⚡ electrical engineering

Synergistic Blood Pressure Estimation via Contactless mmWave Radar and Imaging Photoplethysmography: A Feasibility Study

This feasibility study demonstrates that a dual-modality system combining contactless facial imaging photoplethysmography (iPPG) and mmWave radar, processed by a novel BiLSTM-MS-DiCNN deep learning architecture, achieves accurate, non-invasive blood pressure estimation with a mean absolute difference of approximately 4.6 mmHg across various physiological states.

Original authors: Boyuan Gu, Haiyang Sun, Yongjie Liu, Guangji Ma, Guanghao Sun, Xiaorong Ding

Published 2026-06-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Boyuan Gu, Haiyang Sun, Yongjie Liu, Guangji Ma, Guanghao Sun, Xiaorong Ding

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 trying to measure the pressure in a garden hose without touching it. If you only look at the water splashing out the end (the "distal" signal), you might get a good idea, but if the sun is too bright or the water is murky, your guess could be way off. If you only listen to the pump at the start of the hose (the "proximal" signal), you know when the water starts, but you don't know how hard it's pushing by the time it reaches the end.

This paper presents a new way to measure blood pressure without touching the person at all. Instead of relying on just one method, the researchers combined two different "senses" to get a much clearer picture.

The Two "Senses"

The researchers built a system that uses two cameras/sensors working together:

The "Face Camera" (iPPG): This looks at the person's face. It uses a standard video camera to detect tiny changes in skin color caused by blood pumping through the face. Think of this as watching the water splashing at the end of the hose to see how fast it's flowing.
The "Back Radar" (mmWave): This is a radar sensor placed behind the person's back. It can detect incredibly tiny vibrations in the chest caused by the heart beating. Think of this as listening to the pump at the start of the hose to see exactly when the water is pushed out.

By combining the "when" from the back radar and the "what" from the face camera, the system can calculate the time it takes for the blood pulse to travel from the heart to the face. In physics, this travel time is a key clue for figuring out blood pressure.

The "Brain" of the System

The tricky part is that the signal from the face (a video) and the signal from the back (a radar wave) look very different. It's like trying to mix a painting with a song; they don't naturally fit together.

To solve this, the researchers created a special computer brain (an AI called BiLSTM-MS-DiCNN).

The "Time-Traveler" (BiLSTM): This part of the brain looks at the signals over time, remembering what happened a second ago to understand what's happening now. It's like a conductor keeping the rhythm of the music.
The "Zoom Lens" (MS-DiCNN): This part looks at the details. It can zoom in on tiny blips in the signal and zoom out to see the big wave shape. It helps the system understand the unique shape of the heartbeat, whether it's a calm beat or a fast one after running.

The Experiment: Testing in the Wild (Sort Of)

The team tested this system on 15 healthy young volunteers in a quiet lab. They didn't just sit still; they put the volunteers through three different scenarios to see if the system could keep up:

Sitting Still: The baseline test.
Deep Breathing: Breathing slowly and deeply to change the pressure in the chest.
After Running: A quick jog to make the heart beat fast and the blood pressure change rapidly.

They also tested how well the system handled "bad weather" (metaphorically). They simulated:

Bad Lighting: Making the video too bright (overexposed) or too dark (underexposed).
Static Noise: Adding "snow" to the video like an old TV.
Compression: Making the video low-quality, like a shaky video call.

What They Found

The results were promising, especially when compared to using just one sensor:

The Power of Two: When the system used both the face camera and the back radar, it was much more accurate than using just the camera or just the radar.
- Analogy: If you try to guess the temperature by only looking at a thermometer, you might be right. But if you look at the thermometer and feel the air with your hand, you get a much better, more reliable answer.
Handling Bad Signals: When the researchers made the face video very poor (too dark or noisy), the system using only the camera failed miserably. However, the system using both sensors kept working well. The radar acted like a safety net, catching the information the camera missed.
Accuracy: Under normal conditions, the system was off by an average of about 4.7 mmHg for systolic (top number) and 4.6 mmHg for diastolic (bottom number). This is considered very good for a non-contact method.
No Calibration Needed (Sort of): Usually, these machines need to be "calibrated" with a traditional cuff on the arm first. This system could still work without any calibration, though it got slightly better if it was allowed to "learn" the person's baseline for just a few seconds.

The Bottom Line

This paper proves that it is possible to measure blood pressure without touching the skin by combining a face camera and a back radar. The "brain" they built successfully learned how to mix these two very different signals to track blood pressure changes, even when the person was moving, breathing deeply, or when the lighting was bad.

