Remote Blood Pressure Estimation from Facial Videos using Transfer Learning: Leveraging PPG to rPPG Conversion

Chun-Hong Cheng, Jing Wei Chin, Kwan Long Wong, Tsz Tai Chan, Hau Ching Lo, Kwan Lok Pang, Richard So, Bryan YanWACV 2025

TL;DR

This paper tackles the challenge of remote blood pressure (BP) estimation from facial videos, a crucial but difficult task due to limited labeled data. They propose a transfer learning approach, leveraging knowledge from readily available photoplethysmography (PPG) signals to improve the accuracy of remote photoplethysmography (rPPG) based BP estimation. Their best model achieved state-of-the-art results, demonstrating promising potential for non-invasive, continuous BP monitoring.

What Problem Does This Paper Address?

Accurate blood pressure (BP) monitoring is vital for cardiovascular health assessment. Traditional cuff-based methods are inconvenient for long-term monitoring. While remote photoplethysmography (rPPG) offers a promising contactless alternative, its application for BP estimation is hindered by a scarcity of publicly available datasets with corresponding BP labels. This data limitation makes it challenging to train robust deep learning models. Existing methods often rely on PPG signals, which are still contact-based, or struggle with the noise and variability inherent in rPPG signals captured from facial videos.

Methods

The core of this paper is a transfer learning methodology. The authors first train various deep learning models on existing, larger datasets of PPG signals to predict BP. This pre-training phase allows the models to learn generalizable features related to BP from a more abundant data source. Subsequently, these pre-trained models are fine-tuned using a smaller, publicly available dataset of rPPG signals. This fine-tuning process adapts the learned features to the specific characteristics of rPPG data.

They experimented with several deep learning architectures, including Alexnet, ResNet, BiLSTM, and BP-Net. They also investigated the impact of incorporating heart rate (HR) as an additional feature, explored different loss functions (MAE and MSE), and compared two min-max normalization strategies (per-signal and per-dataset). The study also examined the effectiveness of multi-task learning (predicting SBP and DBP simultaneously) versus single-task learning.

Datasets

The paper utilizes the following datasets:

  • MIMIC-III Subset from Kaggle: A PPG dataset used for pre-training. It contains ABP signals, ECG signals, and PPG signals.
  • MIMIC-III Subset from PulseDB: Another PPG dataset used for pre-training.
  • rPPG-LUH (Leipzig University Hospital): A rPPG dataset collected from patients in the ICU, used for fine-tuning and evaluation. This dataset was collected using an industrial USB camera.

Key Insights and Ideas

  • Transfer Learning is Key for Data Scarcity: The paper strongly advocates for transfer learning as a viable solution to overcome the limited availability of labeled rPPG data for BP estimation. Pre-training on PPG signals significantly improves performance on rPPG data.
  • Bidirectional LSTMs Excel: BiLSTM models consistently outperformed unidirectional LSTMs, suggesting that capturing both forward and backward temporal dependencies in PPG/rPPG signals is crucial for accurate BP prediction.
  • HR as a Feature: A Mixed Bag: While HR is known to correlate with BP, its inclusion as an additional feature did not consistently improve performance, especially when models were pre-trained on large PPG datasets. This suggests that the rich information in PPG signals might already implicitly capture HR-related dynamics. However, for noisier rPPG signals, HR might offer more distinct benefits.
  • Multi-task Learning is Efficient: The study found that multi-task learning (predicting SBP and DBP together) was sufficient and more computationally efficient than single-task learning, without a significant drop in performance.
  • Normalization Matters: The choice of normalization strategy (per-signal vs. per-dataset) can influence model performance, highlighting the importance of careful data preprocessing.

Key Results

The paper reports significant improvements in remote BP estimation using their transfer learning approach. Their best model, a fine-tuned BiLSTM, achieved the following Mean Absolute Errors (MAE) on the rPPG-LUH dataset:

  • Systolic Blood Pressure (SBP): 8.721 mmHg
  • Diastolic Blood Pressure (DBP): 8.653 mmHg

These results represent a substantial reduction in MAE compared to previous state-of-the-art methods on the same dataset, with improvements of 4.879 mmHg for SBP and 1.647 mmHg for DBP. The study also found that pre-training significantly reduced MAE for various models compared to training from scratch.

Limitations and Future Work

Limitations:

  • rPPG Signal Quality: The rPPG-LUH dataset, while valuable, contains noisy signals due to illumination variations and motion artifacts. This noise can impact the accuracy of extracted rPPG signals and subsequent BP estimation.
  • Dataset Imbalance: The rPPG-LUH dataset exhibits an imbalance in SBP and DBP distributions, with a prevalence of pre-hypertension.
  • Potential Data Leakage in PPG Datasets: The MIMIC-III subset from Kaggle might have had instances where data from the same patient was present in both training and testing sets, potentially inflating performance metrics.
  • Generalizability of Combined Datasets: Combining rPPG datasets from different sources can be challenging due to variations in camera setups, illumination, and rPPG extraction algorithms.

Future Work:

  • Addressing Data Imbalance: Exploring diverse resampling techniques and data augmentation methods to improve model generalization.
  • Combining Datasets: Investigating methods for effectively combining rPPG datasets from various sources, potentially by establishing standardized data collection protocols.
  • Advanced Architectures: Incorporating cutting-edge deep learning architectures like transformers for BP estimation.
  • Improved rPPG Extraction: Exploring novel deep learning methodologies for more robust and realistic rPPG signal extraction.

Summary updated

BibTeX
@inproceedings{cheng2025remote,
  author = {Chun-Hong Cheng and Jing Wei Chin and Kwan Long Wong and Tsz Tai Chan and Hau Ching Lo and Kwan Lok Pang and Richard So and Bryan Yan},
  title = {{Remote Blood Pressure Estimation from Facial Videos using Transfer Learning: Leveraging PPG to rPPG Conversion}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2025},
}
Figure