Tesi etd-05242024-111320
Link copiato negli appunti
Tipo di tesi
Dottorato
Autore
BACHI, LORENZO
URN
etd-05242024-111320
Titolo
Traditional and Machine Learning Novel Methods for Processing and Simulation of Ambulatory Electrocardiogram
Settore scientifico disciplinare
ING-INF/06
Corso di studi
Istituto di Scienze della Vita - PHD IN MEDICINA TRASLAZIONALE
Commissione
relatore Prof. PASSINO, CLAUDIO
Parole chiave
- ecg processing
- ecg simulation
- electrocardiogram
- machine learning
- noise detection
- qrs detection
- signal processing
Data inizio appello
18/12/2024;
Disponibilità
parziale
Riassunto analitico
Introduction: The electrocardiogram (ECG) is an established diagnostic technique with relatively low complexity and cost. Since the introduction of the Holter monitor, automatic analysis of ECG recordings has become an essential component of medical devices involving ECG. Automatic ECG analysis comprises many software procedures involving signal processing techniques, which are used to measure target quantities examined by the clinician. While effective processing techniques have indeed been proposed over the years by researchers worldwide, there are instances where typical problems of automatic ECG analysis could be solved with greater accuracy, efficiency, and generalizability on real-world data. This is especially true for the particularly challenging context of ambulatory, long-term ECG recording. Advances in automatic ECG analysis in recent years have been fueled by three main factors: an increasing number of large, public databases have been released; the available processing power of devices keeps increasing; the rapidly evolving knowledge on machine learning and artificial intelligence techniques and related best practices, which may yield substantial performance improvements.
Aims: The three research activities described in this thesis proposed novel methods to improve automatic ECG analysis or develop useful tools for ECG analysis algorithms. More specifically:
• To develop and test a novel QRS detection algorithm to correctly detect heartbeats on ambulatory, long-term ECG with low computational complexity.
• To develop and test a novel ECG simulator tool where the presence of arrhythmias, realistic noise, and the effects of respiration are modeled to offer a data augmentation valuable tool in machine learning applications.
• To develop a novel noise detection procedure based on supervised machine learning in a novel dataset that includes ventricular arrhythmias, a rare event of prime importance often diagnosed through long-term, ambulatory ECG recordings.
These three research activities will be hereafter referred to as Study 1 (S1), Study 2 (S2), and Study 3 (S3), respectively.
Methods: In S1, the QRS detection algorithm was first designed with a knowledge-based approach, using established knowledge of the ECG pattern to develop the three stages of the algorithm. These stages include noise rejection, QRS enhancement, and decision logic. The first two stages of the algorithm were based on the combination of moving average filters, an efficient and robust technique. The algorithm was tested on six public databases and compared to three other public QRS detector algorithms, as this allowed for direct performance comparison. In S2, the ECG simulation modeling was advanced by modeling arrhythmia episodes, noise, the influence of respiration, and heart rate using statistics from publicly available datasets and mathematical modeling of known phenomena. The simulation model was tested in three different evaluations: first, medical doctors with experience in ECG interpretation evaluated a collection of real and simulated ECGs, blinded to the origin of each record. Secondly, the usefulness of simulated data was demonstrated in the context of quality control in atrial fibrillation (AF) detection, where short bursts of noise can be mistakenly detected as an arrhythmia episode. Lastly, simulated data was used with real data in atrial fibrillation detection in the presence of other arrhythmias through a convolutional neural network. In S3, classification models for noise detection were trained on data from four different databases, one of which was simulated through the simulator proposed in S2. Classification models were trained using a feature set of twelve noise indexes, ten of which are established or promising indexes from the scientific literature. Two of the noise indexes were novel measures we proposed. Noise indexes were computed on annotated data manually labeled through a MATLAB visual interface. Models were trained through a multi-stage optimization procedure, including feature selection and hyperparameter optimization through cross-validation.
