Prospects of Cuffless Pulse Pressure Estimation from a Chest-Worn Accelerometer Contact Microphone

This study explores the correlation between pulse pressure (PP) and heart sounds on the chest wall. The proposed framework leverages a sensitive accelerometer contact microphone (ACM) to record chest vibrations. A discrete wavelet transform (DWT) decomposes the chest vibration recordings into a set of sub-bands, from which several time-domain features are extracted. An extreme gradient boosting (XGBoost) regressor is trained on the feature space for PP estimation, and the estimated values are compared with PP readings from a standard cuff-based blood pressure monitor. The performance of the model is evaluated on 20 patients with cardiovascular diseases (CVDs). Average root mean square error (RMSE), mean absolute error (MAE), and accuracy of 11.41 (pm 6.42) mmHg, 10.49 (pm 6.73) mmHg, and 77.14% (pm 19.09%) are achieved, respectively, for a leave-subject-out validation. Additionally, the performance of the model is assessed through a 10-fold cross validation where an average accuracy of 95.68% is obtained, implying high consistency with ground-truth values. The most significant signal sub-bands for PP estimation are found to be high-frequency bands such as 1-2 kHz and 512-1,024 Hz, as well as medium-frequency bands of 32-64 Hz and 64-128 Hz. It is also demonstrated that the most contributive sub-band to PP estimation is 1-2 kHz.