The goal of this research is to create a deep learning algorithm capable of accurately predicting the likelihood of cardiovascular death for individual patients by analyzing their 16-lead electrocardiograms (ECGs) from the Swiss Atrial Fibrillation cohort study. The model will be optimized to detect abnormalities in the ECG and provide a risk score based on the severity of the abnormalities. The risk score will be useful for identifying patients at high risk for heart disease and tracking the cardiovascular risk of existing patients in real time. The provided dataset consists of a 5-minute ECG recording of 2404 patients, each with different ECG derivations, and a corresponding CSV table with patient information. The table contains two columns for each subject: the duration of observation in days and the presence or absence of a cardiovascular event, where a value of 1 indicates cardiovascular death and 0 means no event was observed. The 16-lead ECGs were filtered and segmented into several windows that corresponded to the inputs of the neural network. A loss function resembling the log-max partial likelihood of a Cox model was created, and a ResNet with 28 layers of architecture was used. The study involved training 40 models on different subsets of patients with the same neural network. The results show a Kaplan Meier curve and an association between the predicted cardiovascular risk by the model and 3 age ranges of patients. The validation metrics include the log-rank test and Concordance index. The study concluded that ECG data can be effectively used by neural network models to identify patients with high risk of cardiovascular events. This demonstrates the potential value of ECG data in medical diagnoses.
L’obiettivo di questa ricerca è creare un algoritmo di deep learning in grado di prevedere in modo accurato la probabilità di morte cardiovascolare per singoli pazienti analizzando i loro elettrocardiogrammi a 16 derivazioni (ECG) dallo studio "Swiss Atrial Fibrillation". Il modello verrà ottimizzato per individuare anomalie nell’ECG e fornire un punteggio di rischio basato sulla gravità di quest’ultime. Il punteggio di rischio sarà utile per iden- tificare pazienti ad alto rischio di malattie cardiache e seguire in tempo reale il rischio cardiovascolare dei pazienti esistenti. Il dataset fornito consiste in una registrazione ECG di 5 minuti di 2404 pazienti, ognuno con diverse derivazioni, e una tabella CSV corrispondente con le informazioni del paziente. La tabella contiene due colonne per ogni soggetto: la durata dell’osservazione in giorni e la presenza o assenza di un evento cardiovascolare, dove un valore di 1 indica la morte cardiovascolare e 0 significa che non è stato osservato alcun evento. Gli ECG a 16 derivazioni sono stati filtrati e segmentati in diverse finestre e corrispondono agli ingressi della rete neurale. È stata creata una loss function simile alla log-max partial likelihood del modello di Cox, ed è stata utilizzata una ResNet con 28 livelli di profondità. Lo studio ha coinvolto l’addestramento di 40 modelli su diversi sottoinsiemi di pazienti con la stessa rete. I risultati mostrano una curva di Kaplan Meier e un’associazione tra il rischio cardiovascolare previsto dal modello e 3 fasce d’età dei pazienti. Le metriche di validazione includono il test di log-rank e il Concordance index. Lo studio ha concluso che i dati ECG possono essere efficacemente utilizzati dai modelli di rete neurale per identificare pazienti ad alto rischio di eventi cardiovascolari. Ciò dimostra il potenziale valore dei dati ECG nelle diagnosi mediche.
Development of a deep neural network for predicting cardiovascular risk from 16-lead ecgs in patients with atrial fibrillation
FELICETTI, FEDERICO
2022/2023
Abstract
The goal of this research is to create a deep learning algorithm capable of accurately predicting the likelihood of cardiovascular death for individual patients by analyzing their 16-lead electrocardiograms (ECGs) from the Swiss Atrial Fibrillation cohort study. The model will be optimized to detect abnormalities in the ECG and provide a risk score based on the severity of the abnormalities. The risk score will be useful for identifying patients at high risk for heart disease and tracking the cardiovascular risk of existing patients in real time. The provided dataset consists of a 5-minute ECG recording of 2404 patients, each with different ECG derivations, and a corresponding CSV table with patient information. The table contains two columns for each subject: the duration of observation in days and the presence or absence of a cardiovascular event, where a value of 1 indicates cardiovascular death and 0 means no event was observed. The 16-lead ECGs were filtered and segmented into several windows that corresponded to the inputs of the neural network. A loss function resembling the log-max partial likelihood of a Cox model was created, and a ResNet with 28 layers of architecture was used. The study involved training 40 models on different subsets of patients with the same neural network. The results show a Kaplan Meier curve and an association between the predicted cardiovascular risk by the model and 3 age ranges of patients. The validation metrics include the log-rank test and Concordance index. The study concluded that ECG data can be effectively used by neural network models to identify patients with high risk of cardiovascular events. This demonstrates the potential value of ECG data in medical diagnoses.File | Dimensione | Formato | |
---|---|---|---|
2023_5_Felicetti.pdf
non accessibile
Descrizione: Thesis report
Dimensione
12.42 MB
Formato
Adobe PDF
|
12.42 MB | Adobe PDF | Visualizza/Apri |
2023_5_Felicetti_EX_SUM.pdf
non accessibile
Descrizione: Executive summary
Dimensione
1.46 MB
Formato
Adobe PDF
|
1.46 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/10589/211286