Rapid Adaptation of SpO2 Estimation to Wearable Devices via Transfer Learning on Low-Sampling-Rate PPG

TL;DR

The paper tackles the challenge of accurate SpO2 estimation on energy-constrained wearables by eliminating clinical calibration and using low-sampling-rate dual-channel PPG. It introduces a transfer-learning pipeline that pretrains a BiLSTM with self-attention on a public clinical dataset and fine-tunes it on a small private wearable dataset at 25 Hz, enabling rapid device adaptation. The approach yields MAEs of 2.967% on public data and 2.624% on private data, with robust instantaneous SpO2 tracking (MAE_ins = 3.284%), and achieves about 40% power savings over 100 Hz operation. This work demonstrates practical, low-power SpO2 monitoring for wearables, suitable for long-term monitoring and rapid deployment without device-specific clinical calibration.

Abstract

Blood oxygen saturation (SpO2) is a vital marker for healthcare monitoring. Traditional SpO2 estimation methods often rely on complex clinical calibration, making them unsuitable for low-power, wearable applications. In this paper, we propose a transfer learning-based framework for the rapid adaptation of SpO2 estimation to energy-efficient wearable devices using low-sampling-rate (25Hz) dual-channel photoplethysmography (PPG). We first pretrain a bidirectional Long Short-Term Memory (BiLSTM) model with self-attention on a public clinical dataset, then fine-tune it using data collected from our wearable We-Be band and an FDA-approved reference pulse oximeter. Experimental results show that our approach achieves a mean absolute error (MAE) of 2.967% on the public dataset and 2.624% on the private dataset, significantly outperforming traditional calibration and non-transferred machine learning baselines. Moreover, using 25Hz PPG reduces power consumption by 40% compared to 100Hz, excluding baseline draw. Our method also attains an MAE of 3.284% in instantaneous SpO2 prediction, effectively capturing rapid fluctuations. These results demonstrate the rapid adaptation of accurate, low-power SpO2 monitoring on wearable devices without the need for clinical calibration.

Figures (7)

• Figure 1: SpO$_2$ estimation framework using transfer learning
• Figure 2: Private wearable data collection. (a) We-Be band for PPG signals (b) Fingertip sensor attached to Masimo Rad-G (c) Masimo Rad-G for reference SpO$_2$
• Figure 3: Data processing (a) PPG preprocessing (b) Machine learning method
• Figure 4: SpO$_2$ prediction results on a representative test case from public OpenOximetry Dataset using 25Hz PPG
• Figure 5: Power consumption of the PPG sensor on the We-Be band