DE-AGE Confounder Based Causal Representation Learning for Cuffless Blood Pressure Estimation

Noninvasive electrocardiogram (ECG) and photoplethy-smogram (PPG) have been extensively studied for cuffless blood pressure (BP) measurement, with the assumption that ECG and PPG signals would cause changes in BP or vice versa. However, in this prediction task, there is a category of features that simultaneously impact both the model's input and output, namely confounder. Neglecting confounder is one of the main reasons affecting the model's accuracy. This study endeavors to remove the most prevalent confounding factor, the bias caused by age, in the extraction of features. A novel method of de-Age confounder based causal representation learning for cuffless BP estimation (de-AgeBP) was proposed. It can potentially eliminate the interference induced by age on BP estimation. The de-AgeBP model is based on the artificial neural network and gated recurrent unit (ANN-GRU) with a causal intervention
$P$
(
$Y$
|do(
$X$
)), in which X and Y represent age and BP, respectively. The causal intervention can remove the confounding factor of age as mentioned above. We implement this intervention using representation learning to obtain features that remove the confounding factor. With causal representation learning, features related to BP changes can be extracted without the effect of age, and further applied to estimate BP. To validate the feasibility of the proposed model, data from 96 subjects (age range from 20 to 89) in the MIMIC III database was used. The result showed that the de-AgeBP model achieved an estimation accuracy of mean absolute error (MAE) being 10.68 mmHg and 7.62 mmHg for SBP and DBP, respectively, which outperformed the pure ANN-GRU model without causal intervention.