🫀 Robust_Blood_Pressure_Benchmark

A Curated, Balanced, and High-Fluctuation MIMIC-II Subset for Continuous Blood Pressure Estimation

---

Language: English | 简体中文

📌 Overview

One of the major bottlenecks in developing Machine Learning models for continuous non-invasive blood pressure (cNIBP) estimation is the extreme class imbalance in raw clinical databases like MIMIC-II. Most segments represent normal blood pressure, making models perform poorly on critical hypertensive or hypotensive events.

Robust-BP-Bench solves this by providing a rigorous 4-bin stratified sampling protocol. We extracted a highly curated subset from MIMIC-II that forces an equal representation across all BP ranges, ensuring your models learn robust features rather than just predicting the mean.

Key Features

• 4-Bin Stratification: Perfectly balanced classes across Hypotension (<110), Normal (110-130), Pre-hypertension (130-150), and Hypertension (>150 mmHg).
• Strict Quality Control: Built-in algorithms to filter out low-variance signals, severe artifacts, and poor linearity segments.
• Ready for Deep Learning: High Signal-to-Noise Ratio (SNR) segments ready for feature extraction or end-to-end training.

---

🚀 Zero-Setup Quickstart (Demo Mode)

Want to see how the selection protocol works without downloading the 50GB raw MIMIC-II database? We provide a zero-dependency synthetic data generator.

# 1. Clone the repository
git clone [https://github.com/phish-tech/Robust-BP-Bench.git](https://github.com/phish-tech/Robust-Blood_Pressure-Benchmark.git)
cd Robust-BP-Bench

# 2. Run the synthetic data generator in MATLAB
>> generate_demo_dataset

# 3. Execute the core selection protocol (Ensure IS_DEMO_MODE = true)
>> run_data_selection

The script will instantly parse the demo data and output a perfectly balanced demo_sampled_balanced.mat.

💾 Download the Full Dataset

The complete, 1GB curated "Golden Subset" derived from the actual MIMIC-II database is hosted on Zenodo for permanent open access.

👉 Download [Robust-BP-Bench]{https://zenodo.org/records/19912053} (dataset_sampled_balanced.mat) via Zenodo

📊 Dataset Demographics

Our protocol guarantees a balanced distribution, which is crucial for training unbiased regression models.

Overall Dataset Distribution

Selected Sample Dataset Distribution

📚 Citation

If you use this data selection protocol or the curated dataset in your research, please cite our upcoming EMBC 2026 paper:

From Elastic to Viscoelastic: An EEMD-Enhanced Pulse Transit Time Model for Robust Blood Pressure Estimation

---

🔬 Research Based on This Dataset: Accepted by EMBC 2026

The open-sourced Robust-BP-Bench dataset in this repository was curated specifically to support our group's latest breakthrough in continuous non-invasive blood pressure (cNIBP) estimation. This work has been officially accepted by IEEE EMBC 2026.

📄 Paper Title: From Elastic to Viscoelastic: An EEMD-Corrected PTT Model for Precise Blood Pressure Tracking

💡 Why Do Traditional PTT Models Fail?

Existing Pulse Transit Time (PTT) models are widely based on the Moens-Korteweg equation, which fundamentally assumes human blood vessels are "purely elastic" rigid tubes. However, real biological tissues exhibit Viscoelasticity. During severe blood pressure fluctuations, this viscoelasticity introduces a significant "Hysteresis" effect between PTT and actual BP, causing the accuracy of traditional models to drop precipitously.

🚀 Our Approach (The Secret Sauce)

We propose a novel EEMD-corrected PTT physical hybrid model, achieving a paradigm shift from "purely elastic" to "viscoelastic":

1. Intersecting Tangent Method for Foot Localization: Abandoning highly unstable peak detection, we utilize 10x Makima high-fidelity interpolation and the maximum rising slope tangent to provide a stable benchmark for PTT calculation that strictly aligns with hemodynamic definitions.
2. Viscoelastic Compensation via EEMD: Utilizing Ensemble Empirical Mode Decomposition (EEMD), we "dismantle" the PPG signal and extract the differential energy of high-frequency modes as the "viscoelastic compensation feature." This perfectly quantifies the signal's kinematic intensity, drastically offsetting motion artifacts and hysteresis errors.

🏆 Experimental Performance

Tested on the extreme clinical dataset open-sourced in this repository (which includes 23.4% hypertensive samples), our algorithm demonstrated excellent performance:

• Robust Beat-to-Beat Tracking: Whether the blood pressure is experiencing severe fluctuations, sharp ascents, or downward trends, the EEMD-corrected model closely tracks the true arterial blood pressure (Ground Truth).

🔔 Full Core Code Teaser: The algorithmic framework, including EEMD signal decomposition and tangent foot localization, will be open-sourced in this repository upon the official publication of the paper.
[⭐ Star this repository] to get notified of updates instantly!

🌐 For more information about the EMBC 2026 conference, please visit: EMBC Official Website