The present work presents the design, implementation, and validation of embedded electronic systems for multimodal biomedical signal acquisition and processing, with a focus on cardiovascular and pulmonary monitoring. The work addresses the growing demand for compact, low-power, and intelligent hardware platforms capable of performing continuous physiological measurements in real-world settings. Through two complementary case studies—cuffless blood pressure estimation from photoplethysmographic (PPG) signals and automatic B-line detection in lung ultrasound (LUS) imaging—the thesis demonstrates how the co-design of hardware, signal processing, and machine learning can enable clinically meaningful performance under embedded constraints. The first part of the research focuses on the development of a real-time optical acquisition framework for non-invasive blood pressure monitoring. The system was based on the MAX86916 Evaluation System by Analog Devices, integrating a multi-wavelength optical sensor (blue, green, red, and infrared LEDs) with a high-sensitivity photodiode and a MAX32630FTHR ARM Cortex-M4F microcontroller. Data were sampled at 100 Hz across all wavelengths, preprocessed for noise reduction, and synchronized with reference oscillometric blood pressure measurements. Three machine learning models were evaluated for systolic and diastolic blood pressure estimation: (1) a Multi-Layer Perceptron (MLP) trained on 84 handcrafted temporal and spectral features, (2) a PCA-reduced MLP with 10 principal components, and (3) a Convolutional Neural Network (CNN) trained on 21 features derived from the L2-norm of the PPG signal. In addition, a hybrid CNN–Support Vector Regression (SVR) model was developed, combining deep feature extraction with kernel-based regression. Experimental validation on a dedicated dataset (collected from 81 healthy volunteers) demonstrated that the hybrid CNN–SVR model achieved the best overall performance, reaching a mean absolute error (MAE) of 1.84 mmHg for systolic and 1.98 mmHg for diastolic pressure estimation, with standard deviations of 3.27 mmHg and 3.47 mmHg, respectively. These results satisfy the AAMI/ISO 81060-2 accuracy criteria for non-invasive blood pressure measurement and outperform both traditional regression-based and single-feature models. The second part of the thesis targets automated pulmonary analysis through the detection of B-line artifacts in lung ultrasound images—vertical, hyperechoic reverberations that indicate interstitial fluid accumulation. A new algorithm based on Variational Mode Decomposition (VMD) was developed to enhance and segment B-lines by decomposing ultrasound images into intrinsic mode components characterized by distinct frequency and orientation profiles. The proposed block-wise VMD approach enabled efficient parallel processing and reduced computational cost by 40% compared with standard two-dimensional VMD implementations. The algorithm was validated on a clinical dataset of 116 lung ultrasound frames, annotated by expert radiologists. It achieved a precision of 0.87, recall of 0.79, F1-score of 0.83, and a Dice coefficient of 0.79 when compared to the manual ground truth. Compared with deep learning–based models, the proposed method maintained comparable accuracy but required over 10× less memory and 3× faster inference time, confirming its suitability for embedded deployment. Collectively, these results validate the feasibility of achieving clinically relevant accuracy and efficiency in biomedical signal processing using energy-constrained embedded systems. By integrating sensing hardware, signal conditioning, and data-driven modeling, the thesis establishes a generalizable design framework applicable to a wide range of physiological monitoring tasks. The demonstrated systems—based on real hardware and validated under realistic conditions—bridge the gap between laboratory prototypes and practical, wearable medical technologies. Ultimately, this work advances the field of embedded biomedical engineering by showing that with appropriate co-design methodologies, multi-modal sensing and edge-level intelligence can be effectively combined to realize continuous, interpretable, and personalized healthcare solutions.
