A Seismocardiogram (SCG) is the recording of the dorso-ventral accelerations, usually those on the sternum, of a subject’s cardiac activity. From this trace it is possible to extract some contractility intervals useful for evaluating cardiac activity. With this method we want to introduce an online approach, that makes use of deep neural networks and classical signal processing, to extract four fiducial points (Mitral valve closure (MC), Atrial valve opening (AO), Atrial valve closure (AC), Mitral valve opening (MO)) in the cardiac cycle, fundamental to evaluate these intervals. The method consists in the segmentation, after an initial filtering phase and using the Electrocardiogram (ECG) signal as reference, of the SCG trace in single beats which will then be used as input of a recurrent neural network that will extract the aforementioned fiducial points. With a dataset from 8 different healthy subjects made of 9244 beats, and using a Leave-one-subject-out cross-validation (LOSO) technique to evaluate the algorithm, the network is able to de- tect 98.72% of the fiducial points with a total RMSE of 5.57 ms (MC: 99.56%/3.54 ms, AO: 99.40%/3.85 ms, AC: 97.73%/7.18 ms, MO: 98.18%/7.77 ms) and 94.16% of the fiducial points within 10 ms from the expert annotated one (MC: 96.88%, AO: 96.24%, AC: 89.94%, MO: 93.57%), therefore, also including those not found, a total of 92.95%.
Un Seismocardiogramma (SCG) è la registrazione delle accelerazioni dorso-ventrali, solitamente quelle sullo sterno, dell’attività cardiaca di un soggetto. Da questa registrazione è possibile estrarre alcuni intervalli di contrattilità utili per valutare l’attività cardiaca. Con questo metodo vogliamo introdurre un approccio online, che utilizza reti neurali profonde e l’elaborazione classica dei segnali, per estrarre quattro punti fiduciali (Chiusura della valvola mitrale (MC), Apertura della valvola atriale (AO), Chiusura della valvola atriale (AC), Apertura della valvola mitrale (MO)) del ciclo cardiaco, fondamentali per valutare questi intervalli. Il metodo consiste nella segmentazione, dopo una prima fase di filtraggio e utilizzando un Elettrocardiogramma (ECG) come riferimento, della traccia seismocardiografica in singoli battiti che verranno poi utilizzati come input di una rete neurale ricorrente che estrarrà i suddetti punti fiduciali. Con un dataset di 8 soggetti sani che include 9244 battiti e utilizzando una tecnica di cross validazione Leave-one-subject-out (LOSO) per valutare l’algoritmo, la rete è in grado di rilevare il 98,72% dei punti fiduciali con un Errore quadratico medio (RMSE) totale di 5,57 ms (MC: 99,56%/3,54 ms, AO: 99,40%/3,85 ms, AC: 97,73%/7,18 ms, MO: 98,18%/7,77 ms) e il 94,16% dei punti fiduciali entro 10 ms da quelli annotati da un esperto (MC: 96,88%, AO: 96,24%, AC: 89,94%, MO: 93,57%), quindi, includendo anche quelli non trovati, un totale di 92,95%.
A deep learning approach for online fiducial points detection in seismocardiographic recordings
Siega Battel, Lorenzo
2020/2021
Abstract
A Seismocardiogram (SCG) is the recording of the dorso-ventral accelerations, usually those on the sternum, of a subject’s cardiac activity. From this trace it is possible to extract some contractility intervals useful for evaluating cardiac activity. With this method we want to introduce an online approach, that makes use of deep neural networks and classical signal processing, to extract four fiducial points (Mitral valve closure (MC), Atrial valve opening (AO), Atrial valve closure (AC), Mitral valve opening (MO)) in the cardiac cycle, fundamental to evaluate these intervals. The method consists in the segmentation, after an initial filtering phase and using the Electrocardiogram (ECG) signal as reference, of the SCG trace in single beats which will then be used as input of a recurrent neural network that will extract the aforementioned fiducial points. With a dataset from 8 different healthy subjects made of 9244 beats, and using a Leave-one-subject-out cross-validation (LOSO) technique to evaluate the algorithm, the network is able to de- tect 98.72% of the fiducial points with a total RMSE of 5.57 ms (MC: 99.56%/3.54 ms, AO: 99.40%/3.85 ms, AC: 97.73%/7.18 ms, MO: 98.18%/7.77 ms) and 94.16% of the fiducial points within 10 ms from the expert annotated one (MC: 96.88%, AO: 96.24%, AC: 89.94%, MO: 93.57%), therefore, also including those not found, a total of 92.95%.File | Dimensione | Formato | |
---|---|---|---|
Executive_Summary.pdf
solo utenti autorizzati dal 19/05/2023
Descrizione: Executive summary
Dimensione
549.1 kB
Formato
Adobe PDF
|
549.1 kB | Adobe PDF | Visualizza/Apri |
Thesis.pdf
solo utenti autorizzati dal 19/05/2023
Descrizione: Thesis
Dimensione
9.36 MB
Formato
Adobe PDF
|
9.36 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/10589/188913