Continuous and non-invasive blood pressure monitoring remains a key challenge in the biomedical field, particularly for the early diagnosis and management of cardiovascular diseases. Although conventional measurement techniques are widely used in clinical practice, they are often limited by their intermittent nature, the need for specialized personnel, and low patient compliance during long-term monitoring. These limitations have driven the development of alternative methods that are less invasive, more comfortable, and capable of providing real-time data. This thesis presents a machine learning-based approach for estimating systolic and diastolic blood pressure from the Pulse Pressure Wave (PPW) signal acquired through the Soundi device, developed by Biocubica S.r.l., where this thesis work was carried out. The device integrates a differential pressure sensor placed on the radial artery to acquire PPW signals in a wearable and non-invasive manner. The experimental protocol includes one hour of data recording per subject, divided into postural blocks (standing and lying down), synchronized with reference blood pressure values measured every five minutes by a GIMA ABPM Holter. From each heartbeat, 22 characteristic features were extracted, calculated from the waveform and its derivatives, and used as input for a fully connected neural network designed to predict systolic and diastolic pressure on a beat-by-beat basis through regression. In addition to the Fully Connected Network (FNN), a hybrid CNN+LSTM+FNN architecture was developed and tested, using sequences of consecutive heartbeats to better capture the temporal dynamics of the data. To assess the generalizability of the model across different individuals, a Leave-One-Subject-Out (LOSO) cross-validation strategy was adopted. The LSTM-based approach demonstrated improved predictive accuracy, confirming the effectiveness of deep learning models for time series in continuous blood pressure estimation using Soundi signals.
Il monitoraggio continuo e non invasivo della pressione arteriosa rappresenta ancora una sfida fondamentale nel campo biomedico, soprattutto per la diagnosi precoce e la gestione delle patologie cardiovascolari. Le tecniche di misurazione convenzionali, sebbene ampiamente utilizzate nella pratica clinica, sono spesso limitate dalla loro natura intermittente, dalla necessità di personale specializzato e dalla scarsa compliance del paziente durante il monitoraggio a lungo termine. Queste limitazioni hanno spinto allo sviluppo di metodi alternativi meno invasivi, più confortevoli e in grado di fornire dati in tempo reale. Questa tesi presenta un approccio basato sull’apprendimento automatico per la stima della pressione sistolica e diastolica a partire dal segnale Pulse Pressure Wave (PPW) acquisito tramite il dispositivo Soundi, sviluppato da Biocubica S.r.l presso cui la tesi è stata sviluppata. Il dispositivo integra un sensore di pressione differenziale posizionato sull’arteria radiale per acquisire segnali PPW in modalità indossabile e non invasiva. Il protocollo sperimentale prevede la registrazione di un’ora di dati per soggetto, suddivisa in blocchi posturali (in piedi e sdraiati), sincronizzati con i valori di pressione arteriosa di riferimento forniti ogni cinque minuti da un Holter Pressorio GIMA ABPM. Da ogni battito cardiaco sono state estratte 22 caratteristiche tipiche, calcolate sulla forma d’onda e le sue derivate, che costituiscono l’input per una rete neurale completamente connessa progettata per prevedere la pressione sistolica e diastolica battito per battito tramite regressione. Oltre alla rete Fully Connected (FNN), è stata sviluppata e testata un’architettura ibrida CNN+LSTM+FNN che utilizza sequenze di battiti cardiaci consecutivi per catturare meglio la dinamica temporale dei dati. Per valutare la generalizzabilità del modello tra i diversi individui è stata adottata una strategia di validazione incrociata Leave-One-Subject-Out (LOSO). L’approccio basato su LSTM ha dimostrato un miglioramento dell’accuratezza predittiva, confermando la validità dei modelli deep learning per serie temporali nella stima continua della pressione arteriosa mediante segnali Soundi.
A cuffless based device for the continuous monitoring of blood pressure exploiting a neural network model for new generation microcontrollers
Di Giacomo, Roberto
2024/2025
Abstract
Continuous and non-invasive blood pressure monitoring remains a key challenge in the biomedical field, particularly for the early diagnosis and management of cardiovascular diseases. Although conventional measurement techniques are widely used in clinical practice, they are often limited by their intermittent nature, the need for specialized personnel, and low patient compliance during long-term monitoring. These limitations have driven the development of alternative methods that are less invasive, more comfortable, and capable of providing real-time data. This thesis presents a machine learning-based approach for estimating systolic and diastolic blood pressure from the Pulse Pressure Wave (PPW) signal acquired through the Soundi device, developed by Biocubica S.r.l., where this thesis work was carried out. The device integrates a differential pressure sensor placed on the radial artery to acquire PPW signals in a wearable and non-invasive manner. The experimental protocol includes one hour of data recording per subject, divided into postural blocks (standing and lying down), synchronized with reference blood pressure values measured every five minutes by a GIMA ABPM Holter. From each heartbeat, 22 characteristic features were extracted, calculated from the waveform and its derivatives, and used as input for a fully connected neural network designed to predict systolic and diastolic pressure on a beat-by-beat basis through regression. In addition to the Fully Connected Network (FNN), a hybrid CNN+LSTM+FNN architecture was developed and tested, using sequences of consecutive heartbeats to better capture the temporal dynamics of the data. To assess the generalizability of the model across different individuals, a Leave-One-Subject-Out (LOSO) cross-validation strategy was adopted. The LSTM-based approach demonstrated improved predictive accuracy, confirming the effectiveness of deep learning models for time series in continuous blood pressure estimation using Soundi signals.| File | Dimensione | Formato | |
|---|---|---|---|
|
25_12_diGiacomo_ExecutiveSummary.pdf
solo utenti autorizzati a partire dal 06/11/2026
Descrizione: Executive Summary
Dimensione
586.42 kB
Formato
Adobe PDF
|
586.42 kB | Adobe PDF | Visualizza/Apri |
|
25_12_diGiacomo_Tesi.pdf
solo utenti autorizzati a partire dal 06/11/2026
Descrizione: Thesis Work
Dimensione
5.38 MB
Formato
Adobe PDF
|
5.38 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/245837