MIMIC-BP: A curated dataset for blood pressure estimation

Area of Science:

• Biomedical Engineering
• Cardiovascular Health
• Machine Learning

Background:

• Blood pressure (BP) monitoring is crucial for cardiovascular health but traditional cuff-based methods are inconvenient.
• Deep learning offers potential for pervasive BP estimation, yet existing datasets have limitations like data imbalance and lack of subject identification.
• These limitations can lead to data leakage and algorithmic bias in BP estimation models.

Purpose of the Study:

• To introduce a novel, large-scale, and well-annotated biomedical signal dataset for advancing blood pressure estimation.
• To address the limitations of existing datasets, including data imbalance and lack of subject identification.
• To facilitate the development and validation of robust, calibration-free deep learning models for blood pressure monitoring.

Main Methods:

• A derivative dataset was created comprising 380 hours of biomedical signals (arterial blood pressure, photoplethysmography, electrocardiogram).
• The dataset includes data from 1,524 anonymized subjects, with each subject contributing 30 segments of 30-second signal recordings.
• State-of-the-art deep learning methods were employed to validate the dataset's utility.

Main Results:

• The proposed dataset was validated using advanced deep learning techniques.
• Experiments demonstrated the dataset's suitability for training and testing blood pressure estimation algorithms.
• The validation highlighted the importance of standardized benchmarks for calibration-free BP estimation.

Conclusions:

• The newly curated dataset provides a valuable resource for research in non-invasive blood pressure estimation.
• This dataset can help overcome limitations of previous datasets, reducing bias and improving model generalizability.
• It supports the development of more accurate and accessible tools for continuous cardiovascular health monitoring.