Artificial Intelligence for the Prediction of Weaning Readiness Outcome in a Multi-centrical Clinical Cohort of Mechanically Ventilated Patients - Tesi di Laurea

Aug 6, 2026

Upload by Logic

Mechanical ventilation is used to manage acute respiratory failure until independent breathing is possible, with daily clinical screening to decide whether it can be safely stopped. The process involves Readiness Testing (RT) followed by a 30-minute spontaneous breathing trial (SBT), yielding three mutually exclusive daily scenarios: SBT not attempted, SBT failure, or SBT success. This thesis develops a deep-learning, multi-source neural network to predict these outcomes early in the morning using clinical data, prior-day diary information, and minute-by-minute ventilator parameter histories from a retrospective multicenter cohort in Italy over 27 months, optimizing hyperparameters via cross-validation and evaluating final performance on held-out test patients.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages89

Reading Time3 h 44 min

AccessFree to download

Abstract

Structured summary

Introduction

Rationale

Objectives

Methods

Setting

Prediction problem definition

Data preparation

Prediction model building

Model selection

Results

Preliminary models performance

Final model performance

Discussion

Limitations of the model

Clinical implications of model adoption

Interpretability and explainability of the model

Clinical implications of broader AI tools adoption

Conclusions

Bib liografia

Supplementary material

UNIVERSITÀ DEGLI STUDI DI PADOVA Corso di Laurea Magistrale a ciclo unicoin Medicina e Chirurgia DIPARTIMENTO DI SCIENZE CARDIO-TORACO-VASCOLARI E SANITÀ PUBBLICA Direttore: Prof. Federico Rea UNITÀ DI BIOSTATISTICA, EPIDEMIOLOGIA E SALUTE PUBBLICA Direttore: Prof. Dario Gregori TESI DI LAUREA Artificial Intelligence for the prediction of weaning readiness outcome in a multi-centrical clinical cohort of mechanically ventilated patients Relatore: Corrado Lanera, Ph.D. Laureando/a: Andrea Pedot ANNO ACCADEMICO 2022/2023 UNIVERSITÀ DEGLI STUDI DI PADOVA Corso di Laurea Magistrale a ciclo unicoin Medicina e Chirurgia DIPARTIMENTO DI SCIENZE CARDIO-TORACO-VASCOLARI E SANITÀ PUBBLICA Direttore: Prof. Federico Rea UNITÀ DI BIOSTATISTICA, EPIDEMIOLOGIA E SALUTE PUBBLICA Direttore: Prof. Dario Gregori TESI DI LAUREA Artificial Intelligence for the prediction of weaning readiness outcome in a multi-centrical clinical cohort of mechanically ventilated patients Relatore: Corrado Lanera, Ph.D. Laureando/a: Andrea Pedot TABLE OF CONTENTS Table of Contents ............................................................................................................ 1 Abstract [structured summary]....................................................................................... 2 Riass unto ......................................................................................................................... 3 Introduction .................................................................................................................... 4 Rationale ..................................................................................................................... 4 Objectives.................................................................................................................... 7 Methods .......................................................................................................................... 8 Setting ......................................................................................................................... 8 Prediction problem definition ................................................................................... 11 Data preparation ....................................................................................................... 14 Prediction model building ......................................................................................... 22 Model Selection ........................................................................................................ 25 Results ........................................................................................................................... 27 Preliminary models performance ............................................................................. 27 Final model performance .......................................................................................... 28 Discussion...................................................................................................................... 33 Limitations of the model ........................................................................................... 34 Clinical implications of model adoption.................................................................... 36 Interpretability and Explainability of the model ....................................................... 36 Clinical implications of broader AI tools adoption .................................................... 37 Conclusions ................................................................................................................... 40 Bib liografia .................................................................................................................... 41 Supplementary material ............................................................................................... 49 1 ABSTR

FAQs

How does the mechanical ventilation weaning decision process work in this study?

Patients undergo Readiness Testing (RT) followed by a 30-minute spontaneous breathing trial (SBT) if RT is successful. Depending on RT and SBT results, the day falls into one of three mutually exclusive scenarios: SBT not attempted, SBT failure, or SBT success.

What data sources are used to make the early-morning AI predictions?

Predictions are based on clinical data, information from the previous day’s clinical diary, and whole minute-by-minute recording histories of mechanical ventilator parameters collected from a retrospective observational multicenter study.

How is the final deep-learning model evaluated and what is its reported performance?

Hyperparameters are optimized via cross-validation, with 36 out of 182 patients held out for testing. The final AI model reports 79% accuracy and improves on comparison models such as XG Boost trained on daily and baseline data from the previous day.

Aug 6, 2026

Upload by Logic