Neurological outcome prediction in comatose cardiac arrest survivors remains a critical challenge for clinicians seeking to balance early prognostication with minimizing the risk of premature withdrawal of life-sustaining therapy. This project focuses on evaluating the potential of artificial intelligence in predicting the neurological outcome in comatose patients post cardiac arrest after returning to spontaneous circulation (ROSC) from electrocardiogram (ECG) signals, at different time windows after the cardiac arrest event happened, to assess the predictive capabilities of artificial intelligence over time, relying on ECG signals alone. Unsupervised and supervised machine learning models, and a combination of them, were used for the task, given a set of features extracted from the electrocardiogram activity data as well as patient metadata. The works from Zabihi et al. (2023) and Takahashi et al. (2023) were used as a baseline, with the difference of using just ECG signals to feed the models instead of the combination of EEG and ECG signals. A systematic review of 61 studies was conducted in parallel to assess the current landscape of machine learning (ML) and deep learning (DL) approaches using EEG and ECG data for neurological outcome prediction, revealing a scarcity of ECG-only methods. Motivated by this gap, multiple ML/DL models were developed and evaluated. Using a wide and heterogeneous set of neurological correlated ECG-related features and classifiers, limited results were obtained with a peak of 0.646+-0.080 in balanced accuracy, slightly comparable with the results of EEG-based and EEG-ECG based models. The success of the deep neural networks models fed with just ECG showed difficulties in classifying patients without obtaining lower levels of FPR, and didn’t produce outcomes comparable with the other models, likely due to the lack of available data, high noise, predictive patterns not strong enough to outweigh noise or variability, and data imbalances. Most of the predictive power is already in the clinical features. This project was conducted in collaboration with the Swiss Federal Institute of Technology Zurich in Switzerland, where the algorithm was developed and experiments were carried out.
La previsione dell’esito neurologico nei pazienti in coma sopravvissuti ad un arresto cardiaco rappresenta ancora oggi una sfida cruciale per i medici, i quali devono bilanciare la necessità di una prognosi precoce con il rischio di un’interruzione prematura delle terapie di supporto vitale. Questo progetto si concentra sulla valutazione del potenziale dell’intelligenza artificiale nella predizione dell’esito neurologico in pazienti comatosi dopo il ritorno alla circolazione spontanea (ROSC) in seguito ad arresto cardiaco, utilizzando esclusivamente i segnali dell’elettrocardiogramma (ECG), analizzati in diverse finestre temporali dopo l’evento, per valutare come varia la capacità predittiva dell’intelligenza artificiale nel tempo. Per questo scopo sono stati impiegati modelli di machine learning supervisionati, non supervisionati, e combinazioni degli stessi, sfruttando un insieme di caratteristiche estratte dai segnali ECG e metadati clinici dei pazienti. I lavori di Zabihi et al. (2023) e Takahashi et al. (2023) sono stati presi come riferimento, con la differenza che in questo progetto si è fatto affidamento esclusivamente sui segnali ECG, anziché su una combinazione di segnali EEG ed ECG. Parallelamente, è stata condotta una revisione sistematica di 61 studi per valutare lo stato dell’arte dei metodi di machine learning (ML) e deep learning (DL) basati su EEG ed ECG nella previsione dell’esito neurologico, evidenziando una carenza di approcci basati unicamente sull’ECG. Motivati da questa lacuna, sono stati sviluppati e valutati diversi modelli ML/DL. Utilizzando un ampio e variegato insieme di caratteristiche ECG correlate a parametri neurologici e differenti classificatori, si sono ottenuti risultati limitati, con una accuratezza bilanciata massima di 0.646+-0.080, lievemente comparabile con quella dei modelli basati su EEG o su combinazioni EEG-ECG. I modelli a rete neurale profonda alimentati solo con segnali ECG hanno mostrato difficoltà nel classificare correttamente i pazienti, risultando in tassi di falsi positivi elevati e performance inferiori rispetto agli altri approcci, probabilmente a causa della scarsità di dati disponibili, dell’alto rumore nei segnali, della debolezza dei pattern predittivi rispetto alla variabilità e agli squilibri nei dati. La maggior parte del potere predittivo risulta risiedere già nei dati clinici. Questo progetto è stato condotto in collaborazione con il Politecnico Federale di Zurigo (ETH Zurich), dove è stato sviluppato l’algoritmo e sono stati eseguiti gli esperimenti.
AI for predicting neurological reovery post-cardiac arrest
Cirrincione, Salvatore
2024/2025
Abstract
Neurological outcome prediction in comatose cardiac arrest survivors remains a critical challenge for clinicians seeking to balance early prognostication with minimizing the risk of premature withdrawal of life-sustaining therapy. This project focuses on evaluating the potential of artificial intelligence in predicting the neurological outcome in comatose patients post cardiac arrest after returning to spontaneous circulation (ROSC) from electrocardiogram (ECG) signals, at different time windows after the cardiac arrest event happened, to assess the predictive capabilities of artificial intelligence over time, relying on ECG signals alone. Unsupervised and supervised machine learning models, and a combination of them, were used for the task, given a set of features extracted from the electrocardiogram activity data as well as patient metadata. The works from Zabihi et al. (2023) and Takahashi et al. (2023) were used as a baseline, with the difference of using just ECG signals to feed the models instead of the combination of EEG and ECG signals. A systematic review of 61 studies was conducted in parallel to assess the current landscape of machine learning (ML) and deep learning (DL) approaches using EEG and ECG data for neurological outcome prediction, revealing a scarcity of ECG-only methods. Motivated by this gap, multiple ML/DL models were developed and evaluated. Using a wide and heterogeneous set of neurological correlated ECG-related features and classifiers, limited results were obtained with a peak of 0.646+-0.080 in balanced accuracy, slightly comparable with the results of EEG-based and EEG-ECG based models. The success of the deep neural networks models fed with just ECG showed difficulties in classifying patients without obtaining lower levels of FPR, and didn’t produce outcomes comparable with the other models, likely due to the lack of available data, high noise, predictive patterns not strong enough to outweigh noise or variability, and data imbalances. Most of the predictive power is already in the clinical features. This project was conducted in collaboration with the Swiss Federal Institute of Technology Zurich in Switzerland, where the algorithm was developed and experiments were carried out.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_10_Cirrincione_Executive_Summary.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Executive summary
Dimensione
541.92 kB
Formato
Adobe PDF
|
541.92 kB | Adobe PDF | Visualizza/Apri |
|
2025_10_Cirrincione_Thesis.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Tesi finale
Dimensione
7.57 MB
Formato
Adobe PDF
|
7.57 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/10589/243901