The work presented in this thesis comprises the elaboration and evaluation of an ECG automatic classifier, which is able to assign multiple pathologies to recordings with a variable number of leads. The aim is to explore a Neural Network structure and its training procedure which are able to successfully exploit both Machine Learning and Deep Learning features in the same model. The data were selected from the multi-source dataset of the CinC/PhysioNet Challenge 2021, considering only 10 seconds, 500 Hz signals for a total of 84176 recordings. Up to 26 classes could be assigned to each track simultaneously. Some pathologies occurred more frequently than others, making the learning of the model biased towards them. To reduce this unbalancing, three subsets were created by dividing in three the Normal Sinus Rhythm and Sinus Bradycardia ECGs, which were the most predominant labels. Such subsets were exploited to train as many different models with the same architecture, used to create an ensemble which labels the samples by majority voting. The models comprised a "deep" branch, which extracted the signal's morphological features, and a "wide" branch, responsible for compressing and selecting the 20 handcrafted features. The output features of both branches were concatenated and fed to the final layer to produce 26 predictions, one for each class. A three step training was employed, inspecting its capability to prevent overfitting in the developed architecture: at first, only the deep part was trained, secondly the wide branch was instructed taking the deep features into account during the process, and finally a finetuning of the deep components was performed considering the wide parameters. The results obtained in the local test set, measured in terms of Challenge Metric (ranging from -1 to 1, worst to best) after the three step training, was 0.705 and 0.674 for 12 and 2-lead models respectively. This evaluation allows to assert that the integration of Deep Learning and handcrafted features is successful in improving the generalization capability, as well as in making the classification more explainable with respect to a pure Deep Learning process.
Il lavoro presentato in questa tesi riguarda l'elaborazione e la valutazione di un classificatore automatico di ECG, capace di assegnare più patologie simultaneamente a tracciati con numero variabile di canali. Lo scopo è l'esplorazione della struttura di una Rete Neurale e di una procedura di addestramento della stessa tali da riuscire a integrare efficacemente elementi di Machine Learning e Deep Learning nello stesso modello. I dati usati sono stati selezionati tra i segnali provenienti dai diversi dataset considerati nella CinC/PhysioNet Challenge 2021, includendo soltanto le acquisizioni di 10 secondi campionate a 500 Hz, per un totale di 84176 tracciati. Ad ogni ECG sono state associate fino a 26 classi contemporaneamente. Alcune patologie erano presenti più frequentemente di altre, rendendo l'addestramento del modello sbilanciato verso di esse. Per ridurre tale sbilanciamento, sono stati realizzati tre sottoinsiemi di dati riducendo a un terzo il numero di segnali associati al Ritmo Normale Sinusale e alla Bradicardia Sinusale. Tali sottogruppi sono stati usati per allenare altrettanti modelli diversi con la stessa architettura, così da sfruttarli per fare un ensemble che assegna l'anomalia cardiaca per "majority voting". Il modello era formato da un ramo deep, che estrae le caratteristiche morfologiche dell'ECG, e da un ramo wide, responsabile della compressione e selezione delle 20 proprietà estratte a mano in precedenza. Gli attributi così identificati e restituiti dalle due parti sono concatenati e dati al layer finale per produrre 26 predizioni, una per ogni classe. Per allenare la rete è stata impiegata una procedura a tre fasi, indagandone la capacità di prevenire l'overfitting nell'architettura sviluppata: per prima cosa, è stato istruito solo il ramo "deep", quindi la parte "wide" è stata addestrata tenendo conto delle caratteristiche identificate nel ramo "deep", ed infine è stato eseguito un finetuning delle componenti "deep" considerando i parametri "wide". I risultati ottenuti sul test set locale, misurati in termini di Metrica della Challenge (che varia tra -1 e 1, dal peggiore al migliore), sono di 0.705 e 0.674 rispettivamente per i modelli a 12 e 2 canali. Questa valutazione permette di sostenere il vantaggio apportato dall'integrazione del Deep Learning con delle proprietà estratte a mano, poiché migliorano le capacità di generalizzazione del modello, e rendono la classificazione più comprensibile rispetto ad un processo di puro Deep Learning.
Combining Machine Learning and Deep Learning techniques for multi-label classification of ECG with variable number of leads
Sansonetti, Andrea
2020/2021
Abstract
The work presented in this thesis comprises the elaboration and evaluation of an ECG automatic classifier, which is able to assign multiple pathologies to recordings with a variable number of leads. The aim is to explore a Neural Network structure and its training procedure which are able to successfully exploit both Machine Learning and Deep Learning features in the same model. The data were selected from the multi-source dataset of the CinC/PhysioNet Challenge 2021, considering only 10 seconds, 500 Hz signals for a total of 84176 recordings. Up to 26 classes could be assigned to each track simultaneously. Some pathologies occurred more frequently than others, making the learning of the model biased towards them. To reduce this unbalancing, three subsets were created by dividing in three the Normal Sinus Rhythm and Sinus Bradycardia ECGs, which were the most predominant labels. Such subsets were exploited to train as many different models with the same architecture, used to create an ensemble which labels the samples by majority voting. The models comprised a "deep" branch, which extracted the signal's morphological features, and a "wide" branch, responsible for compressing and selecting the 20 handcrafted features. The output features of both branches were concatenated and fed to the final layer to produce 26 predictions, one for each class. A three step training was employed, inspecting its capability to prevent overfitting in the developed architecture: at first, only the deep part was trained, secondly the wide branch was instructed taking the deep features into account during the process, and finally a finetuning of the deep components was performed considering the wide parameters. The results obtained in the local test set, measured in terms of Challenge Metric (ranging from -1 to 1, worst to best) after the three step training, was 0.705 and 0.674 for 12 and 2-lead models respectively. This evaluation allows to assert that the integration of Deep Learning and handcrafted features is successful in improving the generalization capability, as well as in making the classification more explainable with respect to a pure Deep Learning process.| File | Dimensione | Formato | |
|---|---|---|---|
|
Combining Machine Learning and Deep Learning Techniques for Multi-Label Classification of ECG with Variable Number of Leads.pdf
Open Access dal 07/04/2023
Descrizione: Tesi con Executive Summary
Dimensione
5.56 MB
Formato
Adobe PDF
|
5.56 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/10589/187954