Patient-ventilator Asynchronies (PVAs) arise from a mismatch between patient breaths and ventilator-assisted breaths. Individuals requiring Invasive Mechanical Ventilation (IMV) are unable to autonomously perform full breathing, hindering their ability to absorb oxygen and expel carbon dioxide. Mechanical ventilators help in this process by supplying the right amounts of oxygen and facilitating carbon dioxide elimination. Despite the life-saving benefits of mechanical ventilation, the implications of PVAs are still not fully understood. Research on PVAs has revealed their association with prolonged mechanical ventilation, increased respiratory muscle workload, compromised cardiocirculatory function, patient discomfort and elevated mortality rates. Despite the importance of accurate identification and quantification of asynchronies, they often go undetected and underestimated and consequently, inappropriately treated. While prior studies have explored Machine Learning (ML) and neural network methods for Patient-ventilator Asynchrony (PVA) detection, these methods often focused on a limited number of asynchronies. This thesis aims to automatically detect and classify a comprehensive set of events, including ineffective effort, double trigger, auto-trigger, delayed cycling, flow starvation, leak and normal, using Artificial Intelligence (AI) methods. The study compares the performance of ML models, such as Random Forest and XGBoost, with Neural Network models, specifically one-branch and three-branch One-Dimensional Convolutional Neural Networks (1D-CNNs). Leveraging a dataset of 3583 samples from 20 patients, characterized by an unbalanced distribution, preprocessing techniques like temporal extension, zero-padding and resampling were employed to standardize input data across diverse models. Exploring various lengths of uniformity, ranging from 375 frames (3 seconds) to 750 frames (6 seconds), was necessitated by the requirement for input arrays of the same length across all Convolutional Neural Network (CNN) models. Through a rigorous 5-fold cross-validation, the one-branch and three-branch CNNs exhibited weighted F1-scores of 90.0% and 89.0%, respectively, underscoring their effectiveness in PVA classification. Additionally, Random Forest and XGBoost demonstrated F1 scores of 87.2% and 87.4%, respectively. For PVA detection, the one-branch CNN was exclusively tested on four distinct patients, yielding notable weighted F1 scores of 75.0%, 73.1%, 74.2% and 70.7%. These findings highlight the considerable potential of AI in advancing the detection and classification of PVAs, offering implications for enhancing mechanical ventilation strategies and ultimately improving patient outcomes.
Le asincronie paziente ventilatore (PVAs) si verificano quando si presenta uno squilibrio tra i respiri iniziati dal paziente e quelli assistiti dal ventilatore. Gli individui che necessitano di ventilazione meccanica invasiva (IMV) non sono in grado di respirare completamente in modo autonomo, il che compromette la loro capacità di ossigenare ed eliminare efficacemente l’anidride carbonica. I ventilatori meccanici aiutano in questo processo fornendo livelli precisi di ossigeno e assistendo nella rimozione dell’anidride carbonica. Nonostante i benefici salvavita della ventilazione meccanica, le implicazioni delle PVAs non sono ancora completamente comprese. Le ricerche sulle PVAs hanno mostrato la loro associazione con una ventilazione meccanica prolungata, un aumento del lavoro a carico dei muscoli respiratori, compromissione della funzione cardiovascolare, disagio al paziente e un tasso di mortalità elevato. Nonostante la loro importanza, le PVAs spesso rimangono non rilevate, sottovalutate e, di conseguenza, non affrontate adeguatamente. Mentre studi esistenti hanno esplorato metodi di ML e reti neurali per la rilevazione delle PVAs, il loro focus è stato limitato a specifiche asincronie. Questa tesi si propone di rilevare e classificare automaticamente un insieme completo di eventi, comprendenti ineffective effort, double trigger, auto-trigger, delayed cycling, flow starvation, leak e normal utilizzando metodi di intelligenza artificiale. Lo studio confronta attentamente le performance di modelli di ML, come Random Forest e XGBoost, con modelli di reti neurali, in particolare reti neurali convoluzionali (CNN) a uno e tre rami. Utilizzando un dataset di 3583 campioni provenienti da 20 pazienti, caratterizzato da una distribuzione sbilanciata, sono state impiegate tecniche di pre-elaborazione come estensione temporale, zero-padding e ricampionamento per uniformare i dati di input tra modelli diversi. Esplorare diverse lunghezze di uniformità, comprese tra 375 frame (3 secondi) e 750 frame (6 secondi), è stato motivato dalla necessità di avere array di input della stessa lunghezza per tutti i modelli CNN. Dopo aver eseguito una cross-validation a 5 fold, le CNN a uno e tre rami hanno mostrato F1 score pesati di 90.0% e 89.0%, rispettivamente, sottolineando la loro efficacia nella classificazione delle PVAs. Inoltre, Random Forest e XGBoost hanno ottenuto F1 score di 87.2% e 87.4%, rispettivamente. Per la rilevazione delle PVAs, è stata impiegata esclusivamente la CNN a un ramo, testata su quattro pazienti distinti, ottenendo F1 score pesati di 75.0%, 73.1%, 74.2%, e 70.7%. Questi risultati sottolineano il potenziale dell’intelligenza artificiale nel migliorare la rilevazione e la classificazione delle PVAs, con implicazioni per il perfezionamento delle strategie di ventilazione meccanica e il miglioramento dei risultati clinici.
