Causal Inference for Hypertension Prediction with Wearable Devices - Aurora Project ECG/PPG Dataset

Overview

The Aurora Project Dataset for Hypertension Prediction published in JMIR Cardio (January 2025) provides high-quality wearable sensor data specifically curated for causal inference and machine learning research on hypertension prediction from non-invasive signals.

Dataset Characteristics

• Large cohort of participants balanced across key demographic and clinical factors: gender (approximately equal male/female representation), age groups (young, middle-aged, elderly), and hypertension status (normotensive vs. hypertensive).
• ECG and PPG signals simultaneously acquired using wrist-worn wearable devices during controlled conditions to ensure signal quality and synchronization.
• Reference blood pressure measurements obtained using validated oscillometric or auscultatory methods, providing ground truth for hypertension classification (systolic BP ≥ 140 mmHg or diastolic BP ≥ 90 mmHg).
• Multi-session data collection enabling longitudinal analysis and within-subject variability assessment.

Feature Extraction and Engineering

• 205 features extracted from ECG and PPG waveforms using signal processing and domain knowledge, including:
• Time-domain features: R-R intervals, pulse arrival time (PAT), pulse transit time (PTT), systolic/diastolic peaks, pulse width, inter-beat intervals.
• Frequency-domain features: Power spectral density components, heart rate variability frequency bands (LF, HF, LF/HF ratio).
• Morphological features: Waveform slopes, area under curves, symmetry metrics, and fiducial point characteristics.
• Statistical metrics (6 categories): Mean, standard deviation, median, skewness, kurtosis, and percentiles computed for each of the 205 features to capture distribution properties.

Causal Inference Methodology

• Rather than purely correlational machine learning, the research applies causal discovery to identify features that have genuine causal relationships with hypertension (not just statistical associations).
• Six different causal graph structures explored to model relationships among extracted features and hypertension outcomes.
• Selected causal features used to train predictive models, improving interpretability and potentially enhancing generalization to new populations.
• Evaluation using multiple metrics (accuracy, sensitivity, specificity, AUC) to assess both prediction performance and clinical utility.

Use Cases

• Cuffless blood pressure estimation: Developing algorithms to predict hypertension status from wearable ECG/PPG without traditional cuff-based measurements.
• Causal modeling in health AI: Advancing interpretable and trustworthy AI by identifying features with causal (not just correlational) links to cardiovascular outcomes.
• Wearable device validation: Benchmarking the accuracy of consumer and medical-grade wearables for hypertension screening.
• Personalized risk assessment: Creating individualized hypertension risk models based on longitudinal wearable data.
• Clinical decision support: Integrating continuous wearable monitoring into primary care for early detection and management of hypertension.