⚡ electrical engineering

Generalizable Blood Pressure Estimation from Multi-Wavelength PPG Using Curriculum-Adversarial Learning

This paper proposes a generalizable blood pressure estimation framework that combines curriculum learning and domain-adversarial training on multi-wavelength PPG data to achieve robust, subject-invariant performance with mean absolute errors of 14.2 mmHg for systolic and 6.4 mmHg for diastolic blood pressure under strict subject-level splitting.

Original authors: Zequan Liang, Ruoyu Zhang, Wei Shao, Mahdi Pirayesh Shirazi Nejad, 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, Mahdi Pirayesh Shirazi Nejad, 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

Imagine you want to teach a robot how to guess a person's blood pressure just by looking at a video of their finger pulse. This is what the paper is about. The researchers built a smart system that uses light to "see" the pulse and then uses math to guess the blood pressure numbers.

Here is the story of how they did it, broken down into simple parts:

1. The Problem: Cheating and Blind Spots

In the past, many scientists tried to build these blood-pressure guessers, but they had two big flaws:

The "Blind Spot": Most only used one color of light (like a single flashlight). The researchers realized that using four different colors of light (like a rainbow of flashlights) gives the robot much more information, just like seeing an object in color is better than seeing it in black and white.
The "Cheating" Problem: Many studies tested their robots by giving them data from the same people they had already trained on. It's like a student taking a test where they've already seen the answers. This made the robots look smarter than they really were. The researchers fixed this by strictly separating the "students" (training data) from the "test-takers" (test data) so no one could cheat.

2. The Solution: A Smart Training Camp

To make the robot truly smart and generalizable (able to guess well for anyone, not just the people it practiced on), they used a special two-step training method they call "Curriculum-Adversarial Learning." Think of this as a two-part training camp:

Part A: The "School Curriculum" (Curriculum Learning)
Instead of throwing the hardest math problems at the robot immediately, they taught it in stages:

First Grade: They taught the robot a simple job: "Is this person's blood pressure high (hypertension) or normal?" This is like learning to tell if a fruit is ripe or unripe.
Graduation: Once the robot got good at that, they slowly upgraded the task. They asked it to stop just guessing "High/Low" and start guessing the exact numbers (like 120 over 80).

Why? It's easier to learn the big picture first, then fill in the tiny details. This helps the robot understand the "shape" of the problem before trying to solve it perfectly.

Part B: The "Identity Thief" (Adversarial Training)
This is the cleverest part. The researchers added a second robot, a "detective," whose only job is to guess who the person is based on their pulse.

The main robot tries to guess the blood pressure.
The detective tries to guess the person's name.
The Trick: The main robot is punished if the detective can guess the name. So, the main robot learns to strip away all the unique "personality" traits of the pulse and focus only on the universal features that relate to blood pressure.
The Result: The robot learns to ignore who the person is and focus purely on the physics of the blood flow, making it much better at guessing for new, unknown people.

3. The Teamwork: Mixing the Lights

The system uses four different light channels (660nm, 730nm, 850nm, and 940nm).

Imagine each light color is a different musician in a band. One might be great at hearing the bass (diastolic pressure), while another is great at the drums (systolic pressure).
The researchers built a "conductor" (an attention mechanism) that listens to all four musicians and decides how much to listen to each one.
The Result: The full band (all four lights) played much better music (more accurate predictions) than any single musician playing alone.

4. The Scorecard

When they tested this system on a strict, no-cheating test:

Systolic (Top number): The robot was off by an average of 14.2 mmHg.
Diastolic (Bottom number): The robot was off by an average of 6.4 mmHg.
Comparison: This was better than other methods that didn't use the "School Curriculum" or the "Identity Thief" trick.

Summary

The paper claims that by using four colors of light, teaching the AI step-by-step (from simple to hard), and forcing it to forget who the person is (so it learns universal rules), they created a blood pressure estimator that is more accurate and fair than previous attempts. They proved that this method works better than older methods that used less light or let the AI cheat during training.

Technical Summary: Generalizable Blood Pressure Estimation from Multi-Wavelength PPG Using Curriculum-Adversarial Learning

Problem Statement
Accurate and generalizable blood pressure (BP) estimation is critical for managing cardiovascular diseases, particularly hypertension. While wearable photoplethysmography (PPG) offers a non-invasive, cuff-less alternative to conventional oscillometric methods, existing machine learning approaches face two significant limitations. First, many studies rely solely on single-wavelength PPG signals, neglecting the complementary information available in multi-wavelength data. Second, and more critically, many studies employ improper data splitting strategies, such as random segment-level splitting. This often results in data leakage where training and test sets share subject-specific information, leading to overestimated performance and poor generalization to new individuals.