Il presente lavoro presenta la progettazione, l’implementazione e la validazione di sistemi elettronici embedded per l’acquisizione e l’elaborazione multimodale di segnali biomedici, con particolare attenzione al monitoraggio cardiovascolare e polmonare. La ricerca affronta la crescente esigenza di piattaforme hardware compatte, a basso consumo energetico e dotate di intelligenza, in grado di eseguire misure fisiologiche continue in contesti reali. Attraverso due casi di studio complementari — la stima non invasiva della pressione arteriosa senza bracciale a partire da segnali fotopletismografici (PPG) e il rilevamento automatico delle B-line nelle immagini di ecografia polmonare (LUS) — la tesi dimostra come la co-progettazione di hardware, tecniche di elaborazione del segnale e modelli di machine learning possa consentire prestazioni clinicamente significative nel rispetto dei vincoli tipici dei sistemi embedded. La prima parte della ricerca è dedicata allo sviluppo di un framework di acquisizione ottica in tempo reale per il monitoraggio non invasivo della pressione arteriosa. Il sistema è stato realizzato utilizzando l’Evaluation System MAX86916 di Analog Devices, che integra un sensore ottico multi-lunghezza d’onda (LED blu, verde, rosso e infrarosso), un fotodiodo ad alta sensibilità e un microcontrollore ARM Cortex-M4F MAX32630FTHR. I dati sono stati campionati a 100 Hz su tutte le lunghezze d’onda, preprocessati per la riduzione del rumore e sincronizzati con misure di riferimento della pressione arteriosa ottenute tramite metodo oscillometrico. Sono stati valutati tre modelli di machine learning per la stima della pressione arteriosa sistolica e diastolica: (1) un Multi-Layer Perceptron (MLP) addestrato su 84 caratteristiche temporali e spettrali estratte manualmente; (2) un MLP con riduzione dimensionale tramite PCA basato su 10 componenti principali; (3) una Convolutional Neural Network (CNN) addestrata su 21 caratteristiche derivate dalla norma L2 del segnale PPG. Inoltre, è stato sviluppato un modello ibrido CNN–Support Vector Regression (SVR), che combina l’estrazione di caratteristiche profonde con una regressione basata su kernel. La validazione sperimentale, condotta su un dataset dedicato raccolto su 81 volontari sani, ha dimostrato che il modello ibrido CNN–SVR offre le migliori prestazioni complessive, raggiungendo un errore assoluto medio (MAE) di 1,84 mmHg per la pressione sistolica e di 1,98 mmHg per la pressione diastolica, con deviazioni standard pari rispettivamente a 3,27 mmHg e 3,47 mmHg. Tali risultati soddisfano i criteri di accuratezza AAMI/ISO 81060-2 per la misurazione non invasiva della pressione arteriosa e superano le prestazioni sia dei modelli di regressione tradizionali sia di quelli basati su singole caratteristiche. La seconda parte della tesi è focalizzata sull’analisi automatica del polmone attraverso il rilevamento delle B-line nelle immagini di ecografia polmonare, ovvero artefatti verticali iperecogeni che indicano l’accumulo di fluido interstiziale. È stato sviluppato un nuovo algoritmo basato sulla Variational Mode Decomposition (VMD) per l’enhancement e la segmentazione delle B-line, mediante la decomposizione delle immagini ecografiche in componenti di modo intrinseco caratterizzate da profili distinti in termini di frequenza e orientamento. L’approccio VMD a blocchi proposto ha consentito un’elaborazione parallela efficiente e una riduzione del costo computazionale del 40% rispetto alle implementazioni standard di VMD bidimensionale. L’algoritmo è stato validato su un dataset clinico di 116 frame di ecografia polmonare, annotati da radiologi esperti, ottenendo una precisione di 0,87, un recall di 0,79, un F1-score di 0,83 e un coefficiente di Dice pari a 0,79 rispetto al ground truth manuale. Rispetto ai modelli basati su deep learning, il metodo proposto ha mantenuto un’accuratezza comparabile, richiedendo però oltre 10 volte meno memoria e tempi di inferenza circa 3 volte più rapidi, confermandone l’idoneità all’implementazione su sistemi embedded. Nel complesso, questi risultati validano la fattibilità di ottenere accuratezza ed efficienza clinicamente rilevanti nell’elaborazione di segnali biomedici mediante sistemi embedded a risorse energetiche limitate. Integrando hardware di sensing, condizionamento del segnale e modellazione data-driven, la tesi definisce un framework di progettazione generalizzabile, applicabile a un’ampia gamma di compiti di monitoraggio fisiologico. I sistemi dimostrati — basati su hardware reale e validati in condizioni realistiche — colmano il divario tra i prototipi di laboratorio e le tecnologie mediche indossabili di uso pratico. In definitiva, questo lavoro contribuisce al progresso dell’ingegneria biomedica embedded dimostrando che, attraverso adeguate metodologie di co-progettazione, il sensing multimodale e l’intelligenza a livello edge possono essere efficacemente integrati per realizzare soluzioni di assistenza sanitaria continue, interpretabili e personalizzate.
