Intensive care outcome prediction using machine learning

Weaning from mechanical ventilation represents a major clinical challenge in the Intensive Care Unit (ICU), where both premature and delayed extubation significantly increase morbidity and mortality. This thesis develops machine-learning models to predict the outcome of 30-minute pressure-support Spontaneous Breathing Trials (SBTs) using physiological time-series data. The study is based on a retrospective dataset that is broader than most prior weaning-specific ML studies, encompassing all three weaning-difficulty groups from a general ICU population rather than a narrow clinical cohort. The input data consist solely of routinely collected monitoring signals, enabling future clinical implementation without requiring laboratory results or historical medical information. The dataset required extensive reconstruction to address missing patient identifiers, fragmented stays, variable sampling rates, and gaps in monitoring data. After homogenizing signals to a 30-second sampling grid and extracting valid SBTs, three feature-engineering strategies were evaluated: handcrafted statistical descriptors, PCA-derived components, and latent features from convolutional Autoencoders and Variational Autoencoders. A deliberately compact Feed-Forward Neural Network served as classifier across all feature types, allowing the comparison to isolate the effect of representation learning. Models were trained on an artificially balanced dataset and evaluated on a chronologically independent, imbalanced test set using AUC-ROC, precision, accuracy, and threshold optimization. Results show that handcrafted features provide limited predictive value, whereas PCA and especially AE/VAE embeddings achieve higher ROC-AUC and improved generalization. These findings highlight the importance of unsupervised feature extraction for capturing nonlinear physiological dynamics. By relying on short pre-SBT physiological windows (often as brief as 15 minutes) making directly at the clinical point of extubation, offering a lightweight, readily deployable tool for ICU weaning assessment.