Determining the Respiratory State From a Seismocardiographic Signal - A Machine Learning Approach

Seismocardiography (SCG) is a non-invasive method for measurement of vibrations on the chest wall originating from the heart. Respiration changes the morphology of the SCG-signal and analyzing these changes could improve the diagnostic value of SCG. This study aimed to determine the nasal respiration signal amplitude at mitral closure (MC) and aortic opening (AO) using SCG features. The three proposed methods for this were multiple regression analysis (MRA), support vector regression (SVR), and a neural network (NN). SCG, Electrocardiography and nasal-catheter flow signals were acquired from 18 healthy subjects (age $29\pm 6$). SCG-signal fiducial points were used as features and were found using an automatic algorithm followed by manual verification. Fiducial points amplitudes, timings between these and frequency components formed 12 features. All models were trained on 80% of the data, underwent 10-fold cross-validation and were tested on the remaining 20% of the data. Predictions on test data for MC and AO time points, the Pearson correlations coefficient, and sum of squared errors of prediction were: $(r_{MC},\ r_{AO},\ SSE_{MC},\ SSE_{AO})$ for the following models: NN (0.908, 0.904, 11.71, 12.05), SVR (0.881, 0.833, 18.95, 19.76) and MRA (0.450, 0.437, 51.21, 51.48). These predictive models show a strong correlation between the SCG-signal and respiration.