Seismocardiogram Classification Based on Symmetric Projection Attractor Reconstruction Analysis to Enhance Hemodynamic Variables Estimation

Hemodynamic variables derived from electrocardiography (ECG) and seismocardiography (SCG) provide complementary information for assessing and monitoring cardiovascular health. ECG measures the electrical activity of the heart, while SCG reflects the mechanical response to this activity. Despite its potential, SCG is particularly prone to noise, with signals often affected by artifacts caused by motion, respiration, and external vibrations. These challenges make it difficult to extract reliable hemodynamic variables and to compute the associated cardiac time intervals (CTIs), emphasizing the need for a signal quality index to support automated analysis. Symmetric Projection Attractor Reconstruction (SPAR) analysis offers a promising approach for evaluating signal quality by transforming signals into representations that capture waveform morphology and underlying dynamics. This study introduces an innovative methodology that combines SPAR, machine learning, and transfer learning into a meta-model able to classify SCG signals into three quality levels – good, medium, and low – thus establishing a robust assessment framework. By effectively filtering out unreliable data and preserving high-fidelity signals, this approach enhances the accuracy and reliability of SCG-based cardiovascular monitoring, leveraging key factors affecting SPAR projections and signal quality, as verified through tests performed on a dataset collected from healthy volunteers through a medical device. The results show that the proposed meta-model achieves 91.5% accuracy in high (good and medium) vs low quality classification and 86.9% in good vs medium classification when using SPAR images with a Tau Factor (TF) set at six. Our meta-model also demonstrates robust scalability, maintaining adequate accuracy on novel data, and providing reliable results on long-term SCG recordings. This framework highlights the potential of combining SPAR analysis and machine learning to enhance SCG signal quality classification, facilitating automated cardiovascular health monitoring.