Sepsis is life-threatening organ dysfunction caused by a dysregulated host response to infection. This thesis presents clinical decision support systems (CDSS) for the management of sepsis in intensive care and uses four studies developed on the MIMIC database as examples of machine learning predictive models. The methods and the reported performance are described: in all four studies the best model is an ensemble of decision trees, with an AUC between 0.75 and 0.88 on data not used for development. Evidence of robustness is however limited: a single external validation, partial assessment of calibration and clinical utility, and retrospective designs. The role of the SHAP method for interpretability is discussed, and the conditions still missing for these models to become part of a CDSS usable in the ward are identified: external and prospective validation, impact studies, integration with the electronic health record, and management of the risks of use.
La sepsi è una disfunzione d’organo, potenzialmente letale, causata da una risposta disregolata dell’ospite a un’infezione. Questo elaborato presenta i sistemi di supporto alla decisione clinica (CDSS) nella gestione della sepsi in terapia intensiva ed illustra quattro studi sviluppati sul database MIMIC come esempi di modelli predittivi di machine learning. Sono descritti i metodi e le prestazioni riportate: in tutti e quattro gli studi il miglior modello è un ensemble di alberi decisionali, con AUC da 0,75 a 0,88 su dati non usati per lo sviluppo. Le evidenze sulla robustezza sono però limitate: una sola validazione esterna, calibrazione e utilità clinica valutate in modo parziale, disegni retrospettivi. Si discute il ruolo del metodo SHAP per l’interpretabilità e si individuano le condizioni che mancano perché questi modelli diventino parte di un CDSS utilizzabile in reparto: validazione esterna e prospettica, studi di impatto, integrazione con la cartella clinica, gestione dei rischi d’uso.
Sistemi di supporto alla decisione clinica nella gestione del paziente con sepsi in reparti di terapia intensiva
PERINI, GIACOMO
2025/2026
Abstract
Sepsis is life-threatening organ dysfunction caused by a dysregulated host response to infection. This thesis presents clinical decision support systems (CDSS) for the management of sepsis in intensive care and uses four studies developed on the MIMIC database as examples of machine learning predictive models. The methods and the reported performance are described: in all four studies the best model is an ensemble of decision trees, with an AUC between 0.75 and 0.88 on data not used for development. Evidence of robustness is however limited: a single external validation, partial assessment of calibration and clinical utility, and retrospective designs. The role of the SHAP method for interpretability is discussed, and the conditions still missing for these models to become part of a CDSS usable in the ward are identified: external and prospective validation, impact studies, integration with the electronic health record, and management of the risks of use.| File | Dimensione | Formato | |
|---|---|---|---|
|
Perini_Giacomo.pdf
accesso aperto
Dimensione
2.18 MB
Formato
Adobe PDF
|
2.18 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/120028