Cascade forest regression algorithm for non-invasive blood pressure estimation using PPG signals

Abstract

The regression and classification performances of deep-learning algorithms depend on the tuning of the model and require significant intervention to train the model. A blood pressure prediction (BP) method – which is an improved cascade forest regression method based on a deep forest framework – is proposed in this study to mitigate the adverse effects of hyperparameters on deep learning algorithms and human errors. The impact of hyperparameters on the model can be reduced using random forest regression and extra tree regression as base estimators and by applying them to bootstrap aggregation strategies for learning data. The photoplethysmography (PPG) data and ambulatory blood pressure obtained from the MIMIC II database were evenly divided into 5 s segments to accurately estimate blood pressure. Nonlinear indices of the time domain, frequency domain, and heart rate variability were obtained from the PPG signal as training characteristic values. This process is similar to collecting data directly from a wearable device for rapid blood pressure prediction. This improves the predictive performance of the model and reduces additional memory consumption. Our proposed algorithm achieved mean absolute errors of 1.760 and 2.896 mmHg for systolic blood pressure (SBP) and diastolic blood pressure (DBP), respectively. Additionally, our best model achieved R2 scores of 0.948 and 0.926 for SBP and DBP, respectively. According to the Association for the Advancement of Medical Instrumentation standards, the standard deviation and mean error of the predicted results for systolic and diastolic BPs were within the standard range. According to the British Hypertension Society standard, the results of the proposed algorithm reached grade A. This study confirmed the possibility of developing an algorithm that can accurately estimate blood pressure.