Ivelin Karageorgiev

Hysteresis defines a system’s state as a function of its historical trajectory, manifesting as path-dependent loops where variables diverge based on the direction of change. Intrinsic to both information storage and irreversible thermodynamic dissipation, this phenomenon serves as a critical fingerprint for characterizing complex behaviors and dynamic system responses. The automated classification and labeling of dense hysteresis signals is essential for systems engineering, electronics, bio-signal processing, and meteorology. This study presents a comprehensive comparison between two unsupervised learning pipelines, to which a main objective is to discover the capabilities for phenotype discovery between automated feature extraction and expert-driven feature engineering. These methodologies are examined by clustering 7736 respiratory cycles from critical care patients who have undergone invasive mechanical ventilation. Methodology I utilizes the extraction of 17 geometric and spectral features, reduced via Principal Component Analysis (PCA) to ease the choice of high variance parameters for subsequent clustering. Methodology II proposes a Deep Learning (DL) framework utilizing a 1D-convolutional autoencoder (CAE) for automated feature discovery. Both methods utilize UMAP and HDBSCAN for the final clustering stage. After evaluation both models performed with high trustworthiness (0.9974; 0.9853) and Silhouette Scores of 0.6441 and 0.6415. Here, both methodologies obtained a surprisingly low noise ratio (1.5% and 0.74%) by the aformentioned HDBSCAN. The strategy of narrowing discrete features allows for better model interpretability, and the resulting similarity measures reasoned trough a Davies–Bouldin Index are higher. However, a peculiar difference was observed in favor to the automated extraction (CAE), where the method assessed a better quality of clustering when compared for dispersion between clusters and the dispersion of reference clusters.