The study is a "feasibility study," meaning it shows the idea works in a controlled lab with healthy people. It doesn't claim this is ready for hospitals yet, but it proves the concept is solid and could be a big step toward painless, continuous blood pressure monitoring in the future.

Technical Summary: Synergistic Blood Pressure Estimation via Contactless mmWave Radar and Imaging Photoplethysmography

1. Problem Statement

Continuous, non-contact blood pressure (NCBP) monitoring is critical for managing cardiovascular diseases, yet existing single-modality approaches face inherent physical limitations. Imaging Photoplethysmography (iPPG), which captures distal optical volumetric fluctuations, is highly susceptible to environmental artifacts (e.g., unstable illumination) and skin-tone sensitivity. Conversely, Frequency-Modulated Continuous Wave (FMCW) millimeter-wave (mmWave) radar, which captures proximal cardiac micro-motions (Radar Motion Signals, RMS), lacks the peripheral arterial visibility required to independently compute transit-based blood pressure surrogates. Furthermore, existing dual-modality frameworks often rely on rudimentary feature-level fusion (e.g., scalar Pulse Transit Time calculation via peak detection), which is vulnerable to phase misalignment and transient motion artifacts. There is a lack of systematic investigation into deep learning paradigms capable of fusing these heterogeneous signals to overcome morphological disparities and environmental constraints.

2. Methodology

2.1 Physiological Basis and Sensing Paradigm

The study proposes a synergistic dual-modality sensing paradigm that integrates facial iPPG with posterior-facing mmWave radar.

Physiological Model: The approach is grounded in the Moens-Korteweg and Hughes equations, which relate Blood Pressure (BP) to Pulse Wave Velocity (PWV) and Pulse Transit Time (PTT). The framework aims to capture a non-contact PTT (ncPTT) surrogate by aligning the proximal cardiac ejection event (detected via RMS) with the distal pulse arrival (detected via iPPG).
Hardware Configuration:
- iPPG: A high-precision RGB camera captures facial video (30 Hz) from 50 cm in front of the subject.
- mmWave Radar: A 24 GHz FMCW radar is placed 50 cm behind the subject, facing the posterior torso. This "back-facing" configuration leverages the subject's contact with a chair to suppress macroscopic body sway, isolating subtle cardiomechanical vibrations.
- Ground Truth: A clinical-grade Continuous Non-Invasive Arterial Pressure (CNAP) system provides beat-to-beat SBP and DBP references.

2.2 Signal Processing

iPPG Extraction: The forehead Region of Interest (ROI) is processed using the CHROM color space combination. The signal undergoes zero-phase filtering and Successive Variational Mode Decomposition (SVMD) to mitigate illumination artifacts and stabilize the volumetric baseline.
RMS Extraction: The radar's complex I/Q stream is converted to an unwrapped phase trajectory. A zero-phase 4th-order Butterworth low-pass filter (4 Hz cutoff) and SVMD are applied to isolate the cardiomechanical component (0.8–2 Hz), effectively reconstructing the initial systolic ejection event.
Synchronization: Both streams are downsampled to 30 Hz and temporally aligned. A 5-second buffer is truncated from signal extremities to avoid edge effects during filtering.

2.3 Deep Learning Architecture: BiLSTM-MS-DiCNN

To address the morphological disparity between 1D mechanical radar waves and high-dimensional optical fluctuations, the authors propose an end-to-end deep learning network:

Input: Synchronized dual-channel sequences (iPPG and RMS) windowed into 120 time steps (4 seconds).
BiLSTM Branch: Models long-range temporal dependencies and hemodynamic memory by processing sequences in both forward and backward directions, followed by Global Average Pooling.
MS-DiCNN Branch: Addresses spatial morphological disparities using Multi-Scale Dilated Convolutions (dilation rates $d=1, 2, 4$ ) to capture features at varying scales without resolution loss. A Squeeze-and-Excitation (SE) attention mechanism recalibrates channel responses to focus on physiologically relevant features.
Adaptive Fusion & Subject Adaptation Module (SAM): The latent vectors from both branches are fused adaptively. The SAM utilizes a brief initial calibration phase (first 10% of a new subject's data) to fine-tune the network to individual arterial compliance and baseline vascular tones, addressing inter-subject variability.