Results: S1. The proposed QRS detection approach proved more effective and efficient than the other considered QRS detectors. Our approach reached a gross F1 score of 99.2% over six public databases and a gross sensitivity over atypical beats (not annotated as normal, sinus beats) of 94.8%. Additionally, we measured execution time with the same initial conditions for all algorithms, where our approach proved to be the fastest.
S2. In study 2, an advanced ECG simulator was proposed and evaluated in three experiments. First, the realism of simulated ECG was assessed by medical doctors with experience in ECG interpretation, who were presented with a collection of 100 real ECGs and 100 simulated ECGs and were blinded to the origin of each ECG. On average, the three doctors labeled as realistic 79 out of 100 simulated ECGs, while 84 real ECGs were correctly labeled as realistic. The simulator was also used to improve signal quality control in detecting short episodes of AF detection, where simulated ECGs were used to train a convolutional neural network for AF detection. Quality control improved the ability of the neural network to detect AF, even when simulated data was used, reducing the false positive rate from 87.5% to 65.0% and increasing the positive predictive value from 4.6% to 5.8%. Lastly, the usefulness of simulated data proved beneficial in training a convolutional neural network for AF detection in the presence of other arrhythmias that may trigger false alarms. When trained on real ECG data and tested on simulated data, the convolutional neural network's detection performance was similar to when the opposite was done, i.e., when the network was trained on simulated data and tested on real data.
S3. In study 3, noise detection in the presence of ventricular arrhythmias was conducted by training six different classification models on data, which resulted from merging four data sets, one of which was simulated using the ECG simulator. Noise detection was performed on both 2 s and 10 s windows, with comparable results, reaching a test performance Matthew’s Correlation Coefficient (MCC) as high as 0.85 on 2 s window data and 0.82 on 10 s window data for the Support Vector Machine with Gaussian kernel, with an accuracy of 92.0% and 92.3%, respectively. Specificity on records of clen ECG with ventricular arrhythmias was also evaluated, with the SVM reaching 96.1% on 2 s window data and 96.4% on 10 s window data. The feature selection highlighted the most predictive classification indexes for each classification model. The two novel noise indexes proposed by the authors were consistently included in the selected feature set of most classifiers.
Conclusions: In the research work presented here, three contributions to ECG automatic analysis were designed and successfully developed, which aimed at 1) accurate and efficient heartbeat (QRS) detection, 2) advanced ECG simulation in time-varying conditions, and 3) noise detection in the presence of ventricular arrhythmias for ambulatory ECG. These tools, albeit different in scope, are all intended to innovate on ECG automatic analysis. The research activities mainly focused on accurately detecting and modeling abnormal ECG patterns, which often involve rare clinically relevant events, particularly in ambulatory ECG.
Aims: The three research activities described in this thesis proposed novel methods to improve automatic ECG analysis or develop useful tools for ECG analysis algorithms. More specifically:
• To develop and test a novel QRS detection algorithm to correctly detect heartbeats on ambulatory, long-term ECG with low computational complexity.
• To develop and test a novel ECG simulator tool where the presence of arrhythmias, realistic noise, and the effects of respiration are modeled to offer a data augmentation valuable tool in machine learning applications.
• To develop a novel noise detection procedure based on supervised machine learning in a novel dataset that includes ventricular arrhythmias, a rare event of prime importance often diagnosed through long-term, ambulatory ECG recordings.
These three research activities will be hereafter referred to as Study 1 (S1), Study 2 (S2), and Study 3 (S3), respectively.
