Atrial fibrillation (AF) is the most common sustained arrhythmia. It affects around 0.5% of global population. The risk of suffering from AF increases with age and with the ageing of global population, its incidence is continuously increasing. Although AF itself does not represent a life-threatening condition, it is associated with higher mortality and morbidity levels. Despite its high incidence, life-long clinical consequences and high associated costs, we are still lacking with the tools to properly predict and screen the general population. One of the main problems associated with AF is that it increases stroke risk up to a 25%. If a patient is properly diagnosed it undergoes an anticoagulant treatment to prevent such event. However, in many cases AF is diagnosed after a cryptogenic stroke has taken place. With the development of technology and smartphones, more affordable solutions to the general population are being made available, representing a powerful tool for AF screening and monitoring. These devices generate an enormous amount of information daily that cannot be manually annotate by clinicians as typically done in short electrocardiogram (ECG) recordings. There exists a need to develop corresponding software solutions to extract the relevant information from all the data that is being gathered daily. Moreover, it is also imperative to understand what information of these data should be stored. For so the clinical relevance of some cardiac events still needs to be clarified (i.e. premature atrial complexes (PAC). The development of automatic algorithm for ECG signals annotation could play a double role: that of the monitoring of a patients' cardiac activity and that of the annotation of large ECG databases for the study of the clinical relevance of different cardiac events and their relationship with AF. In this thesis algorithms for the automatic annotation of ECG signals of different nature were developed. In a fist stage a methodology for the detection of AF that could be generalized to other arrhythmia based on the RR interval (RRi) analysis was presented. The methodology was based on the observation that each cardiac rhythm produces an specific pattern on the Poincaré Plot and that such pattern can be captured and used to classify further Poincaré Plot patterns. For so, a binarized version of the Poincaré Plot was introduced, the Poincaré Image. Each rhythm's pattern was captured in a Poincaré Atlas which was used to classify further Poincaré Images. Using 4 public PhysioNet databases, the methodology was tuned and tested for AF, atrial bigeminy (AB) and normal sinus rhythm (NSR) in RR signals of 120 s, 60 s, 30 s and 20 s length. The final methodology used normalized mutual information (NMI) and 2D correlation for the classification of new Poincaré Images using a pixel dimension of [40x40] ms. The proposed method could be suitable for its implementation on constant monitoring devices as it was only based on RRi and required a low computational power. Nevertheless, work for the prediction of first-time occurrence of AF still ought to be done to enable the application of prophylactic measures. For so, the role of other atrial abnormal events suspected to be an indicator of future development of AF need of further study. Such is the case of PACs that have been linked in some studies with AF appearance. In order to help elucidate their role we have developed an automatic beat classifier that can detect PACs with high sensitivity. This detector could annotate long-term recording signals which cannot be manually analyzed give their length, and enable the study of their relationship with AF. The proposed model was a random forest ensemble that used morphological and rhythm information from 2-lead ECG Holter signals. Moreover, a complementary detector was developed able to identify ECG segments containing ventricular or atrial beats. It was developed with the aim of reducing the amount of false positives, a problem common to all classifiers present in literature. The segment-wise classifier used only RR intervals which made it more resistant to noise than those classifiers using morphological information. RR intervals were transformed into Poincaré Images and then introduced into a convolutional neural network (CNN). This classifier could aid not only in reducing the false positives on automatic signal annotation, but also in the manual data revision as it could reduce the amount of signal segments that should be analyzed. Finally, artificial intelligence (AI) solutions were explored for the automatic annotation of ECG signals of different lengths and leads acquired in a clinical setting. Two approaches were taken: one based on classic machine learning and another one integrating deep learning with machine learning. For the machine learning model, a random forest ensemble was trained to detect 7 different ECG abnormalities using morphological and temporal ECG features. Regarding the deep learning model, 26 different cardiac abnormalities were targeted. In addition the best training strategy for integrating machine learning features into a deep learning model was explored. Both models were evaluated following the challenge metric proposed by the Computing in Cardiology PhysioNet Challenges of 2020-2021. Automatic detection of numerous cardiac disorders could enable the annotation of large hospital databases. A high amount of cardiac information remains unused because their manual annotation is highly time consuming. However, these data could help in answering diverse clinical questions if analyzed properly. To conclude this thesis has contributed to the automatic annotation of cardiac signals in three different scopes and signal types. The first analysis based on RRi segments transformed into Poincaré Images provided a new methodology that could be applicable for the detection of AF and extended to other arrhythmia. The second analysis instead targeted single beats and provided a tool for the detection of PACs on 2-lead Holter ECGs with the aim of aiding on the study of their role with respect to first-time appearance of AF and stoke. Finally, the third analysis instead provided a tool for the automatic annotation of 26 different cardiac disorders on signals acquired in a clinical setting with variable number of leads. Such tool could be use to take advantage of the huge amount of information that remains unused in hospital's unannotated databases.
