Automated diagnostic tool for hypertension using deep learning model

Hypertension is a systemic disease manifested by persistent high blood pressure, it is an important health problem because it causes serious complications and is common in the society. Untreated hypertension has been shown to increase the rate of heart failure, coronary heart disease, hemorrhagic and thrombotic stroke, renal failure, peripheral artery disease, aortic dissection, and mortality. Complications of hypertension and related mortality increase in direct proportion to high blood pressure. The purpose of the detection and treatment of hypertension is to reduce the risk of heart, brain, vascular and kidney diseases, and the associated mortality rate. Therefore, it requires a continuous, accurate blood pressure measurement system to identify high risk individuals. In the health sector, an uncomfortable blood pressure holter device is usually attached to the patient for the diagnosis of high blood pressure disease. The aim of the study is to present algorithms for blood pressure prediction that do not occlude continuous and non-invasive vascular access. There are many studies in the literature based on machine learning algorithms for blood pressure prediction, but this study has provided estimation of systolic blood pressure (SBP), diastolic blood pressure (DBP) and mean arterial pressure (MAP) values by focusing on deep learning algorithms. It was aimed to make blood pressure estimates with the features extracted from Electrocardiography (ECG) and Photoplethysmography (PPG) signals. The feature set consists of ECG-based features such as Womersley number, SDI, QRS and QTc interval, and PPG-based features such as photoplethysmograph intensity ratio (PIR). The feature set is used to minimize the error in blood pressure estimation and to explore the relationship of these features with blood pressure. The extracted features were trained with deep learning algorithms such as convolutional neural networks and long short-term memory. Evaluation measures such as mean absolute error (MAE), root mean square error (RMSE), mean square error (MSE) and variance score were also calculated. In the blood pressure estimation of our model, which we obtained by using convolutional neural networks and long-short-term memory algorithms together, it was concluded that the accuracy rate increased. At the end of this thesis, an automated version of blood pressure estimation from electrocardiography and photoplethysmography signals has been implemented, allowing the fastest and most accurate diagnosis of hypertension. The methods used in the noninvasive blood pressure estimation we proposed, by integrating into wearable devices that are very popular in today's technology it can also be used to combat cardiovascular diseases and prevent their risky effects.