CBL - Campus del Baix Llobregat

Projecte llegit

Títol: Assessment of Cardiovascular Health Using Non-Invasive IoT Devices and Machine Learning Techniques


Estudiants que han llegit aquest projecte:


Director/a: CASANELLA ALONSO, RAMON

Departament: EEL

Títol: Assessment of Cardiovascular Health Using Non-Invasive IoT Devices and Machine Learning Techniques

Data inici oferta: 06-02-2026     Data finalització oferta: 06-10-2026



Estudis d'assignació del projecte:
    MU EM CODAS 1
    MU EM CODAS 2
Tipus: Individual
 
Lloc de realització: EETAC
 
Segon director/a extern: Luis M. Correia (IST Lisboa)
 
Paraules clau:
Machine learning, cardiovascular health, IOT, Ballistocardiography, impedance plethysmography, cardiovascular monitoring, pulse transit time, pulse wave velocity, cardiovascular ageing, blood pressure estimation
 
Descripció del contingut i pla d'activitats:
Cardiovascular diseases remain the leading cause of mortality worldwide, highlighting the need for accessible and continuous monitoring tools that enable early detection and prevention. In this context, non-invasive measurements performed outside clinical environments using wearable and Internet of Things (IoT) devices represent a promising approach for large-scale cardiovascular health assessment.

This Master's thesis will focus on the analysis of cardiovascular signal databases acquired from IoT healthcare devices that include electrical and mechanical biosignals such as electrocardiography (ECG), ballistocardiography (BCG) and impedance plethysmography (IPG), collected from subjects of different ages and physical conditions.

Machine learning and artificial intelligence techniques will be applied to extract relevant features and estimate indicators of cardiovascular health and overall cardiovascular status. The project aims to evaluate the feasibility and accuracy of these approaches for non-hospital-based health monitoring.

Additionally, the work will exploree the potential development of a low-power IoT sensing node for the acquisition of complementary physiological parameters, contributing to the design of future wearable or ambient monitoring systems.

The results are expected to provide insights into data-driven cardiovascular assessment and to support the development of scalable, non-invasive health monitoring technologies for preventive medicine.
 
Overview (resum en anglès):
Accessible, non-invasive cardiovascular monitoring could support earlier assessment outside clinical settings, provided unobtrusive sensors yield reliable information. This thesis investigated whether electrocardiography (ECG), scale-based ballistocardiography (BCG), and hand-to-hand and foot-to-foot impedance plethysmography (IPG) could estimate arterial timing, stiffness, age and blood pressure, using carotid-femoral tonometry and continuous blood pressure as references. Resting recordings underwent physiological cleaning, modality-specific quality control and extraction of timing, waveform and heart-rate-variability features. Regression and machine-learning models were assessed using leave-one-subject-out cross-validation, with nested feature selection where applicable and external validation for selected age models. Blood-pressure analysis assessed relative pressure classification, prediction in unseen participants and personalised tracking after calibration.

The hand-to-hand IPG foot and BCG R-I interval estimated carotid pulse arrival time with errors of approximately 6 and 7 to 10 ms, respectively. BCG R-J and R-K intervals and foot-to-foot IPG timing were associated with femoral arrival time. However, the BCG I-J interval did not reliably estimate carotid-femoral pulse transit time. Exploratory pulse-wave-velocity models combining late IPG morphology with body size achieved a nested baseline-relative R squared of 0.39. BCG waveform features with a timing reference achieved a baseline-relative R squared of 0.45, falling to 0.18 under nested feature selection. For chronological-age estimation, ECG heart-rate variability combined with carotid tonometry achieved a nested R squared of 0.47 and a mean absolute error of 6.7 years. A BCG-only model achieved a nested R squared of 0.33 internally; a different BCG-only model transferred to an independent cohort with an R squared of 0.35 and an error of 9.0 years. Absolute blood pressure was not reliably estimated in unseen participants. Following calibration, IPG tracked resting systolic-pressure fluctuations with a within-participant correlation of 0.64 and a mean absolute error of 3.54 mmHg, improving on a constant calibration baseline by 1.26 mmHg.

These findings support BCG and IPG as sources of cardiovascular timing and age-related information, with potential for calibrated short-term blood-pressure tracking. Small cohorts, limited physiological variation and measurement uncertainty constrain interpretation. Larger, more diverse datasets and independent validation are needed before clinical or home monitoring applications.


© CBLTIC Campus del Baix Llobregat - UPC