La fibrillazione atriale (AF) è l'aritmia sostenuta più comune. Colpisce circa lo 0,5% della popolazione mondiale. Il rischio di soffrire di fibrillazione atriale aumenta con l'età e con l'invecchiamento della popolazione mondiale la sua incidenza è in continuo aumento. Sebbene la AF di per sé non rappresenti una condizione di pericolo di vita, è associata a livelli di mortalità e morbilità più elevati. Nonostante l'elevata incidenza, le conseguenze cliniche che si protraggono per tutta la vita e gli alti costi associati, mancano ancora gli strumenti per prevedere e sottoporre a screening in modo adeguato la popolazione generale. Uno dei problemi principali associati alla AF è che aumenta il rischio di ictus fino al 25%. Se un paziente viene diagnosticato correttamente, viene sottoposto a un trattamento anticoagulante per prevenire tale evento. Tuttavia, in molti casi la AF viene diagnosticata dopo che si è verificato un ictus criptogenetico. Con lo sviluppo della tecnologia e degli smartphone, sono state rese disponibili soluzioni più accessibili alla popolazione generale, che rappresentano un potente strumento per lo screening e il monitoraggio della AF. Questi dispositivi generano quotidianamente un'enorme quantità di informazioni che non possono essere annotate manualmente dai medici, come avviene di solito nelle registrazioni di brevi elettrocardiogrammi (ECG). Esiste la necessità di sviluppare soluzioni software corrispondenti per estrarre le informazioni rilevanti da tutti i dati che vengono raccolti quotidianamente. Inoltre, è indispensabile capire quali informazioni di questi dati debbano essere memorizzate. Per questo motivo, la rilevanza clinica di alcuni eventi cardiaci deve ancora essere chiarita (ad esempio, i complessi atriali prematuri (PAC). Lo sviluppo di un algoritmo automatico per l'annotazione dei segnali ECG potrebbe svolgere un duplice ruolo: quello del monitoraggio dell'attività cardiaca del paziente e quello dell'annotazione di grandi database ECG per lo studio della rilevanza clinica di diversi eventi cardiaci e della loro relazione con la AF. In questa tesi sono stati sviluppati algoritmi per l'annotazione automatica di segnali ECG di diversa natura. In una prima AFse è stata presentata una metodologia per il rilevamento della fibrillazione atriale, che potrebbe essere generalizzata ad altre aritmie, basata sull'analisi dell'intervallo RR (RR). La metodologia si basava sull'osservazione che ogni ritmo cardiaco produce un modello specifico sul diagramma di Poincaré e che tale modello può essere catturato e utilizzato per classificare altri modelli di diagrammi di Poincaré. Per questo motivo è stata introdotta una versione binarizzata del diagramma di Poincaré, l'immagine di Poincaré. Il pattern di ogni ritmo è stato catturato in un Atlante di Poincaré che è stato utilizzato per classificare altre immagini di Poincaré. Utilizzando 4 database pubblici di PhysioNet, la metodologia è stata messa a punto e testata per la fibrillazione atriale, la bigeminia atriale (AB) e il ritmo sinusale normale (NSR) in segnali RR di 120 s, 60 s, 30 s e 20 s di lunghezza. La metodologia finale ha utilizzato l'informazione reciproca normalizzata (NMI) e la correlazione 2D per la classificazione di nuove immagini di Poincaré utilizzando una dimensione dei pixel di [40x40] ms. Il metodo proposto potrebbe essere adatto all'implementazione su dispositivi di monitoraggio costante, poiché si basa solo sulla RRi e richiede una bassa potenza di calcolo. Tuttavia, è ancora necessario lavorare per la previsione della prima comparsa di AF per consentire l'applicazione di misure profilattiche. Per questo motivo, il ruolo di altri eventi anomali atriali sospettati