Detekce významných bodů v balistokardiografické křivce pomocí klasických nebo pokročilých metod zpracování signálů a metod strojového učení.

Abstract

This master’s thesis deals with the design and implementation of a system for detecting significant points in the ballistocardiographic signal. Ballistocardiography represents a promising non-invasive alternative to traditional ECG, which is particularly suitable for examinations in a magnetic resonance environment, where standard ECG faces limits associated with strong electromagnetic interference. However, processing the BCG signal is analytically challenging due to its nonlinearity and high morphological variability, which depends on the patient’s weight distribution and position. The thesis presents a methodology including an algorithm for the automatic selection of the highest quality measurement channel and the subsequent selection of highly reliable recording segments using synchronization with a reference ECG and a pulse oximeter. The continuous signal is subsequently transformed into training datasets through the method of overlapping sliding windows. The extracted features were reduced to 15 key parameters supplemented by a broader temporal context using ablation analysis. Machine learning methods were deployed for the final classification and detection of heartbeats, with the Extra Trees and Random Forest models achieving the best results. The extreme data imbalance was successfully addressed using majority class reduction and SMOTE methods. The resulting optimized system is flexible and allows the operator to choose a strategy exactly according to clinical requirements – either prioritizing maximum sensitivity (97.6%) or high precision with a minimum of false detections (F1 score 0.722).