SCG systolic detection in the wild: A static-dynamic cross-dataset analysis

The seismocardiographic signal represents a promising technique for continuous monitoring of cardiac electromechanical activity. The systolic complex is known to be the most informative part of the seismocardiogram and it requires accurate detection to perform further analysis. State-of-the-art solutions based on Deep Learning (DL) have been proven to be effective for systolic complex detection, but existing approaches limit their experimental analysis to controlled settings and consider a single dataset at a time. In our work, an exhaustive experimental analysis was conducted to specifically investigate the problem of SCG systolic detection in the wild, departing from the unrealistic assumptions of existing studies. Specifically, a DL model based on a U-Net was utilized to quantitatively assess the impact of domain shift in different scenarios, and to evaluate the challenges related to the SCG systolic detection in a dynamic real-world setting. For this purpose, a novel experimental framework was designed by combining multiple datasets in different ways. Our experimental framework enabled to perform a two-fold domain shift experimental analysis involving (i) cross-dataset, i.e. different acquisition setting, and (ii) static-dynamic evaluation analysis, considering signals collected during daily-life activities. Our results demonstrated significant differences between tests conducted in a controlled and a real-world environment, with the latter being considerably more challenging. Both cross-dataset and static-dynamic domain shifts resulted in worsening of the detection performance, which could be only partially overcome by traditional model adaptation strategies, such as fine-tuning or personalization. Finally, the benefits of a multi-channel approach in real-world scenarios were demonstrated.