A Novel Approach For Blood Pressure Prediction Using Machine Learning Techniques

Sedentary lifestyle, unbalanced diet, stress and alcohol abuse are the risk factors of many diseases of worldwide population. Among them hypertension is known as ”the silent killer” being the leading cause of premature death according to World Health Organization (WHO). Hypertension is the primary cause of the development of cardiovascular, kidney and brain diseases. After the development of a pathological cardiovascular condition in which blood pressure is constantly too high, the subject does not manifest immediately significant symptoms, and that is why it often leads to further complications. Although the data show a particularly high incidence (1 in 5 women and 1 in 4 men) hypertension is a risk condition widely preventable and modifiable through interventions on the population and individuals at high risk. Thus, it is an objective of contemporary studies to use the most efficient and innovative technologies that allow society to prevent and monitor the risks and the progress of the problem. In recent years, with the development of miniaturized sensors and increasingly refined measurement techniques, this has become possible. This thesis aims to provide a simple and effective method that will lay the foundation for future improvements in the prevention and monitoring of hypertension. In the current state of the art, it is not possible to have direct and non-invasive measurements of pressure in a reliable and continuous way. For this reason, the researchis evolving by exploiting the non-invasive collection of other physiological signals of human body, such as photoplethysmogram (PPG) and electrocardiogram (ECG). In this application, the development of an intelligent system starts from the study of the signals provided by MIMIC III database, which is a widely furnished online database about physiological data collected in Intensive Care Unit (ICU). It is proposed the processing of the input data by exploiting modern signal processing techniques, which have been implemented in Python programming language. A new approach for informative and consistent dataset construction was performed for the application of regression techniques, such as Linear Regression, Ridge Regression, Support Vector Regression and Random Forest Regression. The dataset contains data of Heart Rate (HR) and Pulse Transit Time (PTT) related to specific time windows has been used to train and then predict blood pressure of patients in a continuous way. Unfortunately, it is not possible to provide a model that could be suitable for every patient, and due to the dependence of blood pressure on other external factors, such as age and drug intake, the algorithm itself needs periodic recalibrations to maintain its degree of accuracy. The final results show prediction errors that are in accordance with the Association for the Advancement of Medical Instrumentation (AAMI) guidelines, obtaining a Mean Absolute Error (MAE) of 1.96 ± 1.44 mmHg on Diastolic Blood Pressure (DBP) values and 3.11±1.89 mmHg on Systolic Blood Pressure (SBP) values. These encouraging results promise future developments of the algorithm for wearable devices, and lay the groundwork for the improvement of an automatic, generic and high-performanced system.