Tecniche per l'imputazione di valori mancanti in un dataset per lo svezzamento di pazienti da ventilatore automatico derivato da MIMIC-IV = Advanced Imputation Techniques for Missing Values in a Weaning-Focused Dataset Derived from MIMIC-IV

Missing values are a significant challenge in the application of Artificial Intelligence (AI), particularly in healthcare. Incomplete datasets can lead to biased models and unreliable predictions, making it essential to develop robust methods for data imputation. Traditional approaches, such as replacing missing values with the mean or median, and more advanced statistical methods, have been widely applied. However, recent advancements in Machine Learning (ML) and Deep Learning (DL) offer promising alternatives. In healthcare, time is a crucial factor, especially in Intensive Care Units (ICUs) where timely decisions can impact patient outcomes. Health-related datasets often incorporate time information, which can be leveraged to improve the accuracy of imputation methods. The correlation between variables and their temporal patterns provides a valuable resource for reconstructing missing values in time-series data. This thesis focuses on developing a dataset for ICU patients undergoing weaning from mechanical ventilation, using data derived from MIMIC-IV. The dataset includes physiological variables that are useful for assessing a patient’s readiness to initiate the weaning process. Weaning is a crucial phase in critical care, requiring precise timing to avoid complications associated with prolonged mechanical ventilation, which is linked to higher morbidity and mortality. Studies confirm that the timing of weaning significantly affects the success patient outcomes. To address the issue of missing data, this work employs state-of-the-art deep learning models, specifically the Bidirectional Recurrent Imputation for Time Series (BRITS) network, designed to handle time-series data with missing values. BRITS utilizes both forward and backward sequences to capture temporal dependencies, making it well-suited for imputation in healthcare settings. Additionally, BRITS has been compared with the Multidirectional Recurrent Neural Network(MRNN), another advanced technique for handling sequential data with missing values. The comparison focuses on evaluating the performance of each model using metrics such as Mean Absolute Error (MAE) and Mean Relative Error (MRE) calculated on different percentages of missing values. The ultimate goal of this research is to create a fully imputed, accurate dataset that can assist clinicians in making informed decisions about weaning. By comparing the performance of BRITS and MRNN, this work aims to identify the most effective approach for handling missing data in critical care. Furthermore, the methodology developed can be adapted for other time-sensitive clinical applications, highlighting the broader potential impact of machine learning-based imputation in healthcare.