Abstract Designing automatic cardiovascular disease (CVD) diagnostic systems specifically for signals acquired using wearable electrocardiogram (ECG) sensors becomes a challenge specifically requiring solutions for signal distortions caused by high level of motion artifacts and efficient CVD diagnosis. Hence the aim of this thesis is to develop an adaptation of Segmented Beat Modulation Method (SBMM, a template-based method for denoising of ECG signals) using wearable ECG data to additionally account for non-sinus rhythms and to increase the usability of modern wearable sensors in comparison to traditional in-clinic machines for CVD diagnosis. SBMM has currently failed to work with abnormal or arrhythmic (rare but critical events often leading to sudden cardiac death) heartbeats which hugely limits its applicability to cardiovascular disease diagnosis in a real-world scenario. To this aim, this work presents Extended Segmented Beat Modulation Method with a heartbeat classification function using convolutional neural network (CNN) that first separates the normal (N) from supraventricular (S) and ventricular (V) heartbeats and secondly uses separate median representative templates to denoise and reconstruct the clean ECG recording. Overall, the CNN classification accuracy (Ac) was 91.5% while the positive predictive (PP) values were 92.8%, 95.6%, and 83.6%, for N, S, and V beat classes, respectively. Eventually, signal-to-noise (SNR) improvement was less than 2 dB in the absence of noise but increased in the presence of noise until exceeding 5 dB in the presence of electrode motion artifacts. Hence, ESBMM proved to be a reliable tool to classify cardiac beats into N, S, and V classes and to denoise ECG tracings characterized by both sinus and non-sinus rhythms maintaining the morphological variability in the pseudo-periodic ECG signal. Other improvements proposed to SBMM are a preliminary compression test using discrete cosine transform. The method is evaluated using SNR and compression ratio (CR) considering varying levels of signal energy in the reconstructed ECG signal. For denoising, an average SNR of 4.56 dB was achieved representing an average overall decline of 1.68 dBs (37.9%) as compared to the uncompressed signal processing while 95% of signal energy is intact and quantized at 6 bits for signal storage (CR=2) compared to the original 12 bits, hence resulting in 50% reduction in storage size. Another improvement dynamic-template SBMM adapts SBMM to heart rate and generates the template in a dynamic fashion every 20 seconds and is particularly targeted and tested for long-term ECG data acquisitions. Another presented improvement adapts SBMM to modern fast hardware using vectorization technique and graphical processing units called GPU-SBMM. GPU-SBMM application yielded a significant increase of SNR (from 1±5 dB to 19±5 dB; p<10E-10). Additionally, a considerable speed up in the algorithm runtime (3.56x on average on an NVIDIA GeForce GPU) was achieved. In a secondary domain, an automated arrhythmia detection system is presented that is designed to produce maximum diagnostic accuracy with minimum amount of data (removing redundant and noisy data) using differential evolution (DE) and a less computationally intense probabilistic neural network (PNN). All tests are performed for ambulatory and long term ECG signals acquired using wearable sensing modality. The proposed DE-PNN scheme provides better classification accuracy considering 8 classes with only 41 features optimized from a 253 element feature set implying an 83.7% reduction in direct amplitude features compared to the other evolutionary and statistical schemes. In conclusion, this work has proved beneficial for improving the quality and efficiency of automatic cardiovascular disease diagnosis system on a modern and evolving cardiovascular health monitoring platform i.e. wearable ECG sensors.