di essere un indicatore del futuro sviluppo della AF necessita di ulteriori studi. È il caso delle PAC, che in alcuni studi sono state collegate alla comparsa di FA. Per contribuire a chiarire il loro ruolo, abbiamo sviluppato un classificatore automatico di battiti in grado di rilevare le PAC con un'elevata sensibilità. Questo rilevatore potrebbe annotare i segnali di registrazione a lungo termine che non possono essere analizzati manualmente, data la loro lunghezza, e consentire lo studio della loro relazione con la FA. Il modello proposto è un ensemble di foreste casuali che utilizza informazioni morfologiche e ritmiche provenienti da segnali ECG Holter a 2 derivazioni. Inoltre, è stato sviluppato un rilevatore complementare in grado di identificare i segmenti ECG contenenti battiti ventricolari o atriali. È stato sviluppato con l'obiettivo di ridurre la quantità di falsi positivi, un problema comune a tutti i classificatori presenti in letteratura. Il classificatore segment-wise utilizza solo gli intervalli RR che lo rendono più resistente al rumore rispetto ai classificatori che utilizzano informazioni morfologiche. Gli intervalli RR sono stati trasformati in immagini di Poincaré e quindi introdotti in una rete neurale convoluzionale (CNN). Questo classificatore potrebbe aiutare non solo a ridurre i falsi positivi nell'annotazione automatica del segnale, ma anche nella revisione manuale dei dati, in quanto potrebbe ridurre la quantità di segmenti di segnale da analizzare. Infine, sono state esplorate soluzioni di intelligenza artificiale (AI) per l'annotazione automatica di segnali ECG di diverse lunghezze e derivazioni acquisiti in ambito clinico. Sono stati adottati due approcci: uno basato sull'apprendimento automatico classico e un altro che integra l'apprendimento profondo con l'apprendimento automatico. Per il modello di apprendimento automatico, è stato addestrato un ensemble di foreste casuali per rilevare 7 diverse anomalie ECG utilizzando caratteristiche ECG morfologiche e temporali. Per quanto riguarda il modello di deep learning, sono state individuate 26 diverse anomalie cardiache. Inoltre, è stata esplorata la migliore strategia di addestramento per integrare le caratteristiche di apprendimento automatico in un modello di apprendimento profondo. Entrambi i modelli sono stati valutati in base alla metrica di sfida proposta dal Computing in Cardiology PhysioNet Challenges del 2020-2021. Il rilevamento automatico di numerosi disturbi cardiaci potrebbe consentire l'annotazione di grandi database ospedalieri. Un'elevata quantità di informazioni cardiache rimane inutilizzata perché la loro annotazione manuale richiede molto tempo. Tuttavia, se analizzati correttamente, questi dati potrebbero aiutare a rispondere a diverse domande cliniche. In conclusione, questa tesi ha contribuito all'annotazione automatica dei segnali cardiaci in tre diversi ambiti e tipi di segnale. La prima analisi, basata su segmenti RRi trasformati in immagini di Poincaré, ha fornito una nuova metodologia che potrebbe essere applicata per il rilevamento della AF ed estesa ad altre aritmie. La seconda analisi, invece, si è rivolta a singoli battiti e ha fornito uno strumento per il rilevamento di PAC su ECG Holter a 2 derivazioni, con l'obiettivo di contribuire allo studio del loro ruolo rispetto alla prima comparsa di AF e stoke. Infine, la terza analisi ha fornito uno strumento per l'annotazione automatica di 26 diversi disturbi cardiaci su segnali acquisiti in ambiente clinico con un numero variabile di derivazioni. Tale strumento potrebbe essere utilizzato per sfruttare l'enorme quantità di informazioni che rimangono inutilizzate nei database non annotati degli ospedali.
