Development of an innovative Machine Learning algorithm for single-site PWV assessment

Cardiovascular disease (CVD) is the leading cause of mortality worldwide, responsible for about 18 million deaths per year. Early detection and treatment of CVD can significantly reduce the risk of premature mortality and support patients in leading healthier lives. Among the several predictive parameters, arterial stiffness represents a valuable indicator of cardiovascular risk, since it describes the rigidity of the arterial wall. This parameter is strongly related to Pulse Wave Velocity (PWV), which refers to the propagation speed of the pressure wave through the circulatory system. Clinically, PWV is determined non-invasively by dividing the distance between two arterial points by the time the pressure wave takes to travel that distance. The carotid-femoral PWV (cfPWV) is the most widely used clinical method, as it provides a comprehensive assessment of the cardiovascular system's overall condition. However, the conventional PWV measurement process is time-consuming and requires considerable operator expertise, particularly for positioning the device at the femoral site, presenting challenges that this project aims to simplify. For this purpose, the thesis focuses on developing an innovative Machine Learning (ML) algorithm for assessing the PWV from a single measurement site. The objective is to develop a pattern recognition framework that analyzes individual waveforms from the carotid artery, providing insights into arterial compliance and enabling an accurate estimation of PWV. The initial dataset has already been constructed by eLiONS research group in collaboration with the Echolab medical team from Città della Salute e della Scienza in Turin. The dataset includes the data from 118 subjects, the features extracted from the pulse wave signal, and the PWV values. The dataset’s features were refined and expanded to enhance study insights. Real signals were compared to ideal ones to detect anomalies and explore their potential correlation with abnormal PWV values. These features and PWV values were then used as inputs and targets in supervised learning models. Two machine learning models were implemented: Bagged Trees (ensembles of decision trees) and Neural Networks. For each model, both regression and classification tasks were performed, with a detailed analysis of their performance. The final pipeline for the ML algorithm utilizes a two-class (normal and anomaly) classifier trained on an augmented dataset through data augmentation with gaussian noise, to achieve balanced classes. Additionally, for each class, the most accurate regressor was chosen from those trained on the complete not-augmented dataset and those trained only on the data from each class. Once the best models were identified, the pipeline was implemented. The two-class classification accuracy obtained with Bagged Trees classifier was 85.71%, and the Mean Absolute Error (MAE) obtained from the final pipeline was found to be 1.138 m/s, using Bagged Trees regressors for each class. Although the preliminary results of this study are not yet optimal, they suggest that this approach is promising and merits further research. Future work should focus on expanding the dataset and incorporating additional subjects from underrepresented classes to achieve class balance without relying on data augmentation techniques. Furthermore, exploring other ML models could further enhance predictive performance for PWV assessment.