⚡ electrical engineering

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

This paper proposes a transfer learning framework using a BiLSTM model with self-attention to enable rapid, accurate, and low-power SpO2 estimation on wearable devices via 25Hz PPG signals, achieving high precision without the need for complex clinical calibration.

Original authors: Zequan Liang, Ruoyu Zhang, Wei Shao, krishna Karthik, Ehsan Kourkchi, Setareh Rafatirad, Houman Homayoun

Published 2026-07-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zequan Liang, Ruoyu Zhang, Wei Shao, krishna Karthik, Ehsan Kourkchi, Setareh Rafatirad, Houman Homayoun

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: Teaching a Watch to "Read" Your Blood Without a Doctor's Help

Imagine you want to build a smartwatch that can tell you how much oxygen is in your blood (SpO2). Usually, to get this right, you need a medical-grade machine that costs a lot of money, uses a lot of battery, and needs a doctor to calibrate it first.

The authors of this paper wanted to solve three problems:

Battery Drain: High-speed sensors eat up battery life.
Calibration Hassle: New devices usually need a doctor to "teach" them how to work.
Missing the Moment: Most devices are too slow to catch sudden drops in oxygen (like when someone stops breathing for a second).

Their solution? A "Transfer Learning" framework that acts like a student who learns from a textbook and then does a quick internship.

1. The "Textbook" Phase (Pre-training)

First, the researchers took a massive, high-quality dataset from a hospital (the "textbook"). This data was collected in a controlled environment with perfect equipment.

They trained a complex AI model (a BiLSTM with Self-Attention) on this data. Think of this model as a musical student who has spent years listening to perfect recordings of a symphony. They know exactly what a "perfect" heartbeat signal looks like.

The Trick: They taught this student to listen to the music at a slower tempo (25Hz) instead of a fast one (100Hz). Usually, slowing down music makes it sound muddy, but they found that for heart signals, the "slow tempo" still held all the important notes. This is like listening to a song at half-speed; you can still hear the melody, but it uses less data to store.

2. The "Internship" Phase (Fine-tuning)

Now, the student is ready to work in the real world. But real-world smartwatches (like their "We-Be band") are different from hospital machines. The skin is different, the wrist moves, and the sensors aren't perfect.

If they just used the "textbook" student, they would fail because the real world is messy. So, they used Transfer Learning.

The Analogy: Imagine sending that musical student to a local jazz club. The music is similar, but the acoustics and instruments are different. Instead of starting over, the student takes a short internship (fine-tuning) using just a few hours of recordings from the jazz club.
The Result: The student quickly adapts. They don't need a whole new education; they just tweak their ear to fit the new room. This allows the watch to work accurately without needing a doctor to calibrate it every time.

3. The "Energy Saving" Hack

The paper highlights a major win for battery life.

The Old Way: Most sensors take 100 snapshots of your blood flow every second. This is like taking a photo every second of a moving car; it uses a lot of memory and battery.
The New Way: Their watch only takes 25 snapshots per second.
The Metaphor: It's like switching from a high-definition 4K video to a smooth, efficient 1080p video. You still see the car moving clearly, but you save 40% of the battery power. This means the watch can last much longer on a single charge.

4. Catching the "Sudden Drops"

One of the coolest claims is that this system can catch instantaneous changes in oxygen.

The Problem: If you hold your breath or have a sleep apnea event, your oxygen drops fast. Old models are like a slow-moving camera; by the time they focus, the moment is gone.
The Solution: The AI model is fast enough to act like a high-speed shutter. It can see the oxygen level drop in real-time. In their tests, when the oxygen dropped suddenly, their model tracked it almost perfectly, whereas traditional methods missed the drop entirely.

Summary of Results

Accuracy: On public hospital data, their model was very accurate (error rate of about 3%). On their own wearable data, after the "internship," it got even better (error rate of about 2.6%).
Speed vs. Power: They proved you don't need a fast, power-hungry sensor to get good results.
No Doctor Needed: The biggest win is that the watch learns to work on your wrist using just a tiny bit of your own data, skipping the need for expensive clinical calibration.

In a nutshell: They built a smartwatch that learns from a medical textbook, does a quick internship on your wrist, saves 40% battery by slowing down its "heartbeat" camera, and is fast enough to catch sudden health scares in real-time.

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

Problem Statement

Blood oxygen saturation (SpO2) is a critical physiological marker for cardiovascular health, traditionally estimated via pulse oximetry using photoplethysmography (PPG) sensors. However, deploying accurate SpO2 estimation on energy-constrained wearable devices faces three primary challenges:

High Power Consumption: Existing machine learning approaches often require high PPG sampling rates (e.g., 100Hz), which drain battery life.
Clinical Calibration Dependency: Traditional calibration methods for new devices require controlled clinical environments and extensive data collection, hindering rapid deployment.
Poor Representation of Fluctuations: Clinical datasets often maintain stable saturation levels, failing to capture the spontaneous, rapid SpO2 fluctuations (e.g., during sleep apnea or respiratory distress) essential for real-world monitoring.

