposed 1D-CNN and Bi-LSTM models. PCA reduced the dataset dimensionality which lowered the 1D-CNN and Bi-LSTM prediction times to 68.8 and 74.8 ms, respectively. Finally, ANN, 1D-CNN, and Bi-LSTM models are stacked in an ensemble model to reduce the diagnosis error of each one. The stacked ensemble model improved the overall diagnosis accuracy to 99.7%. Consequently, this model was deployed on the Amazon cloud computing platform to diagnose HD from sensors data and display the diagnosis result on Android application in a complete IoT environment.