Computational tools for atrial arrhythmia detection on ECG signals : the contribution of different leads
Garcia Isla, Guadalupe
2022/2023
Abstract
Atrial fibrillation (AF) is the most common sustained arrhythmia. It affects around 0.5% of global population. The risk of suffering from AF increases with age and with the ageing of global population, its incidence is continuously increasing. Although AF itself does not represent a life-threatening condition, it is associated with higher mortality and morbidity levels. Despite its high incidence, life-long clinical consequences and high associated costs, we are still lacking with the tools to properly predict and screen the general population. One of the main problems associated with AF is that it increases stroke risk up to a 25%. If a patient is properly diagnosed it undergoes an anticoagulant treatment to prevent such event. However, in many cases AF is diagnosed after a cryptogenic stroke has taken place. With the development of technology and smartphones, more affordable solutions to the general population are being made available, representing a powerful tool for AF screening and monitoring. These devices generate an enormous amount of information daily that cannot be manually annotate by clinicians as typically done in short electrocardiogram (ECG) recordings. There exists a need to develop corresponding software solutions to extract the relevant information from all the data that is being gathered daily. Moreover, it is also imperative to understand what information of these data should be stored. For so the clinical relevance of some cardiac events still needs to be clarified (i.e. premature atrial complexes (PAC). The development of automatic algorithm for ECG signals annotation could play a double role: that of the monitoring of a patients' cardiac activity and that of the annotation of large ECG databases for the study of the clinical relevance of different cardiac events and their relationship with AF. In this thesis algorithms for the automatic annotation of ECG signals of different nature were developed. In a fist stage a methodology for the detection of AF that could be generalized to other arrhythmia based on the RR interval (RRi) analysis was presented. The methodology was based on the observation that each cardiac rhythm produces an specific pattern on the Poincaré Plot and that such pattern can be captured and used to classify further Poincaré Plot patterns. For so, a binarized version of the Poincaré Plot was introduced, the Poincaré Image. Each rhythm's pattern was captured in a Poincaré Atlas which was used to classify further Poincaré Images. Using 4 public PhysioNet databases, the methodology was tuned and tested for AF, atrial bigeminy (AB) and normal sinus rhythm (NSR) in RR signals of 120 s, 60 s, 30 s and 20 s length. The final methodology used normalized mutual information (NMI) and 2D correlation for the classification of new Poincaré Images using a pixel dimension of [40x40] ms. The proposed method could be suitable for its implementation on constant monitoring devices as it was only based on RRi and required a low computational power. Nevertheless, work for the prediction of first-time occurrence of AF still ought to be done to enable the application of prophylactic measures. For so, the role of other atrial abnormal events suspected to be an indicator of future development of AF need of further study. Such is the case of PACs that have been linked in some studies with AF appearance. In order to help elucidate their role we have developed an automatic beat classifier that can detect PACs with high sensitivity. This detector could annotate long-term recording signals which cannot be manually analyzed give their length, and enable the study of their relationship with AF. The proposed model was a random forest ensemble that used morphological and rhythm information from 2-lead ECG Holter signals. Moreover, a complementary detector was developed able to identify ECG segments containing ventricular or atrial beats. It was developed with the aim of reducing the amount of false positives, a problem common to all classifiers present in literature. The segment-wise classifier used only RR intervals which made it more resistant to noise than those classifiers using morphological information. RR intervals were transformed into Poincaré Images and then introduced into a convolutional neural network (CNN). This classifier could aid not only in reducing the false positives on automatic signal annotation, but also in the manual data revision as it could reduce the amount of signal segments that should be analyzed. Finally, artificial intelligence (AI) solutions were explored for the automatic annotation of ECG signals of different lengths and leads acquired in a clinical setting. Two approaches were taken: one based on classic machine learning and another one integrating deep learning with machine learning. For the machine learning model, a random forest ensemble was trained to detect 7 different ECG abnormalities using morphological and temporal ECG features. Regarding the deep learning model, 26 different cardiac abnormalities were targeted. In addition the best training strategy for integrating machine learning features into a deep learning model was explored. Both models were evaluated following the challenge metric proposed by the Computing in Cardiology PhysioNet Challenges of 2020-2021. Automatic detection of numerous cardiac disorders could enable the annotation of large hospital databases. A high amount of cardiac information remains unused because their manual annotation is highly time consuming. However, these data could help in answering diverse clinical questions if analyzed properly. To conclude this thesis has contributed to the automatic annotation of cardiac signals in three different scopes and signal types. The first analysis based on RRi segments transformed into Poincaré Images provided a new methodology that could be applicable for the detection of AF and extended to other arrhythmia. The second analysis instead targeted single beats and provided a tool for the detection of PACs on 2-lead Holter ECGs with the aim of aiding on the study of their role with respect to first-time appearance of AF and stoke. Finally, the third analysis instead provided a tool for the automatic annotation of 26 different cardiac disorders on signals acquired in a clinical setting with variable number of leads. Such tool could be use to take advantage of the huge amount of information that remains unused in hospital's unannotated databases.File | Dimensione | Formato | |
---|---|---|---|
Reviewers_response_submit.pdf
non accessibile
Descrizione: Reviewers response
Dimensione
4.62 MB
Formato
Adobe PDF
|
4.62 MB | Adobe PDF | Visualizza/Apri |
PhD_Thesis_GGI_r_final.pdf
accessibile in internet per tutti
Descrizione: PhD Thesis
Dimensione
3.03 MB
Formato
Adobe PDF
|
3.03 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/10589/194924