الفهرس 
Chapter 1 IntroductionOnly 14 pages are availabe for public view
 |  | from | 167 |  |  |
| | | from | 167 | | |
Abstract
Medical and healthcare systems are critical in everyday life. Their roles are not only for providing medical treatment and care for patients, but also for their roles in preventing difficult medical conditions by providing early diagnosis of these conditions. These systems contain a lot of data, in which the data include electronic records stored in database tables, images produced by radiology devices, signals produced by heart monitoring devices, and data collected from different monitoring devices such as intensive care unit monitoring devices. These data require strong intelligent analytical tools to use it in early prediction of diseases to make immediate medical intervention. The more collected clinical and medical healthcare data, the more knowledge the medical support system may support. Artificial Intelligence (AI) provides these strong intelligent analytical tools, especially Machine Learning (ML) and Deep Learning (DL) algorithms. Integrating AI algorithms with internet of things (IoT) technology enhances the capabilities of medical and healthcare systems in very immediate detection of difficult medical conditions for making very responsive interventions. Hence, real monitoring clinical and healthcare data for patients is the trend of this decade based on IoT. IoT models facilitate human life by easily collecting clinical data remotely for recognizing diseases that are easily treatable if it is diagnosed early. A lot of diseases may be preventable if they can be analyzed or predicted from patient historical and family data. Predicting diagnosis depends on the gathered clinical and physiological data of patients. This thesis provides four different models in which AI algorithms are used in enhancing the performance of medical and health care systems. The first model discusses the heart diseases and other cardiovascular disorders are among the prominent reasons for mortality on the globe. As the foundation of diagnosing cardiovascular disease is arrhythmia detection, this thesis first proposes an IoT-based framework for heart conditions. This framework consists of two models: Heart Attack Detection Model (HADM) and Electrocardiography (ECG) Heartbeat Multiclass-Classification Model (ECG-HMCM). HADM uses ML algorithms, namely k nearest neighbor (KNN), support vector machines (SVM), and random forest (RF) using a dataset which consists of 1190 patients and 14 features. By combining multiple datasets that were previously available separately, one dataset was created. This dataset, which includes five heart datasets including Cleveland: 303 observations; Hungarian: 294 observations; Switzerland: 123 observations; Long Beach VA: 200 observations; and 270 observations in the Stalog (Heart) (CHSLBS) datasets. The total number of observations is 1190, but 272 duplicated rows were found. Hence, the preprocessing step is to DROP this duplication to finally get 918 rows. The results show that this model achieves accuracies of 92.3%, 89.51% and 90.91% when using KNN, SVM, and RF, respectively with hyperparameters optimization using grid search technique. ECG-HMCM used the MIT-BIH Arrhythmia and PTB Diagnostic ECG signals datasets which contain 5 categories with 109446 samples. k Nearest Neighbor (KNN) technique is applied to build ECG-HMCM in addition to the using of grid search algorithm for hyperparameter optimization aiming to improve the accuracy of classification which achieved 97.5%. The proposed framework provides a method that can effectively identify heart attacks and help professionals to choose the best decision for saving human life. The contact methods are the most common for measuring these biological aspects because of their high accuracy. Commonly, choosing the contact techniques requires using Electrocardiography (ECG) or Photoplethysmography (PPG) and other on-body sensors. However, using such sensors requires some preliminary steps to obtain valid measurements. Despite these contact methods are commonly used and having high accuracy, they nonetheless have several drawbacks. Furthermore, these methods are inconvenient and uncomfortable for patients, especially those with severe skin burns that prevent the measurement process in traditional ways. Also, these contact-based methods are inappropriate for older people, adults with sensitive skin, and neonates. In addition to the limitation of the patient’s mobility, these methods also lead to infection spreading. So, adopting non-contact methods for measuring biological aspects has become more popular because of their apparent benefits and low cost. Hence, the thesis proposes a second proposed model for Heart Rate (HR) and Respiration Rate (RR) estimation in a contactless environment. This contactless environment enables the estimation of HR and RR without any need for connected devices to the patient preventing any possibility of infection. This proposed system depends on video-based techniques. These techniques include Eulerian Video Magnification (EVM), MediaPipe pose detection, and MediaPipe face mesh. The model achieves a Mean Absolute Error (MAE) of 2.05 and a Pearson Correlation Coefficient (PCC) of 0.91 for HR estimation using (Reg, Green, Blue) RGB color modes. The model achieved MAE of 2.03 and PCC of 0.86 for HR estimation using HSV color modes. The model also achieved MAE of 1.62 and PCC of 0.45 for RR estimation. The third proposed model is used to make a multi-class classification of three different types of thyroid disease. This model uses extreme gradient boosting (XGBoost) technique which is one of the most potent ML algorithms. This technique is employed with hyperparameters optimization. The model achieves an accuracy of 99%. This accuracy is acceptable compared to the state-of-the-art models. The fourth proposed model uses Convolutional Neural Network (CNN) for the early detection of Alzheimer’s Disease (AD). This model classifies Magnetic Reasoning Imaging (MRI) into four different classes. This model achieved an accuracy of 99.94% which is higher than when compared with achieved accuracy by previous works.