Single-Beat Cuffless Blood Pressure Estimation Using Ear-PPG and ECG with a Lightweight Hybrid Learning Framework

Continuous, cuffless blood pressure (BP) monitoring is often hindered by motion artifacts, the need for long temporal windows to stabilize signal processing, and the computational intensity of deep learning models. This paper investigates whether BP-related information can be reliably extracted at the single-beat level to enable efficient, robust, and real-time monitoring on wearable devices.

The authors propose a hybrid learning architecture that combines two distinct data streams. First, a one-dimensional convolutional neural network (CNN) processes individual PPG beats to extract 64-dimensional morphological embeddings. Second, 30 physiology-grounded features—including pulse transit time (PTT) statistics and heart rate variability (HRV)—are calculated from synchronized chest ECG and ear-clip PPG signals. These two inputs are fused and fed into a LightGBM regressor to estimate systolic and diastolic blood pressure. The system was evaluated using a multi-phase stress protocol (rest, exercise, recovery, and cold pressor stress) on 10 healthy volunteers and validated against the PulseDB public dataset.

The proposed framework achieved a mean absolute error (MAE) of 4.02 mmHg for systolic BP and 1.79 mmHg for diastolic BP across 30 independent subject-disjoint validation runs. This represents a 28.2% reduction in combined MAE compared to baseline models. By focusing on single-beat estimation, the system avoids the need for long temporal context, making it highly resilient to intermittent signal corruption and suitable for low-power wearable deployment.

This work demonstrates that high-fidelity BP estimation does not require long, stable recording windows, which are often impossible to maintain during daily activities. By integrating motion-resilient ear-clip sensors with a computationally efficient hybrid model, this approach provides a viable path toward continuous, non-invasive hypertension management that can operate under the strict resource constraints of modern wearables.