Lo scopo di questa tesi è adattare il Segmented Beat Modulation Method (SBMM), un metodo per il filtraggio di segnali electrocardiografici (ECG), per tenere conto sia dei ritmi cardiaci sinusali che non e per aumentare la sua usabilità includendo i moderni sensori indossabili oltre ai tradizionali dispositivi clinici per la diagnosi di patologie cardiovascolari. Infatti, SBMM non è attualmente in grado di funzionare in presenza di battiti cardiaci anormali o aritmici (eventi critici che potrebbero portare allaa morte cardiaca improvvisa), il che limita enormemente la sua applicabilità alla diagnosi di malattie cardiovascolari in uno scenario reale. A questo scopo, questo lavoro presenta il Extended Segmented Beat Modulation Method (ESBMM) con una funzione di classificazione del battito cardiaco utilizzando la convolutional neural network (CNN) che separa prima i battiti cardiaci normali da quelli sopraventricolari (S) e ventricolari (V), e in secondo luogo utilizza modelli rappresentativi mediani separati per filtrare e ricostruire la registrazione ECG pulita. Nel complesso, l’accuratezza (Ac) della classificazione CNN era del 91,5% mentre i valori di predizione positive erano del 92,8%, 95,6% e 83,6%, rispettivamente per le classi di battito N, S e V. Alla fine, il miglioramento del rapporto segnale-rumore (SNR) è stato inferiore a 2 dB in presenza di livelli di rumore trascurabile, ma è aumentato in presenza di rumore fino a superare i 5 dB in presenza di artefatti da movimento degli elettrodi. Pertanto, ESBMM si è dimostrato uno strumento affidabile per classificare i battiti cardiaci in classi N, S e V e per filtraggio di tracciati ECG caratterizzati da ritmi sia sinusali che non sinusali mantenendo la variabilità morfologica nel segnale ECG pseudo-periodico. Altri miglioramenti proposti a SBMM sono un test di compressione preliminare che utilizza la trasformata coseno discreta. Il metodo viene valutato utilizzando SNR e il rapport di compressione (CR) considerando diversi livelli di energia del segnale ECG ricostruito. Per il filtraggio, è stato raggiunto un SNR medio di 4,56 dB che rappresenta un calo complessivo medio di 1,68 dB (37,9%) rispetto all'elaborazione del segnale non compresso mentre il 95% dell'energia del segnale è intatto e quantizzato a 6 bit per la memorizzazione del segnale (CR=2) rispetto ai 12 bit originali, con conseguente riduzione del 50% delle dimensioni di archiviazione. Un altro miglioramento è l’adattamento dell'SBMM alla frequenza cardiaca in modo dinamico ogni 20 secondi, particolarmente indicato per l'acquisizione di dati ECG a lungo termine. Un altro miglioramento presentato adatta SBMM al moderno hardware veloce utilizzando la tecnica di vettorizzazione e le unità di elaborazione grafica chiamate GPU-SBMM. L'applicazione GPU-SBMM ha prodotto un aumento significativo dell'SNR (da 1±5 dB a 19±5 dB; p<10-10). Inoltre, è stata raggiunta una notevole velocità nel runtime dell'algoritmo (3,56x volte GPU NVIDIA GeForce). In aggiunta, viene presentato un sistema automatico di rilevamento dell'aritmia progettato per produrre la massima precisione diagnostica con una quantità minima di dati utilizzando differential evolution (DE) e una probabilistic neural network (PNN) meno pesante dal punto di vista computazionale. Tutti i test sono stati eseguiti su ECG ambulatoriali e a lungo termine acquisiti utilizzando sensoristica indossabile. Lo schema DE-PNN proposto ha fornito una migliore accuratezza di classificazione considerando 8 classi con solo 41 caratteristiche ottimizzate da un insieme di 253 elementi che hanno causato una riduzione dell'83,7% delle caratteristiche di ampiezza diretta. In conclusione, questo lavoro si è dimostrato utile per migliorare la qualità e l'efficienza del sistema di diagnosi automatica delle malattie cardiovascolari su una piattaforma di monitoraggio della salute cardiovascolare moderna e in evoluzione, ovvero sensori ECG indossabili
Adaptation of the Segmented Beat Modulation Method to support diagnosis of cardiovascular disorders using electrocardiographic tracings acquired by wearable sensors / Nasim, Amnah. - (2021 Mar 18).
