Cardiac arrest represents one of the leading causes of mortality and disability worldwide. Hypoxic brain injury is highly frequent and often leaves patients in a comatose state, making early neurological prognostication essential to guide therapeutic decisions. Continuous electroencephalographic (EEG) monitoring is a valuable tool for assessing brain function, as it allows the detection of pathological patterns such as burst suppression and seizure activity, which are frequently associated with poor prognosis. In parallel, the electrocardiogram (ECG) and the derived Heart Rate Variability (HRV) are emerging as complementary markers, as they reflect the activity of the autonomic nervous system and its connection to brain function. However, manual interpretation of these biosignals is complex, time-consuming and highly dependent on the operator’s expertise, making large-scale systematic analysis challenging. These limitations have therefore motivated the adoption of artificial intelligence (AI) techniques, aimed at providing faster, more reproducible and objective predictions. In this context, the I-CARE database, which collects EEG, ECG and clinical data from comatose post–cardiac arrest patients, is used to develop and validate multimodal algorithms for neurological outcome prediction. A complete pipeline is implemented, including signal pre-processing, feature extraction, quantitative assessment of seizure activity as a prognostic marker and the application of machine learning, deep learning and hybrid models. Results show that machine learning models based on handcrafted features achieve performance comparable to hybrid approaches, with the best network reaching an AUROC of 0.86. Furthermore, models originally designed for seizure detection in EEG signals are adapted for neurological outcome prediction. Among the different data sources, EEG emerges as the most informative modality, while clinical information, especially shockable rhythm, provides complementary insight. Overall, this work introduces an innovative approach that supports neurological prognosis after cardiac arrest through artificial intelligence.
L’arresto cardiaco è una delle principali cause di mortalità e disabilità a livello mondiale. Le lesioni cerebrali da ipossia sono frequenti e spesso lasciano i pazienti in coma, rendendo necessaria una prognosi neurologica precoce per guidare le decisioni terapeutiche. Il monitoraggio elettroencefalografico (EEG) continuo è un importante strumento per la valutazione della funzione cerebrale e la predizione dell’esito neurologico, poiché consente di rilevare pattern patologici come burst suppression e attività epilettica, spesso associati a prognosi sfavorevole. Parallelamente, l’elettrocardiogramma (ECG) e l’Heart Rate Variability (HRV) stanno emergendo come marcatori complementari, poiché riflettono l’attività del sistema nervoso autonomo e il suo legame con la funzione cerebrale. Tuttavia, la valutazione manuale di questi biosegnali è complessa, dispendiosa e dipendente dall’esperienza dell’operatore, rendendo difficile un’analisi sistematica su larga scala. Tali limitazioni hanno motivato l’uso di tecniche di intelligenza artificiale (AI), volte a fornire previsioni più rapide, riproducibili e oggettive. In questo contesto, il database I-CARE, che raccoglie EEG, ECG e dati clinici di pazienti comatosi post-arresto cardiaco, è stato utilizzato per sviluppare e validare algoritmi multimodali di predizione dell’esito neurologico. È stata implementata una pipeline comprendente pre-processing, estrazione delle features, valutazione quantitativa dell’attività epilettica e applicazione di modelli di machine learning, deep learning e approcci ibridi. I modelli basati su features estratte manualmente raggiungono prestazioni paragonabili agli approcci ibridi, con la miglior rete che raggiunge un AUROC di 0,86. Inoltre, modelli originariamente sviluppati per la rilevazione dell’attività epilettica sono stati riadattati alla predizione dell’esito neurologico. Tra i vari dati, l’EEG risulta il più informativo, mentre le informazioni cliniche, in particolare lo shockable rhythm, offrono un contributo complementare. Complessivamente, il lavoro propone un approccio innovativo per supportare la prognosi neurologica post-arresto cardiaco mediante intelligenza artificiale.
Artificial intelligence and seizure features for predicting neurological recovery post-cardiac arrest
Benetti, Diletta
2024/2025
Abstract
Cardiac arrest represents one of the leading causes of mortality and disability worldwide. Hypoxic brain injury is highly frequent and often leaves patients in a comatose state, making early neurological prognostication essential to guide therapeutic decisions. Continuous electroencephalographic (EEG) monitoring is a valuable tool for assessing brain function, as it allows the detection of pathological patterns such as burst suppression and seizure activity, which are frequently associated with poor prognosis. In parallel, the electrocardiogram (ECG) and the derived Heart Rate Variability (HRV) are emerging as complementary markers, as they reflect the activity of the autonomic nervous system and its connection to brain function. However, manual interpretation of these biosignals is complex, time-consuming and highly dependent on the operator’s expertise, making large-scale systematic analysis challenging. These limitations have therefore motivated the adoption of artificial intelligence (AI) techniques, aimed at providing faster, more reproducible and objective predictions. In this context, the I-CARE database, which collects EEG, ECG and clinical data from comatose post–cardiac arrest patients, is used to develop and validate multimodal algorithms for neurological outcome prediction. A complete pipeline is implemented, including signal pre-processing, feature extraction, quantitative assessment of seizure activity as a prognostic marker and the application of machine learning, deep learning and hybrid models. Results show that machine learning models based on handcrafted features achieve performance comparable to hybrid approaches, with the best network reaching an AUROC of 0.86. Furthermore, models originally designed for seizure detection in EEG signals are adapted for neurological outcome prediction. Among the different data sources, EEG emerges as the most informative modality, while clinical information, especially shockable rhythm, provides complementary insight. Overall, this work introduces an innovative approach that supports neurological prognosis after cardiac arrest through artificial intelligence.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_12_Benetti_Executive Summary.pdf
non accessibile
Descrizione: executive summary
Dimensione
431.18 kB
Formato
Adobe PDF
|
431.18 kB | Adobe PDF | Visualizza/Apri |
|
2025_12_Benetti_Tesi.pdf
non accessibile
Descrizione: testo tesi
Dimensione
3.77 MB
Formato
Adobe PDF
|
3.77 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/10589/246876