Deep Learning Algorithms for Estimation of Demographic and Anthropometric Features from Electrocardiograms

Area of Science:

• Cardiology and Artificial Intelligence
• Medical Informatics
• Biomedical Signal Processing

Background:

• Electrocardiograms (ECGs) are influenced by demographic and anthropometric factors.
• Existing methods for assessing health status may not fully capture individual physiological variations.
• There is a need for novel biomarkers to provide a more accurate reflection of an individual's health status.

Purpose of the Study:

• To develop and validate deep learning models for predicting age, sex, body mass index (BMI), and ABO blood type from ECG data.
• To explore the potential of ECG-derived features as physiological biomarkers.
• To assess the performance of convolutional neural networks (CNNs) in these prediction tasks.

Main Methods:

• A retrospective study utilizing 124,415 ECGs from individuals aged 18 years and older.
• Development of CNN models with specific configurations (three convolutional layers, five kernel sizes, two pooling sizes).
• Implementation of both classification and regression models for various parameters, with evaluation using AUROC and MAE.

Main Results:

• High accuracy in predicting age (AUROC: 0.923, MAE: 8.410) and sex (AUROC: 0.947).
• Moderate performance in predicting BMI (AUROC: 0.765, MAE: 2.332).
• Limited performance in predicting ABO blood type (top-1 accuracy: 31.98%).

Conclusions:

• Deep learning models can effectively estimate demographic and anthropometric features from ECGs.
• ECG-derived predictions hold promise for developing advanced physiological biomarkers.
• These biomarkers could offer a more nuanced understanding of health status compared to traditional measures like chronological age.