Adaptation of the Segmented Beat Modulation Method to support diagnosis of cardiovascular disorders using electrocardiographic tracings acquired by wearable sensors
NASIM, AMNAH
2021-03-18
Abstract
Abstract Designing automatic cardiovascular disease (CVD) diagnostic systems specifically for signals acquired using wearable electrocardiogram (ECG) sensors becomes a challenge specifically requiring solutions for signal distortions caused by high level of motion artifacts and efficient CVD diagnosis. Hence the aim of this thesis is to develop an adaptation of Segmented Beat Modulation Method (SBMM, a template-based method for denoising of ECG signals) using wearable ECG data to additionally account for non-sinus rhythms and to increase the usability of modern wearable sensors in comparison to traditional in-clinic machines for CVD diagnosis. SBMM has currently failed to work with abnormal or arrhythmic (rare but critical events often leading to sudden cardiac death) heartbeats which hugely limits its applicability to cardiovascular disease diagnosis in a real-world scenario. To this aim, this work presents Extended Segmented Beat Modulation Method with a heartbeat classification function using convolutional neural network (CNN) that first separates the normal (N) from supraventricular (S) and ventricular (V) heartbeats and secondly uses separate median representative templates to denoise and reconstruct the clean ECG recording. Overall, the CNN classification accuracy (Ac) was 91.5% while the positive predictive (PP) values were 92.8%, 95.6%, and 83.6%, for N, S, and V beat classes, respectively. Eventually, signal-to-noise (SNR) improvement was less than 2 dB in the absence of noise but increased in the presence of noise until exceeding 5 dB in the presence of electrode motion artifacts. Hence, ESBMM proved to be a reliable tool to classify cardiac beats into N, S, and V classes and to denoise ECG tracings characterized by both sinus and non-sinus rhythms maintaining the morphological variability in the pseudo-periodic ECG signal. Other improvements proposed to SBMM are a preliminary compression test using discrete cosine transform. The method is evaluated using SNR and compression ratio (CR) considering varying levels of signal energy in the reconstructed ECG signal. For denoising, an average SNR of 4.56 dB was achieved representing an average overall decline of 1.68 dBs (37.9%) as compared to the uncompressed signal processing while 95% of signal energy is intact and quantized at 6 bits for signal storage (CR=2) compared to the original 12 bits, hence resulting in 50% reduction in storage size. Another improvement dynamic-template SBMM adapts SBMM to heart rate and generates the template in a dynamic fashion every 20 seconds and is particularly targeted and tested for long-term ECG data acquisitions. Another presented improvement adapts SBMM to modern fast hardware using vectorization technique and graphical processing units called GPU-SBMM. GPU-SBMM application yielded a significant increase of SNR (from 1±5 dB to 19±5 dB; p<10E-10). Additionally, a considerable speed up in the algorithm runtime (3.56x on average on an NVIDIA GeForce GPU) was achieved. In a secondary domain, an automated arrhythmia detection system is presented that is designed to produce maximum diagnostic accuracy with minimum amount of data (removing redundant and noisy data) using differential evolution (DE) and a less computationally intense probabilistic neural network (PNN). All tests are performed for ambulatory and long term ECG signals acquired using wearable sensing modality. The proposed DE-PNN scheme provides better classification accuracy considering 8 classes with only 41 features optimized from a 253 element feature set implying an 83.7% reduction in direct amplitude features compared to the other evolutionary and statistical schemes. In conclusion, this work has proved beneficial for improving the quality and efficiency of automatic cardiovascular disease diagnosis system on a modern and evolving cardiovascular health monitoring platform i.e. wearable ECG sensors.File | Dimensione | Formato | |
---|---|---|---|
Tesi_Nasim.pdf
Open Access dal 16/09/2022
Descrizione: Tesi_Nasim
Tipologia:
Tesi di dottorato
Licenza d'uso:
Creative commons
Dimensione
4.81 MB
Formato
Adobe PDF
|
4.81 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.