Age estimation via electrocardiogram from smartwatches

Figures

• 

  Fig. 1. An ideal ECG waveform16.

• 

  Fig. 2. Histogram of age distribution in our dataset.

• 

  Fig. 3. Age-related changes in Heart Rate Variability (HRV). a Standard deviation of heart rate variability (HRV STD) against age. b Root mean square of successive differences (HRV RMSSD) against age.

• 

  Fig. 4. Predicted age vs. actual age, corresponding to the lowest MAE.

• 

  Fig. 5. Predicted age vs. actual age, for male participants.

• 

  Fig. 6. Predicted age vs. actual age, for female participants.

• 

  Fig. 7. Hyperparameter optimization for age estimation using FNN, LSTM, and Inception-1D models. FNN: MAE vs a number of hidden units, b Number of hidden layers, c learning rate, d batch size, e epochs, f dropout rate, g input feature length. LSTM: MAE vs h number of LSTM layers, i number of hidden units, j input sequence length, k learning rate, l batch size, m epochs, n dropout rate. Inception-1D: MAE vs o number of inception blocks, p number of filters per block, q input sequence length, r learning rate, s batch size, t Epochs, u dropout rate.

• 

  Fig. 8. Saliency map revealing key ECG features for age prediction from de-noised ECG using FNN in different ages. a age 3, b age 10, c age 20, d age 30, e age 40, f age 50, g age 60, h age 70.