Age Estimation via Electrocardiogram from Smartwatches

Age estimation, a crucial legal requirement for accessing age-restricted services or products, has recently gained global attention to enhance online safety for children. Electrocardiogram (ECG) signals, which reflect the electrical activity of the heart, have emerged as a potential method for age estimation due to certain age-dependent attributes. Traditionally, studies have relied on clinical ECG data from hospital settings, which is impractical for widespread use. To address this, we develop a dataset using ECG signals collected from a smartwatch worn by 172 individuals of varying ages. We then experimented using a combination of different features and different machine learning schemes. A method combining Discrete Wavelet Transform (DWT) and Long Short-Term Memory (LSTM) yields the best result, providing a Mean Absolute Error (MAE) of 6.58 in age prediction, outperforming previous studies based on clinical ECG data. The incorporation of DWT notably enhanced performance due to its effective morphology matching, frequency localization, and robustness. Age estimation accuracy was significantly higher in younger individuals (below 30 years) compared to older age groups, aligning with established patterns of rapid early-adulthood changes in ECG characteristics and heart rate variability. Additionally, we performed binary classification within specific age thresholds, from 12 to 22 years, achieving accuracy ranging from 72–95%. These promising results highlight the potential for further research and the application of smartwatch-based ECG for anonymous age estimation.