Embedded Electronics for Multimodal Biomedical Signal Acquisition and Processing / Botrugno, Chiara. - ELETTRONICO. - (2026).
Embedded Electronics for Multimodal Biomedical Signal Acquisition and Processing
Botrugno, Chiara
2026
Abstract
The present work presents the design, implementation, and validation of embedded electronic systems for multimodal biomedical signal acquisition and processing, with a focus on cardiovascular and pulmonary monitoring. The work addresses the growing demand for compact, low-power, and intelligent hardware platforms capable of performing continuous physiological measurements in real-world settings. Through two complementary case studies—cuffless blood pressure estimation from photoplethysmographic (PPG) signals and automatic B-line detection in lung ultrasound (LUS) imaging—the thesis demonstrates how the co-design of hardware, signal processing, and machine learning can enable clinically meaningful performance under embedded constraints. The first part of the research focuses on the development of a real-time optical acquisition framework for non-invasive blood pressure monitoring. The system was based on the MAX86916 Evaluation System by Analog Devices, integrating a multi-wavelength optical sensor (blue, green, red, and infrared LEDs) with a high-sensitivity photodiode and a MAX32630FTHR ARM Cortex-M4F microcontroller. Data were sampled at 100 Hz across all wavelengths, preprocessed for noise reduction, and synchronized with reference oscillometric blood pressure measurements. Three machine learning models were evaluated for systolic and diastolic blood pressure estimation: (1) a Multi-Layer Perceptron (MLP) trained on 84 handcrafted temporal and spectral features, (2) a PCA-reduced MLP with 10 principal components, and (3) a Convolutional Neural Network (CNN) trained on 21 features derived from the L2-norm of the PPG signal. In addition, a hybrid CNN–Support Vector Regression (SVR) model was developed, combining deep feature extraction with kernel-based regression. Experimental validation on a dedicated dataset (collected from 81 healthy volunteers) demonstrated that the hybrid CNN–SVR model achieved the best overall performance, reaching a mean absolute error (MAE) of 1.84 mmHg for systolic and 1.98 mmHg for diastolic pressure estimation, with standard deviations of 3.27 mmHg and 3.47 mmHg, respectively. These results satisfy the AAMI/ISO 81060-2 accuracy criteria for non-invasive blood pressure measurement and outperform both traditional regression-based and single-feature models. The second part of the thesis targets automated pulmonary analysis through the detection of B-line artifacts in lung ultrasound images—vertical, hyperechoic reverberations that indicate interstitial fluid accumulation. A new algorithm based on Variational Mode Decomposition (VMD) was developed to enhance and segment B-lines by decomposing ultrasound images into intrinsic mode components characterized by distinct frequency and orientation profiles. The proposed block-wise VMD approach enabled efficient parallel processing and reduced computational cost by 40% compared with standard two-dimensional VMD implementations. The algorithm was validated on a clinical dataset of 116 lung ultrasound frames, annotated by expert radiologists. It achieved a precision of 0.87, recall of 0.79, F1-score of 0.83, and a Dice coefficient of 0.79 when compared to the manual ground truth. Compared with deep learning–based models, the proposed method maintained comparable accuracy but required over 10× less memory and 3× faster inference time, confirming its suitability for embedded deployment. Collectively, these results validate the feasibility of achieving clinically relevant accuracy and efficiency in biomedical signal processing using energy-constrained embedded systems. By integrating sensing hardware, signal conditioning, and data-driven modeling, the thesis establishes a generalizable design framework applicable to a wide range of physiological monitoring tasks. The demonstrated systems—based on real hardware and validated under realistic conditions—bridge the gap between laboratory prototypes and practical, wearable medical technologies. Ultimately, this work advances the field of embedded biomedical engineering by showing that with appropriate co-design methodologies, multi-modal sensing and edge-level intelligence can be effectively combined to realize continuous, interpretable, and personalized healthcare solutions.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

