Bayesian Model Averaging for Improving the Accuracy of Cuffless Blood Pressure Estimation.

In recent decades, many researches have proposed various models for continuous, cuffless blood pressure (BP) estimation. However, due to aleatoric uncertainty and epistemic uncertainty existing in the problem, it is very challenging to evaluate cuffless BP with acceptable accuracy. This paper innovatively proposes a cuffless BP ensemble estimation model based on Bayesian Model Average (BMA) method to reduce the epistemic uncertainty. We combine four most frequently cited physiological models and four regression models based on Photoplethysmogram (PPG) and Electrocardiogram (ECG) signals, and use the BMA method to assign weights to each model to achieve accurate cuffless BP prediction. The proposed method was validated on 17 healthy and 13 hypertensive subjects with continuous Finometer BP as a reference. The results showed that the error mean ± SD (standard deviations) of both SBP and DBP predicted by the proposed method were 2.13 ± 5.68 mmHg and 1.42 ± 5.11 mmHg, respectively, which were both lower than each of the model. And the MAE was 6% and 8% lower than the best member of the model ensemble. We also analyzed the relationship between the number of training epochs and model prediction performance. When 15 cardiac cycles were choosed for training, it could get a good balance between model prediction accuracy and algorithm complexity. Therefore, the proposed BMA method can solve the model uncertainty problem, providing robust and deterministic BP prediction. Clinical relevance- This paper proposes a new method for wearable BP estimation which enables BP monitoring in both clinical settings and home settings. It offers a stable way to monitor BP to help patients detect disease early.