Automatic detection and classification of patient-ventilator asynchronies with artificial intelligence methods
Bianco, Giorgia
2023/2024
Abstract
Patient-ventilator Asynchronies (PVAs) arise from a mismatch between patient breaths and ventilator-assisted breaths. Individuals requiring Invasive Mechanical Ventilation (IMV) are unable to autonomously perform full breathing, hindering their ability to absorb oxygen and expel carbon dioxide. Mechanical ventilators help in this process by supplying the right amounts of oxygen and facilitating carbon dioxide elimination. Despite the life-saving benefits of mechanical ventilation, the implications of PVAs are still not fully understood. Research on PVAs has revealed their association with prolonged mechanical ventilation, increased respiratory muscle workload, compromised cardiocirculatory function, patient discomfort and elevated mortality rates. Despite the importance of accurate identification and quantification of asynchronies, they often go undetected and underestimated and consequently, inappropriately treated. While prior studies have explored Machine Learning (ML) and neural network methods for Patient-ventilator Asynchrony (PVA) detection, these methods often focused on a limited number of asynchronies. This thesis aims to automatically detect and classify a comprehensive set of events, including ineffective effort, double trigger, auto-trigger, delayed cycling, flow starvation, leak and normal, using Artificial Intelligence (AI) methods. The study compares the performance of ML models, such as Random Forest and XGBoost, with Neural Network models, specifically one-branch and three-branch One-Dimensional Convolutional Neural Networks (1D-CNNs). Leveraging a dataset of 3583 samples from 20 patients, characterized by an unbalanced distribution, preprocessing techniques like temporal extension, zero-padding and resampling were employed to standardize input data across diverse models. Exploring various lengths of uniformity, ranging from 375 frames (3 seconds) to 750 frames (6 seconds), was necessitated by the requirement for input arrays of the same length across all Convolutional Neural Network (CNN) models. Through a rigorous 5-fold cross-validation, the one-branch and three-branch CNNs exhibited weighted F1-scores of 90.0% and 89.0%, respectively, underscoring their effectiveness in PVA classification. Additionally, Random Forest and XGBoost demonstrated F1 scores of 87.2% and 87.4%, respectively. For PVA detection, the one-branch CNN was exclusively tested on four distinct patients, yielding notable weighted F1 scores of 75.0%, 73.1%, 74.2% and 70.7%. These findings highlight the considerable potential of AI in advancing the detection and classification of PVAs, offering implications for enhancing mechanical ventilation strategies and ultimately improving patient outcomes.File | Dimensione | Formato | |
---|---|---|---|
2024_04_Bianco_Tesi.pdf
solo utenti autorizzati dal 20/03/2025
Descrizione: Tesi
Dimensione
11.53 MB
Formato
Adobe PDF
|
11.53 MB | Adobe PDF | Visualizza/Apri |
2024_04_Bianco_Executive_Summary.pdf
solo utenti autorizzati dal 20/03/2025
Descrizione: Executive Summary
Dimensione
1.32 MB
Formato
Adobe PDF
|
1.32 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/10589/219826