Methodology

The authors propose a transfer learning-based framework designed to adapt SpO2 estimation models to wearable devices using low-sampling-rate (25Hz) dual-channel PPG signals. The framework consists of three main stages:

1. Data Acquisition and Preprocessing

Public Dataset (Pretraining): The model is initially pretrained on the OpenOximetry Dataset, collected in a controlled clinical environment. Signals (Red 660nm and IR 940nm) recorded at 86Hz are downsampled to 25Hz via Fourier-domain resampling.
Private Dataset (Fine-tuning): A private dataset is collected using a custom "We-Be" wearable band (25Hz, Red 660nm/IR 950nm) alongside an FDA-approved Masimo Rad-G reference oximeter. To induce transient fluctuations, participants perform breath-holding cycles.
Signal Processing: Raw PPG signals undergo band-pass filtering (0.5–12Hz) to separate AC (pulsatile) and DC (baseline) components. The AC/DC ratio is computed for normalization. Data is segmented using a 5-second sliding window with a 1-second stride, resulting in input tensors of shape $(N, 125, 2)$ for $N$ segments.

2. Model Architecture

The core machine learning model comprises:

BiLSTM: A bidirectional Long Short-Term Memory network to capture temporal dependencies in the dual-channel PPG sequence.
Self-Attention Layer: Utilizes internal BiLSTM features as query, key, and value to capture global contextual relationships within the temporal representation.
Fully Connected (FC) Layer: Regresses the final SpO2 value.
Post-Processing: A moving average filter (window size 5) is applied to the sequence of predictions to smooth the output.

3. Transfer Learning Strategy

The authors employ a two-stage fine-tuning strategy to adapt the model from the clinical domain to the wearable domain:

Pretraining: All model layers are trained on the public clinical dataset.
Fine-tuning:
- First, only the FC layer is fine-tuned on the private wearable dataset.
- Second, the BiLSTM and FC layers are unfrozen and jointly fine-tuned on the private dataset.
  This approach aims to leverage the robust feature extraction learned from clinical data while adapting the model to the specific hardware and signal characteristics of the wearable device.

Key Contributions

Energy-Efficient Wearable Monitoring: The system operates at a 25Hz sampling rate, which the authors demonstrate reduces PPG power consumption by approximately 40% compared to standard 100Hz sampling, without significant loss in estimation accuracy.
Rapid Deployment via Transfer Learning: The framework enables fast model adaptation to new wearable devices using small amounts of user-level data, effectively eliminating the need for device-specific clinical calibration.
Instantaneous Fluctuation Capture: The model is capable of detecting rapid SpO2 changes, addressing the limitation of clinical datasets that often lack transient desaturation events.

Experimental Results

Public Dataset Performance

On the OpenOximetry dataset (downsampled to 25Hz), the proposed BiLSTM+Attn model achieved:

Mean Absolute Error (MAE): 2.967%
Root Mean Square Error (RMSE): 4.393%
This performance significantly outperformed traditional calibration methods (MAE 3.286%) and other deep learning baselines (CNN 1d, Transformer, standard BiLSTM). The results indicated that key SpO2 information is preserved even after downsampling from 86Hz to 25Hz.

Private Dataset Performance

Evaluated on the private wearable dataset using a leave-one-subject-out (LOSO) cross-validation strategy:

Overall Performance: The BiLSTM+Attn+Transfer model achieved the lowest MAE of 2.624% and RMSE of 3.214%, outperforming both the traditional method (MAE 4.069%) and the non-transferred BiLSTM+Attn model (MAE 2.717%).
Instantaneous Performance: In "instant zones" (defined as intervals where SpO2 variation exceeds 3% within 10 seconds), the transfer learning model achieved an MAE of 3.284%, demonstrating superior tracking of rapid desaturation events compared to the traditional method (MAE 4.738%).

Power Consumption

Measurements on the We-Be band showed that operating at 25Hz consumes approximately 9mW (excluding baseline draw), a 40% reduction compared to the ~15mW required at 100Hz.

Significance and Claims

The paper claims that this work demonstrates a viable path toward accurate, low-power SpO2 monitoring on wearable devices without the need for clinical calibration. By combining low-sampling-rate PPG with a transfer learning framework, the authors show that:

High sampling rates are not strictly necessary for accurate SpO2 estimation if appropriate temporal models (BiLSTM+Attention) are used.
Transfer learning effectively bridges the domain gap between controlled clinical data and real-world wearable data, allowing for rapid deployment.
The approach is particularly effective for applications requiring the detection of transient events, such as sleep apnea and respiratory distress, where capturing instantaneous fluctuations is critical.

The authors conclude that while the current method is robust, future work will focus on leveraging temporal dependencies across sliding windows and employing more advanced domain adaptation techniques.