Table 2 Evaluation of different training strategies on the Eindhoven PPG data set.

From: A deep transfer learning approach for wearable sleep stage classification with photoplethysmography

Model

Training procedure summary

Cohen’s kappa

Accuracy (%)

ECG-trained model

Traina on Siesta

0.57 ± 0.12

71.88 ± 8.34

PPG-trained model

Trainb on Eindhoven

0.55 ± 0.14

69.82 ± 10.23

Domain retrain

Pre-traina on Siesta + adaptb using Eindhoven

0.62 ± 0.12

75.21 ± 7.82

Decision retrain

Pre-traina on Siesta + adaptb using Eindhoven

0.63 ± 0.12

75.14 ± 8.10

Combined retrain

Pre-traina on Siesta + adaptb using Eindhoven

0.65 ± 0.11

76.36 ± 7.57

  1. aTraining was done on the entire Siesta ECG data set.
  2. bDone in 4-fold cross-validation. In each fold 45 participants of the Eindhoven data set were used for training and 15 were left out for validation. Shown results are aggregated over all folds. All cross-validation experiments used the same folds to enable comparison. Results are presented as mean ± standard deviation. Distribution of performance over participants and statistical significance tests are shown in Fig. 5.
Back to article page