Introduction: Aging is a major risk factor for functional impairment and life-time diseases. The measurement of aging by chronological means is imperfect. As the aging process affects the heart with the electrocardiogram (ECG) being a direct representation of its physiology, deep learning models can predict an individual’s age from the ECG. The difference between predicted and actual (i.e. chronological) age is commonly used to reflect the aging process. Utilizing data from a large population-based cohort study, we determine the predicted age difference (PAD) from the ECG as a marker for slow and accelerated aging of the heart. Furthermore, we explore the impact of smoking on the PAD.

Methods: In this study we utilize the 12-channel-ECGs of 8489 participants (median age 62.0 years, 49.6% men) and corresponding age as well as smoking status levels (pack years)  from the population-based Hamburg City Health Study (HCHS). We split the dataset into a large hold-out test dataset (n=2932) and training data (n=5557) for a Resnet-1D neural network. To further refine the model, we also include the PTB-XL ECG dataset for training (n=18803 individuals), which is a unified collection of public ECG dataset of 10s length. We calculate the predicted age difference (PAD) as the difference of the predicted and actual age of the participants on our hold out test dataset, a measurement commonly used for ageing, which is then associated with the pack years from participants with smoking status in our hold out test dataset (n=462) using spearman correlation and independent t-tests on participants with more than 40 and more than 50 pack years compared to participants, which have never smoked.

Results: The model employing only the HCHS dataset achieved a R-squared [R²] of 0.22 and mean absolute error [mae] of 6.16 predicting chronological age, when incorporating the PTB-XL dataset the mae improves to 5.57 with a R² of 0.325 (Fig. 1). Therefore, the model benefits from the diverse PTB-XL data, which also covers a greater range of represented age values. Using this model, we show that the PAD is associated with pack years (Spearman coefficient 0.2, p<0.001) and according to our results, the first 30 pack years do not significantly impact the ECG-based PAD, while the PAD for 40 and more pack years is significantly increased compared to never-smokers (>40 pack years: t=3.6, p<0.001; >50 pack years: t=3.7, p<0.001) (Fig.2).

Conclusions: Using machine-learning-based ECG analyses we are able to quantify aging. Furthermore, a history of smoking is associated with an increased predicted aging difference. Higher predicted age values occur for those who have a smoking history of >40 pack years. These results allow for the hypothesis that smoking leads to an accelerated aging process in the heart.