Cancer therapies are increasingly associated with cardiovascular complications, including atrial fibrillation (AF), which may remain clinically silent while increasing morbidity and mortality. Early detection of conduction abnormalities and repolarization changes is clinically important. Although the 12-lead electrocardiogram (ECG) is the gold standard, wearable devices recording single-lead ECGs offer opportunities for continuous remote monitoring. However, interval measurements from wearable ECGs are not fully validated. We developed and evaluated a transparent, clinically interpretable deep learning (DL) framework for automated estimation of PR and corrected QT (QTc) intervals from single-lead ECGs in oncology patients. Daily recordings over 12 weeks from 40 lung cancer patients were analysed using a U-Net–based DL architecture and compared with the NeuroKit2 algorithm. The method showed good accuracy for PR (mean absolute error (MAE) 14.3 ms vs 24.5 ms) and QTc (MAE 13.6 ms vs 19.1 ms), and smartwatch ECGs agreed closely with paired 12-lead recordings. These findings support automated cardiac interval quantification from wearable ECGs and their potential in scalable remote cardiac monitoring in oncology patients at increased arrhythmic risk.
Le terapie oncologiche sono sempre più associate a complicanze cardiovascolari, tra cui la fibrillazione atriale (FA), che può rimanere clinicamente silente pur aumentando morbidità e mortalità. L’identificazione precoce di anomalie della conduzione e alterazioni della ripolarizzazione è quindi importante. Sebbene l’ECG a 12 derivazioni sia il gold standard, i dispositivi indossabili che registrano ECG a singola derivazione offrono opportunità per il monitoraggio remoto continuo, ma le misurazioni degli intervalli non sono completamente validate. Questo studio ha sviluppato e valutato un framework di deep learning (DL) trasparente e interpretabile per stimare automaticamente gli intervalli PR e QT corretto (QTc) da ECG a singola derivazione in pazienti oncologici. Registrazioni giornaliere per 12 settimane di 40 pazienti con carcinoma polmonare sono state analizzate con un’architettura DL basata su U-Net e confrontate con NeuroKit2. Il metodo ha mostrato buona accuratezza per PR (errore medio assoluto (EMA) 14.3 ms vs 24.5 ms) e QTc (EMA 13.6 ms vs 19.1 ms), con buono accordo tra smartwatch e ECG standard.
SEGMENTAZIONE DI ECG DA SMARTWATCH A SINGOLA DERIVAZIONE BASATA SU DEEP LEARNING CON U-NET PER LA STIMA AUTOMATIZZATA DEGLI INTERVALLI PR E QTc NEI PAZIENTI ONCOLOGICI
DILDA, MIRIAM
2024/2025
Abstract
Cancer therapies are increasingly associated with cardiovascular complications, including atrial fibrillation (AF), which may remain clinically silent while increasing morbidity and mortality. Early detection of conduction abnormalities and repolarization changes is clinically important. Although the 12-lead electrocardiogram (ECG) is the gold standard, wearable devices recording single-lead ECGs offer opportunities for continuous remote monitoring. However, interval measurements from wearable ECGs are not fully validated. We developed and evaluated a transparent, clinically interpretable deep learning (DL) framework for automated estimation of PR and corrected QT (QTc) intervals from single-lead ECGs in oncology patients. Daily recordings over 12 weeks from 40 lung cancer patients were analysed using a U-Net–based DL architecture and compared with the NeuroKit2 algorithm. The method showed good accuracy for PR (mean absolute error (MAE) 14.3 ms vs 24.5 ms) and QTc (MAE 13.6 ms vs 19.1 ms), and smartwatch ECGs agreed closely with paired 12-lead recordings. These findings support automated cardiac interval quantification from wearable ECGs and their potential in scalable remote cardiac monitoring in oncology patients at increased arrhythmic risk.| File | Dimensione | Formato | |
|---|---|---|---|
|
tesi_magistrale_Miriam_Dilda.pdf
accesso aperto
Dimensione
3.14 MB
Formato
Adobe PDF
|
3.14 MB | Adobe PDF | Visualizza/Apri |
È consentito all'utente scaricare e condividere i documenti disponibili a testo pieno in UNITESI UNIPV nel rispetto della licenza Creative Commons del tipo CC BY NC ND.
Per maggiori informazioni e per verifiche sull'eventuale disponibilità del file scrivere a: unitesi@unipv.it.
https://hdl.handle.net/20.500.14239/34983