Estimating blood pressure from the electrocardiogram: findings of alarge-scale negative results study

Objective. Electrocardiography and blood pressure (BP) measurement are two widely used tools for diagnosis and monitoring cardiovascular diseases. While the electrocardiogram (ECG) and BP have been considered complementary modalities, there are also systematic relationships between them. Therefore, advancements in portable and wearable ECG devices, along with promising results in cuff-less BP measurement using a combination of ECG and other bio-signals have led researchers to hypothesize the possibility of estimating BP and classifying BP categories (e.g. normal vs. hypertensive) using only ECG. However, the literature is divided on this topic: some studies support this hypothesis, while others reject it. Approach. In this study, regression and classification machine learning (ML) models were developed to explore the feasibility of estimating BP and predicting BP categories (normal vs. hypertensive) from 30 s ECGs using an extensive dataset from AliveCor Inc. which includes 124 427 records from 7412 subjects. The ECG and BP recordings were asynchronous with variable counts and time lags. Therefore, a 3.5 min time window before and after each ECG recording was used to calculate the mean BP measurement. Sex-aware ML models were trained using a comprehensive feature vector comprising 280 features: 128 explainable ECG features developed by the research team and 150 ECG features extracted by the Black Swan team, one of the top-performing teams in the PhysioNet Challenge 2017. Additionally, the average time gap between each ECG and the corresponding BP measurement, along with the subject’s age, were included as two supplementary features. Main results. Our best regression ML models achieved a mean absolute error of 12.59 mmHg for estimating systolic BP and 7.43 mmHg for diastolic BP, with correlation coefficients of 0.35 and 0.38 between the predicted and actual values, respectively. The best BP normal-hypertensive classification model achieved an area under the receiver operating characteristic curve of 0.655. Significance. Using a large dataset of ECG and BP recordings, this study found that ML models did not achieve acceptable performance in predicting BP values or classifying BP categories, indicating that BP cannot be reliably estimated from the ECG.