Intro: Aging is linked to structural and mechanical changes in the vascular wall. These changes lead to increased arterial stiffness, which is a primary indicator for assessing risk of cardiovascular diseases. Photoplethysmography (PPG) is a technology to detect blood volume variations in the arteries and can be measured using a smartphone camera. PPG is influenced by arterial stiffness and shows promise as a potential marker for vascular aging. Using deep learning, we aim to predict age from PPG and identify how age is represented in the waveform of the PPG.
Methods: We utilised real world smartphone acquired PPG signals with patient reported age and gender. PPGs were excluded if the quality was too low or there was no record of age and/or gender. The model architecture of InceptionTime was used for the regression task and a grid search was performed to select the optimal model hyperparameters and PPG window size. Saliency plots were calculated to highlight predictive regions for age within the PPG.
Results: We included 23 237 PPG of 5 505 patients with a mean age of 46.8 ± 16.1 years. The grid search indicated the best results for the default hyperparameters and a signal length of 16 consecutive heartbeats. The optimal model resulted in a mean absolute error of 9.52 ± 7.34 years and an r2-score of 0.32. Saliency plots highlighted the region around the dicrotic notch as most determining for age prediction, which has been previously linked to arterial stiffness and vascular age.
Conclusion: Real world smartphone acquired PPG allows for the estimation of aging trends . The gravity of the dicrotic notch in saliency plots provides a rationale for its prediction and suggests that PPG-derived age may serve as a potential marker for vascular aging.