Table 2.
Performance metrics with Test B for MP-CNN model vs SP-CNN model.
| Algorithm | Model | AUROC (95% CI) | Accuracy (95% CI) | Sensitivity (95% CI) | Specificity (95% CI) | PPV (95% CI) | NPV (95% CI) |
|---|---|---|---|---|---|---|---|
| ResNet50 | MP-CNN | 0.913 (0.907–0.919) | 0.830 (0.805–0.855) | 0.784 (0.729–0.839) | 0.877 (0.854–0.900) | 0.866 (0.845–0.887) | 0.803 (0.764–0.842) |
| SP-CNN(Neu) | 0.890* (0.882–0.898) | 0.805 (0.789–0.821) | 0.800 (0.778–0.822) | 0.810 (0.780–0.840) | 0.810 (0.787–0.834) | 0.801 (0.783–0.818) | |
| SP-CNN(Flx) | 0.896* (0.887–0.905) | 0.817 (0.810–0.824) | 0.780 (0.768–0.792) | 0.855 (0.840–0.869) | 0.844 (0.832–0.857) | 0.794 (0.785–0.802) | |
| SP-CNN(Ext) | 0.890* (0.880–0.901) | 0.821 (0.801–0.841) | 0.800 (0.814–0.871) | 0.842 (0.814–0.871) | 0.838 (0.817–0.859) | 0.809 (0.773–0.845) | |
| VGG19 | MP-CNN | 0.914 (0.903–0.925) | 0.829 (0.810–0.848) | 0.782 (0.758–0.806) | 0.877 (0.852–0.901) | 0.865 (0.842–0.889) | 0.802 (0.790–0.813) |
| SP-CNN(Neu) | 0.893* (0.877–0.909) | 0.818 (0.789–0.847) | 0.778 (0.717–0.839) | 0.859 (0.779–0.939) | 0.849 (0.783–0.915) | 0.804 (0.789–0.818) | |
| SP-CNN(Flx) | 0.895* (0.882–0.908) | 0.819 (0.798–0.840) | 0.782 (0.746–0.818) | 0.857 (0.841–0.872) | 0.846 (0.829–0.863) | 0.790 (0.767–0.813) | |
| SP-CNN(Ext) | 0.890* (0.875–0.905) | 0.815 (0.804–0.826) | 0.774 (0.757–0.791) | 0.857 (0.826–0.887) | 0.846 (0.821–0.872) | 0.813 (0.798–0.838) | |
| VGG16 | MP-CNN | 0.910 (0.905–0.915) | 0.835 (0.829–0.842) | 0.782 (0.764–0.800) | 0.889 (0.875–0.903) | 0.877 (0.865–0.889) | 0.803 (0.764–0.842) |
| SP-CNN(Neu) | 0.889* (0.882–0.895) | 0.819 (0.806–0.832) | 0.796 (0.779–0.813) | 0.842 (0.829–0.856) | 0.836 (0.823–0.849) | 0.801 (0.783–0.818) | |
| SP-CNN(Flx) | 0.889* (0.884–0.895) | 0.819 (0.804–0.834) | 0.770 (0.737–0.803) | 0.869 (0.849–0.888) | 0.856 (0.839– 0.873) | 0.794 (0.785–0.802) | |
| SP-CNN(Ext) | 0.892* (0.885–0.898) | 0.824 (0.808–0.841) | 0.808 (0.777–0.839) | 0.840 (0.833–0.848) | 0.836 (0.827–0.845) | 0.809 (0.773–0.845) | |
| EfficientNet-B1 | MP-CNN | 0.899 (0.894–0.905) | 0.800 (0.792–0.808) | 0.714 (0.693–0.735) | 0.887 (0.879–0.894) | 0.865 (0.859– 0.870) | 0.755 (0.742–0.767) |
| SP-CNN(Neu) | 0.883* (0.873–0.893) | 0.800 (0.780–0.820) | 0.734 (0.701–0.767) | 0.867 (0.850–0.884) | 0.848 (0.830–0.866) | 0.764 (0.741–0.787) | |
| SP-CNN(Flx) | 0.877 (0.857–0.898) | 0.798 (0.778–0.818) | 0.726 (0.697–0.755) | 0.871 (0.856–0.885) | 0.850 (0.832–0.868) | 0.759 (0.738–0.780) | |
| SP-CNN(Ext) | 0.866* (0.848–0.884) | 0.790 (0.778–0.802) | 0.735 (0.714–0.755) | 0.846 (0.826–0.867) | 0.829 (0.812–0.847) | 0.759 (0.746–0.773) |
The best values are in bold.
Statistical analysis was performed using a paired t-test to compare the AUROC of each SP-CNN model with that of the MP-CNN model. An asterisk (*) means statistically significant difference (p < 0.05).
AUROC area under the receiver operating characteristics curve, PPV positive predictive value, NPV negative predictive value, CI confidence interval, MP-CNN multi pose-based convolutional neural network model, SP-CNN single pose-based convolutional neural network, Neu neutralposture, Flx flexion posture, Ext extension posture.