. 2021 Feb 25;11:4730. doi: 10.1038/s41598-021-83735-7

Table 4.

The classification accuracy of the proposed BNN-based system compared with the performance of each marker separately.

	Features	Sensitivity	Specificity	F1-score	Accuracy
2 Folds	Gibbs energy	$86.53 % \pm 3.9 %$	$71.04 % \pm 10.6 %$	$79.49 % \pm 6.1 %$	$78.08 % \pm 4.9 %$
	Thickness	$84.75 % \pm 1.4 %$	$76.42 % \pm 10.6 %$	$80.55 % \pm 7 %$	$80 % \pm 5.4 %$
	Tortuosity	$92.13 % \pm 6.1 %$	$71.26 % \pm 12.9 %$	$82.46 % \pm 5.9 %$	$80.77 % \pm 5.4 %$
	Reflectivity	$90.35 % \pm 3.5 %$	$100 % \pm 0 %$	$94.91 % \pm 2 %$	$95 % \pm 2.7 %$
	BNN-Based Fusion	$93.94 % \pm 1.0 %$	$97.3 % \pm 3.8 %$	$95.15 % \pm 1.9 %$	$95.38 % \pm 1.1 %$
4 Folds	Gibbs energy	$85.92 % \pm 13.2 %$	$72.99 % \pm 9.8 %$	$79.78 % \pm 8.5 %$	$78.85 % \pm 8 %$
	Thickness	$79.15 % \pm 7.2 %$	$79.21 % \pm 13 %$	$78.67 % \pm 10.1 %$	$79.23 % \pm 8.6 %$
	Tortuosity	$95.47 % \pm 3.9 %$	$72.32 % \pm 12.7 %$	$85.04 % \pm 7.9 %$	$83.46 % \pm 8.5 %$
	Reflectivity	$92.37 % \pm 5.9 %$	$96.86 % \pm 2.3 %$	$94.44 % \pm 4.5 %$	$94.62 % \pm 3.9 %$
	BNN-based fusion	$94.62 % \pm 4.4 %$	$98.65 % \pm 2.7 %$	$96.33 % \pm 1.4 %$	$96.54 % \pm 0.8 %$
10 Folds	Gibbs energy	$91.36 % \pm 7.4 %$	$71.73 % \pm 21.8 %$	$83.43 % \pm 9.3 %$	$81.92 % \pm 10.4 %$
	Thickness	$79.59 % \pm 15.7 %$	$74.27 % \pm 22.9 %$	$77.39 % \pm 15 %$	$78.08 % \pm 12.4 %$
	Tortuosity	$95.91 % \pm 5.6 %$	$73.67 % \pm 15.9 %$	$85.95 % \pm 10 %$	$84.62 % \pm 10.1 %$
	Reflectivity	$92.94 % \pm 8.3 %$	$96.7 % \pm - 5.7 %$	$94.66 % \pm 5.8 %$	$94.62 % \pm 5.8 %$
	BNN-based fusion	$96.75 % \pm 5.5 %$	$96.41 % \pm 4.9 %$	$96.82 % \pm 3.2 %$	$96.54 % \pm 3.8 %$
LOSO	Gibbs energy	86.92%	71.54%	80.71%	79.23%
	Thickness	83.85%	71.54%	78.99%	77.69%
	Tortuosity	94.62%	73.85%	85.71%	84.23%
	Reflectivity	90.77%	100.00%	95.16%	95.38%
	BNN-Based Fusion	96.15%	99.23%	97.66%	97.69%

Note that: BNN and LOSO stand for backpropagation neural network and leave-one-subject-out, respectively.