Seismocardiography is a noninvasive technique for evaluating the cardiac activity based on the measurement and analysis of the superficial accelerations of the human chest caused by the beating heart. The recording of a seismocardiogram can be nowadays performed by small and lightweight wearable systems embedding an accelerometer. This enables the estimation of a series of cardiac contractility parameters during subjects' spontaneous behaviour and outside a controlled laboratory environment. The aim of this thesis was to develop an algorithm for the classification and analysis of seismocardiographic recordings. Our main focus was to support the well-established signal processing tools generally used by experts in this field with a data-driven solution. The implemented algorithm is composed of three sequential tasks. The first task is the classification of each heart beat (distinguishing between good- and bad- quality beats), needed to exclude beats affected by movements or other types of artefact from the analysis. The classification was based on the comparison of each beat waveform with a recording-specific template. The second task is the automatic estimation, for each beat, of four fiducial points, from which a series of cardiac parameters can be derived. This task consists in identifying the position of the fiducial points inside the beat waveform and was carried out using an alignment technique supported by clustering. The third task is the creation of families of signal shapes, useful to improve the signal inspection, and was implemented as a tree-structured hierarchy of clustering instances. The algorithm performance was evaluated by analysing six recordings from different subjects and taken in different setups, including spontaneous behaviour and cardiovascular stress tests. The classification of beats as good- or bad- quality consistently achieved a Positive Predictive Value (PPV) >= 0.995. The detection of the fiducial points reached a PPV >= 0.8, computed by considering the number of beats in which all the fiducial points were correctly identified. The clustering of signal shapes is promising, even though future refinements might be needed to discriminate the finest details of the waveform. Our performance suggests that seismocardiography could benefit from data-driven approaches in the future and that useful tools could be developed to support the signal interpretation by doctors and researchers.
La sismocardiografia è una tecnica non invasiva per la valutazione dell'attività cardiaca basata sulla misurazione e l'analisi delle accelerazioni superficiali del torace umano causate dal battito cardiaco. La registrazione di un sismocardiogramma può attualmente essere effettuata tramite sistemi wearable di dimensioni e peso ridotti dotati di un accelerometro. Ciò consente la stima di una serie di parametri di contrattilità cardiaca su soggetti in movimento spontaneo e fuori da un contesto controllato di laboratorio. Lo scopo di questa tesi è stato lo sviluppo di un algoritmo per la classificazione e l'analisi di registrazioni sismocardiografiche, cercando di supportare i consolidati strumenti di elaborazione dei segnali, generalmente applicati da esperti in questo campo, con una soluzione data-driven. L'algoritmo implementato si compone di tre procedure sequenziali. La prima procedura è la classificazione di ogni battito cardiaco (distinguendo tra battiti di buona o cattiva qualità), necessaria per escludere i battiti affetti da movimenti o altri tipi di artefatto. La classificazione è stata basata sulla comparazione della forma d'onda di ciascun battito con un template, specifico per ogni registrazione. La seconda procedura è la identificatione automatica, per ogni battito, di quattro punti fiduciali, da cui è possibile derivare alcuni parametri cardiaci. Questa procedura consiste nell'individuazione di tali punti fiduciali all'interno della forma d'onda ed è stata implementata applicando una tecnica di allineamento supportata da clustering. La terza procedura è la creazione di famiglie di forme d'onda, utile per favorire l'indagine del segnale, ed è stata implementata tramite una serie di istanze di clustering, eseguite seguendo una struttura gerarchica ad albero. Le performance dell'algoritmo sono state valutate analizzando sei registrazioni da diversi soggetti, effettuate in diverse condizioni, inclusi il movimento spontaneo e lo svolgimento di uno stress test cardiovascolare. La procedura di classificazione dei battiti in buona o cattiva qualità ha regolarmente raggiunto un PPV >= 0.995. L'identificazione dei punti fiduciali ha raggiunto un PPV >= 0.8, indice calcolato a partire dal numero di battiti in cui tutti i punti fiduciali sono stati individuati correttamente. Il raggruppamento delle forme d'onda è promettente, ma saranno necessarie rifiniture future per discriminare i dettagli più minuti del segnale. Tali risultati suggeriscono che la sismocardiografia potrebbe in futuro trarre vantaggi da approcci data-driven e che utili strumenti potrebbero essere sviluppati a supporto dell'interpretazione del segnale da parte di medici e ricercatori.
Towards an automatic beat-to-beat classification and analysis of seismocardiographic recordings
CIUSA, ANTONIO
2019/2020
Abstract
Seismocardiography is a noninvasive technique for evaluating the cardiac activity based on the measurement and analysis of the superficial accelerations of the human chest caused by the beating heart. The recording of a seismocardiogram can be nowadays performed by small and lightweight wearable systems embedding an accelerometer. This enables the estimation of a series of cardiac contractility parameters during subjects' spontaneous behaviour and outside a controlled laboratory environment. The aim of this thesis was to develop an algorithm for the classification and analysis of seismocardiographic recordings. Our main focus was to support the well-established signal processing tools generally used by experts in this field with a data-driven solution. The implemented algorithm is composed of three sequential tasks. The first task is the classification of each heart beat (distinguishing between good- and bad- quality beats), needed to exclude beats affected by movements or other types of artefact from the analysis. The classification was based on the comparison of each beat waveform with a recording-specific template. The second task is the automatic estimation, for each beat, of four fiducial points, from which a series of cardiac parameters can be derived. This task consists in identifying the position of the fiducial points inside the beat waveform and was carried out using an alignment technique supported by clustering. The third task is the creation of families of signal shapes, useful to improve the signal inspection, and was implemented as a tree-structured hierarchy of clustering instances. The algorithm performance was evaluated by analysing six recordings from different subjects and taken in different setups, including spontaneous behaviour and cardiovascular stress tests. The classification of beats as good- or bad- quality consistently achieved a Positive Predictive Value (PPV) >= 0.995. The detection of the fiducial points reached a PPV >= 0.8, computed by considering the number of beats in which all the fiducial points were correctly identified. The clustering of signal shapes is promising, even though future refinements might be needed to discriminate the finest details of the waveform. Our performance suggests that seismocardiography could benefit from data-driven approaches in the future and that useful tools could be developed to support the signal interpretation by doctors and researchers.| File | Dimensione | Formato | |
|---|---|---|---|
|
TESI_CIUSA_final_with_ack.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: 25_11_2020 23_14
Dimensione
9.66 MB
Formato
Adobe PDF
|
9.66 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/10589/175193