This bachelor's thesis addresses the problem of automatic detection of premature contractions from photoplethysmographic (PPG) signals using machine learning methods. The theoretical part describes cardiac activity, the characteristics of atrial and ventricular extrasystoles, and the diagnostically significant features of the PPG signal. Annotated data from the MIMIC database and a synthetic dataset derived from the CPSC2018 database via a PPG simulator were used for this purpose.
The practical part covers the design and implementation of a systolic peak detector, a morphological filter for spurious extrasystolic waves, and a pipeline for onset-to-onset segmentation and feature extraction. From an initial set of 23 features, 13 were selected for modelling based on visual separability analysis and correlation checks. The classifier was trained exclusively on the synthetic dataset and evaluated on real patient data, with splits performed strictly at the patient level to prevent data leakage.
The classification task was formulated as binary detection of premature contractions (ES class versus normal rhythm). Three machine learning algorithms were compared: Random Forest, XGBoost, and Support Vector Machine with an RBF kernel. Optimal feature subsets were identified by ablation study and hyperparameters were tuned using stratified group cross-validation. The final model, a tuned Random Forest using six temporal and amplitude ratio features, achieves an F1-score of 0.798, sensitivity of 0.807, and specificity of 0.994 on the test set. As a post-hoc analysis, model performance was evaluated separately for atrial and ventricular extrasystoles. Sensitivity reached 83.9 % for atrial and 58.9 % for ventricular extrasystoles, with the lower detection rate for ventricular extrasystoles attributable to insufficient morphological coverage of this arrhythmia type in the synthetic training dataset.