Methods: In S1, the QRS detection algorithm was first designed with a knowledge-based approach, using established knowledge of the ECG pattern to develop the three stages of the algorithm. These stages include noise rejection, QRS enhancement, and decision logic. The first two stages of the algorithm were based on the combination of moving average filters, an efficient and robust technique. The algorithm was tested on six public databases and compared to three other public QRS detector algorithms, as this allowed for direct performance comparison. In S2, the ECG simulation modeling was advanced by modeling arrhythmia episodes, noise, the influence of respiration, and heart rate using statistics from publicly available datasets and mathematical modeling of known phenomena. The simulation model was tested in three different evaluations: first, medical doctors with experience in ECG interpretation evaluated a collection of real and simulated ECGs, blinded to the origin of each record. Secondly, the usefulness of simulated data was demonstrated in the context of quality control in atrial fibrillation (AF) detection, where short bursts of noise can be mistakenly detected as an arrhythmia episode. Lastly, simulated data was used with real data in atrial fibrillation detection in the presence of other arrhythmias through a convolutional neural network. In S3, classification models for noise detection were trained on data from four different databases, one of which was simulated through the simulator proposed in S2. Classification models were trained using a feature set of twelve noise indexes, ten of which are established or promising indexes from the scientific literature. Two of the noise indexes were novel measures we proposed. Noise indexes were computed on annotated data manually labeled through a MATLAB visual interface. Models were trained through a multi-stage optimization procedure, including feature selection and hyperparameter optimization through cross-validation.
Results: S1. The proposed QRS detection approach proved more effective and efficient than the other considered QRS detectors. Our approach reached a gross F1 score of 99.2% over six public databases and a gross sensitivity over atypical beats (not annotated as normal, sinus beats) of 94.8%. Additionally, we measured execution time with the same initial conditions for all algorithms, where our approach proved to be the fastest.
S2. In study 2, an advanced ECG simulator was proposed and evaluated in three experiments. First, the realism of simulated ECG was assessed by medical doctors with experience in ECG interpretation, who were presented with a collection of 100 real ECGs and 100 simulated ECGs and were blinded to the origin of each ECG. On average, the three doctors labeled as realistic 79 out of 100 simulated ECGs, while 84 real ECGs were correctly labeled as realistic. The simulator was also used to improve signal quality control in detecting short episodes of AF detection, where simulated ECGs were used to train a convolutional neural network for AF detection. Quality control improved the ability of the neural network to detect AF, even when simulated data was used, reducing the false positive rate from 87.5% to 65.0% and increasing the positive predictive value from 4.6% to 5.8%. Lastly, the usefulness of simulated data proved beneficial in training a convolutional neural network for AF detection in the presence of other arrhythmias that may trigger false alarms. When trained on real ECG data and tested on simulated data, the convolutional neural network's detection performance was similar to when the opposite was done, i.e., when the network was trained on simulated data and tested on real data.
S3. In study 3, noise detection in the presence of ventricular arrhythmias was conducted by training six different classification models on data, which resulted from merging four data sets, one of which was simulated using the ECG simulator. Noise detection was performed on both 2 s and 10 s windows, with comparable results, reaching a test performance Matthew’s Correlation Coefficient (MCC) as high as 0.85 on 2 s window data and 0.82 on 10 s window data for the Support Vector Machine with Gaussian kernel, with an accuracy of 92.0% and 92.3%, respectively. Specificity on records of clen ECG with ventricular arrhythmias was also evaluated, with the SVM reaching 96.1% on 2 s window data and 96.4% on 10 s window data. The feature selection highlighted the most predictive classification indexes for each classification model. The two novel noise indexes proposed by the authors were consistently included in the selected feature set of most classifiers.
Conclusions: In the research work presented here, three contributions to ECG automatic analysis were designed and successfully developed, which aimed at 1) accurate and efficient heartbeat (QRS) detection, 2) advanced ECG simulation in time-varying conditions, and 3) noise detection in the presence of ventricular arrhythmias for ambulatory ECG. These tools, albeit different in scope, are all intended to innovate on ECG automatic analysis. The research activities mainly focused on accurately detecting and modeling abnormal ECG patterns, which often involve rare clinically relevant events, particularly in ambulatory ECG.
File
| Nome file | Dimensione |
|---|---|
Ci sono 1 file riservati su richiesta dell'autore. |
|