A Robust Neural Network-Based Method to Estimate Arterial Blood Pressure Using Photoplethysmography.

High Blood Pressure can lead to various cardiovascular diseases increasing the risk of death. Photoplethysmography (PPG) can be used as a low cost, optical technique to determine the arterial blood pressure continuously and noninvasively. Features of several different categories can be extracted from PPG signals. The prominent ones include width-based features, frequency domain features and features extracted from the second derivative of the signal (accelerated PPG). Existing methods primarily use one category of features or another but do not use features from multiple categories. We propose a method to extract a combination of characteristics from the PPG signal, which spans across the aforementioned categories and use them to train a neural network in order to estimate the Blood pressure values. Furthermore, most existing methods are not evaluated on PPG signals collected in a nonclinical setting using consumer-grade/wearable devices, which leaves their applicability to such settings untested. We evaluate our method using a benchmark dataset (MIMIC II) collected in a clinical setting as well a dataset collected using a consumer-grade device in a nonclinical setting. The results show that our method using 53 features achieves Mean Absolute Errors of 4.8 mmHg & 2.5 mmHg for Systolic Blood Pressure and Diastolic Blood Pressure respectively while reaching grade A for both the estimates under the standard British Hypertension Society for the MIMIC II dataset. The same methodology applied to the second dataset shows good agreement (MAE 4.1, 1.7 mmHg for SBP and DBP respectively) with readings taken using a standard oscillometric device, which suggests the robustness of our approach.