3. Experimental Setup

Participants: 15 healthy volunteers (9 male, 6 female; age 22 ± 3 years).
Protocols: Data was collected under three distinct hemodynamic states to test robustness:
1. Sitting (Sit): Normal breathing (baseline).
2. Deep Breathing (DB): Induced Respiratory Sinus Arrhythmia (RSA) and intrathoracic pressure variations.
3. Post-Exercise (Ex): High-frequency cardiac peaks and disordered breathing following a 2-minute run.
Validation Strategy: A Leave-One-Subject-Out (LOSO) cross-validation protocol was enforced to prevent data leakage. Synthetic degradations (Gaussian noise, over/underexposure, H.264 compression) were applied to the iPPG channel to evaluate robustness.

4. Key Results

4.1 Performance Under Standard Conditions

Under resting conditions, the dual-modality framework achieved:

Systolic BP (SBP): Mean Absolute Difference (MAD) of 4.71 mmHg, Root Mean Square Difference (RMSD) of 6.13 mmHg, and $R^2$ of 0.85.
Diastolic BP (DBP): MAD of 4.60 mmHg, RMSD of 5.76 mmHg, and $R^2$ of 0.87.
These results fall within commonly cited AAMI-style error bounds (5 ± 8 mmHg). The model maintained consistent performance during physiological perturbations (Deep Breathing: SBP MAD 6.35 mmHg; Post-Exercise: SBP MAD 5.33 mmHg).

4.2 Comparison with Single-Modality and Baselines

Dual vs. Single Modality: The dual-modality approach significantly outperformed iPPG-only and RMS-only baselines. Single-modality models exhibited higher error dispersion (SDE > 9 mmHg for RMS-only), whereas the fused model suppressed artifacts and improved stability.
Model Comparison: The proposed BiLSTM-MS-DiCNN outperformed traditional regressors (XGBoost, SVR, GPR) and a dual-stream Feedforward Neural Network (FNN) in terms of accuracy and computational efficiency (Model size < 1 MB; inference time ~30s on GPU).

4.3 Robustness to Signal Degradation

When iPPG signals were degraded by Gaussian noise, overexposure, or underexposure:

The iPPG-only model suffered significant performance drops (e.g., SBP MAD increased from 5.35 mmHg to 9.98 mmHg under underexposure).
The Dual-Modality model maintained low error margins (SBP MAD ~5.3–5.4 mmHg) and high correlation ( $R^2 > 0.7$ ), demonstrating that the mmWave radar effectively compensates for optical signal degradation.

4.4 Ablation and Calibration Analysis

Ablation: Removing any core component (BiLSTM, MS-DiCNN, dilation, or fusion) resulted in drastic performance degradation (e.g., removing MS-DiCNN increased SBP MAD to 12.54 mmHg), confirming the necessity of each module.
Calibration: The model retains a measurable baseline capability even without calibration (0% calibration ratio: SBP MAD 7.52 mmHg). Performance improves with brief calibration (10% ratio), reducing SBP MAD to 4.71 mmHg, indicating that SAM enhances personalization but is not strictly required for the model to function.

4.5 Physiological Interpretability

A linear probe analysis confirmed that the network's internal fused features encode non-contact PTT information, achieving a Pearson correlation of $r = 0.726$ ( $p < 10^{-30}$ ) with ground-truth ncPTT. This suggests the model implicitly learns the physiologically relevant timing relationship between proximal ejection and distal arrival.

5. Significance and Claims

The paper claims that this study provides a feasibility demonstration of a synergistic dual-modality sensing paradigm for NCBP monitoring. Key contributions include:

Feasibility of Face-Back Fusion: Validating that combining facial iPPG with posterior-facing mmWave radar effectively mitigates single-modality environmental sensitivities while providing complementary boundary conditions of the cardiovascular tree.
Architectural Innovation: The development of the BiLSTM-MS-DiCNN network successfully addresses the morphological disparity between heterogeneous signals, enabling end-to-end mapping without fragile intermediate feature extraction.
Robustness: The framework demonstrates that mmWave radar can act as a compensatory modality when optical signals are compromised, a critical requirement for real-world deployment.
Physiological Alignment: The model's ability to implicitly encode PTT information supports the validity of the deep learning approach as a physiologically grounded method rather than a "black box."

The authors emphasize that while the results are promising and competitive with state-of-the-art non-contact methods, the study is limited to a controlled cohort of 15 young, healthy participants. Future work is identified as necessary to include diverse demographics, pathological cases, and clinical trials with certified medical devices.