An AutoEncoder-Based Semi-Supervised Learning for Efficient Blood Pressure Estimation with Minimal Labeled Data

Using photoplethysmography (PPG) signals to estimate blood pressure (BP) is a feasible approach, easily monitored via wearables, unlike traditional BP measurement methods that require specialized equipment. This paper proposes a semi-supervised learning method using AutoEncoder to predict BP with minimal labeled data. Based on the publicly available VitalDB dataset with 1844 patient data, we first apply a multi-head self-attention AutoEncoder (MHSA-AE) model to extract high-dimensional features from PPG signals and their derivatives without BP. Next, by applying the method of extracting representative pulses, the amount of labeled data is reduced to approximately 13.56%-51.90% of the original labeled data. Then, semi-supervised learning trains the model on the differences in the limited labeled data to generate pseudo-labels, minimizing the reliance on large labeled data. Finally, personalized prediction is achieved by fine-tuning the last layer of a 5-layer multi-layer perceptron (MLP). Experimental results show that using the labeled data filtering and difference processing method (LDF-DP), along with pseudo-labeling, significantly reduces the amount of labeled data and effectively lowers the mean absolute error (MAE) and standard deviation (STD) of systolic blood pressure (SBP) and diastolic blood pressure (DBP) on the test set. In general, the smaller the amount of labeled data, the greater the gain brought by pseudo-labels. When there is a sufficient amount of labeled data, pseudo-labeling offers little improvement. Finally, fine-tuning through transfer learning provides a noticeable gain in BP prediction.