Table 7 The performance for 3 classes (baseline, stress, amusement) prediction on the WESAD dataset under leave-one-subject-out validation.

From: Dynamic clustering via branched deep learning enhances personalization of stress prediction from mobile sensor data

Model

F1-score \(20\%\)

F1-score \(40\%\)

F1-score \(60\%\)

LSTM

CALM-Net

Branched CALM-Net

0.758 ± 0.152

0.772 ± 0.113

0.753 ± 0.137

0.797 ± 0.140

0.832 ± 0.087

0.726 ± 0.159

0.784 ± 0.116

0.799 ± 0.116

0.803 ± 0.109

Transformer46

CATran-Net

Branched CATran-Net

0.761 ± 0.145

0.900 ± 0.100

0.801 ± 0.166

0.791 ± 0.176

0.890 ± 0.118

0.935 ± 0.093

0.944 ± 0.091

0.938 ± 0.089

0.960 ± 0.061

  1. The validation is conducted under leave-one-subject-out, with data from the left out subject being added to training set gradually to simulate the real-world online-learning setting. The percentage represent the amount of total available data included.
  2. [bold] values indicate top performers in each category.
Back to article page