Predicting Cardiovascular Risk with Machine Learning: A Data-Driven Analysis of Blood Pressure and Heart Rate Variability

Blood pressure (BP) is a key vital parameter in assessing cardiovascular diseases. It undergoes continuous dynamic fluctuations across different timescales, spanning durations from seconds to years, driven by a complex interaction of environmental and emotional factors along with cardiovascular regulatory mechanisms. Blood pressure variability (BPV) is defined by the magnitude and patterns of BP fluctuations, quantified through various indices that capture different aspects of BP dynamics. These indices include dispersion measures, which assess the overall spread of blood pressure values; sequence indices, that evaluate patterns of change over time; instability measures, that indicate the variability in blood pressure readings; behavioural indices, that consider factors such as daily activities and stress; frequency domain analyses, that decompose BP fluctuations into their constituent frequencies. Nevertheless, the clinical significance of BPV has not been fully determined, especially when using spot clinic measurements and including diastolic blood pressure in its calculation. To overcome this limitation, several studies recommend the application of Machine Learning (ML) algorithms to predict outcomes in hypertensive patients, leading to early diagnosis of CVD and more personalized treatment plans for patients. The aim of this work is to find a correlation between BPV indices and cardiovascular diseases, in order to use the first as a predictive role in cardiovascular diseases screening. In this study, 24h Ambulatory Blood Pressure Monitoring (ABPM) was conducted for 124 subjects participating to the “Artù” study, performed by Molinette Hospital. Systolic Blood Pressure (SBP), Diastolic Blood Pressure (DBP) and Heart Rate (HR) values were measured by means of Spacelab Ultralite 90207 and Spacelab 90217A (Spacelab Healthcare, USA) for all the subjectsinvolved in the study, of which the information about the occurrence of cardiovascular and/or pressor events was known. Since the sampling frequency of data was very low (one BP value every 15 minutes), frequency analysis and its corresponding indices were excluded from the work. Overall, 30 BPV indices were evaluated, divided in the four groups reported above (dispersion, sequence, instability, behavioural), to which anamnesis data were added. After a preliminary unsuccessful correlation analysis, BPV indices were used as input features to train a supervised Machine Learning model for binary classification. Support Vectors Machine (SVM) and K-Nearest Neighbour (KNN) classifiers have been tested, and finally a Neural Network has been implemented. The output labels for all the employed models were the information about the occurrence of cardiovascular and/or pressor events. The results showed that these features allow a reliable binary classification (Accuracy > 70%, F1-Score > 80%) for the identification of at-risk subjects. In conclusion, machine learning can play a crucial role in disease prevention by identifying patterns in blood pressure variability (BPV), enabling early detection of risk factors and enhancing proactive healthcare strategies. Moreover, these promising results not only suggest a strong foundation for future research, but also indicate that there is ample opportunity for further improvements in predictive accuracy and clinical applications.