Methodology
The authors propose a generalizable BP estimation framework utilizing a public four-wavelength PPG dataset (660nm, 730nm, 850nm, and 940nm). The core of the approach is a Curriculum-Adversarial Learning strategy built upon a multi-channel Convolutional Neural Network (CNN) architecture.

Data Preprocessing and Feature Extraction:
- The pipeline enforces strict subject-level data splitting (4:1 ratio for training/testing) to eliminate data leakage.
- PPG signals undergo band-pass filtering (0.5–8 Hz) and Z-score normalization.
- Features are extracted via first and second-order derivatives and upper envelopes to capture temporal dynamics and peak information.
- Each of the four PPG channels is processed through two parallel CNN branches: one for the normalized signal and derivatives, and another for the envelopes. This results in six input streams per channel.
Multi-Channel Fusion:
- Feature vectors extracted from the CNN branches are fused using an attention mechanism. This layer assigns soft weights to each channel, producing a weighted sum that serves as the input for downstream tasks.
Curriculum-Adversarial Learning:
The model is trained with three distinct heads and a composite loss function:
- Curriculum Learning: The model is guided from coarse-level to fine-grained tasks. Initially, a binary classifier predicts hypertension status (SBP ≥130 mmHg or DBP ≥90 mmHg) using Binary Cross-Entropy (BCE) loss. As training progresses, the focus shifts to BP regression (estimating SBP and DBP values) using Mean Squared Error (MSE) loss. An adaptive weight ( $\lambda_1$ ) transitions the loss function from classification to regression.
- Domain-Adversarial Training: To enhance generalization across subjects, a domain-adversarial discriminator attempts to identify the subject identity. A Gradient Reversal Layer (GRL) is employed during backpropagation to invert gradients, forcing the feature extractor to learn subject-invariant features that confuse the discriminator. This component uses Categorical Cross-Entropy (CE) loss.
- Total Loss: The objectives are combined as $L = \lambda_1 L_{reg} + (1 - \lambda_1) L_{cls} - \lambda_2 L_{adv}$ , where $\lambda_2$ controls the adversarial contribution.

Key Contributions

Strict Subject-Level Evaluation: The authors utilize a public four-wavelength dataset but enforce a strict subject-level split, addressing the data leakage issues prevalent in prior literature.
Multi-Wavelength Feature Fusion: The study demonstrates the efficacy of extracting and fusing complementary features across four distinct wavelengths (660nm, 730nm, 850nm, 940nm) rather than relying on single-channel inputs.
Curriculum-Adversarial Framework: The paper introduces a novel combination of curriculum learning (classification-to-regression transition) and domain-adversarial training to simultaneously capture clinical patterns and subject-invariant features.

Experimental Results
Experiments were conducted on a dataset of 180 subjects (66 with hypertension). The proposed method was compared against single-channel models and existing multi-wavelength baselines (A-BiLSTM, MLP, CNN1D, Multi-CNN).

Performance: Under strict subject-level splitting, the proposed method achieved a Mean Absolute Error (MAE) of 14.2 mmHg for Systolic BP (SBP) and 6.4 mmHg for Diastolic BP (DBP).
Multi-Channel Advantage: The full multi-channel fusion model consistently outperformed single-channel models. While the 660nm channel performed best for SBP and 940nm for DBP individually, the fused model achieved the lowest MAE and highest British Hypertension Society (BHS) grading scores (e.g., 57.6% of SBP predictions within 15 mmHg).
Comparison: The proposed approach outperformed all baselines, including the backbone Multi-CNN model, demonstrating the value of the curriculum and adversarial components.
Ablation Study: Removing either the classification head (curriculum) or the adversarial discriminator resulted in higher MAE, confirming that both components are essential for optimal performance.

Significance and Claims
The paper claims that its proposed framework offers a more realistic and robust assessment of BP estimation capabilities by eliminating data leakage through subject-level splitting. The authors highlight that while the stricter protocol results in higher reported errors compared to some previous studies (which may have suffered from leakage), the model demonstrates strong potential for hypertension screening and long-term trend monitoring.

The authors remain modest regarding the current accuracy, noting that performance still has room for improvement, particularly under extreme blood pressure ranges and degraded signal quality. They conclude that leveraging complementary multi-wavelength information combined with curriculum-adversarial strategies provides a viable path toward accurate and generalizable cuff-less BP estimation. Future work is planned to include additional data collection and signal quality assessment enhancements.