Skip to main content

Table 1 The mean and standard deviation of the classification accuracies, F1-scores, and AUCs over 5 different random training-validation-test data splittings. Our proposed model, which benefits from the incorporation of multiple complementary drawing tasks (clock drawing, cube-copying, and trail-making), self-attention mechanism, and soft labeling approach, achieved much higher mean accuracy, F1-score, and AUC than the baseline model

From: An explainable self-attention deep neural network for detecting mild cognitive impairment using multi-input digital drawing tasks

Models

Accuracy

F1-score

AUC

VGG16 with only clock-drawing test

0.7478 ± 0.0071

0.3573 ± 0.0443

0.7429 ± 0.0131

VGG16 with only cube-copying test

0.7739 ± 0.0096

0.4994 ± 0.0477

0.7813 ± 0.0197

VGG16 with only trail-making test

0.7739 ± 0.0249

0.5283 ± 0.0548

0.7722 ± 0.0240

Multi-input VGG16

0.7986 ± 0.0071

0.5938 ± 0.0207

0.8115 ± 0.0192

Conv-Att with only clock-drawing test

0.7522 ± 0.0125

0.3586 ± 0.0309

0.7337 ± 0.0204

Conv-Att with only cube-copying test

0.7768 ± 0.0168

0.5095 ± 0.0515

0.7791 ± 0.0199

Conv-Att with only trail-making test

0.7696 ± 0.0167

0.5211 ± 0.0272

0.7662 ± 0.0231

Multi-input Conv-Att

0.7986 ± 0.0071

0.5981 ± 0.0221

0.8379 ± 0.0176

Multi-input Conv-Att with a soft label (proposed)

0.8116 ± 0.0103

0.6539 ± 0.0097

0.8375 ± 0.0116