Evaluation of Clustering and Neural Network Approaches for Morphological Analysis in Impedance Cardiography

Impedance cardiography (ICG) is a non-invasive technique used to estimate hemodynamic parameters by analyzing changes in thoracic impedance. To obtain these values – as stroke volume or cardiac output - characteristic points of the ICG waveform must be obtained. Although many studies assume the existence of a single, idealized ICG morphology with fixed distribution of characteristic points, recent research suggests the presence of multiple waveform subtypes. Some of these subtypes may exhibit different distributions or even lack several characteristic points, raising concerns about the accuracy and generalizability of traditional analysis methods. This project explores whether ICG signals exhibit a variety of coexisting morphologies or are dominated by a single representative pattern. To investigate this, signals from two independent databases have been used: ReBeatICG and HeartCycle. ICG and electrocardiography (ECG) signals from different healthy subjects have been processed and ICG beat segments have been obtained. Then, a Random Forest classifier has been trained to automatically discard noisy or invalid segments. Several unsupervised clustering algorithms are applied to the subdataset of valid segments, using dynamic time warping (DTW) as a similarity metric, to group morphologically similar segments. To assess the consistency of the resulting clusters with existing morphological classifications, a supervised neural network has been implemented. This model serves as a reference for patterns previously reported in the literature. As a result, a comparative analysis of ICG waveform subtypes across clusters is conducted by applying the artificial neural networks trained to the clustering results. In the HeartCycle database, multiple ICG subtypes are observed within each cluster, indicating that the clustering algorithm has grouped segments based on criteria different from those established in the literature. In contrast, the ReBeatICG dataset reveals the presence of three predominant ICG subtypes, suggesting a clustering outcome consistent with previously reported classifications.