Skip to main content

Table 2 Average performance of each model according to precision-recall AUC, F1 score, and F1 and PR AUC mean difference from baseline model

From: On the analysis of data augmentation methods for spectral imaged based heart sound classification using convolutional neural networks

 

F1 Score

[95% CI]

F1 Score Difference from Baseline [95% CI]

PR AUC [95% CI]

PR AUC Difference from Baseline Mean [95% CI]

Model 0

Baseline

86.7% [85.7, 87.7]

0.763 [0.734, 0.792]

Model 1

Pitch/time alterations

84.3% [82.4, 86.2] ↓

− 0.025 [− 0.042, − 0.006]

0.748 [0.703, 0.793] ↓

− 0.031 [− 0.067,0.005]

Model 2

Noise injection

84.6% [83.4, 85.8] ↓

− 0.021 [− 0.035, − 0.006]

0.757 [0.722, 0.792] ↓

− 0.032 [− 0.063, − 0.00002]

Model 3.1

Horizontal flip

87.9% [86.8, 89.0] ↑

0.012 [− 0.00030.024]

0.819 [0.792, 0.846] ↑

0.044 [0.013,0.073]

Model 3.2

Vertical flip

84.5% [83.0, 86.0] ↓

− 0.022 [− 0.038, − 0.006]

0.741 [0.695, 0.787] ↓

− 0.030 [− 0.070,0.0102]

Model 4.1

SV perturbations

87.6% [86.4, 88.8] ↑

0.008 [− 0.006,0.023]

0.784 [0.761, 0.807] ↑

0.005 [− 0.033,0.0425]

Model 4.2

PCA color augmentation

86.4% [85.3, 87.5] ↓

− 0.003 [− 0.014,0.008]

0.779 [0.751, 0.807] ↑

0.000 [− 0.029,0.029]

Model 4.3

Random color filters

85.3% [83.1, 87.5] ↓

− 0.014 [− 0.034,0.005]

0.754 [0.703, 0.805] ↓

− 0.029 [− 0.055, − 0.002]

Model 5

Time/frequency masking

85.1% [83.7, 86.5] ↓

− 0.016 [− 0.031, − 0.001]

0.772 [0.741, 0.803] ↑

− 0.007 [− 0.036,0.023]

Model 6

Horizontal flip and PCA

88.7% [87.5, 89.9] ↑

0.020 [0.004,0.034]

0.815 [0.772, 0.858] ↑

0.036 [− 0.0002,0.0712]

Model 7

Horizontal flip and SV perturbations

88.1% [87.2, 89.0] ↑

0.014 [0.005,0.0213]

0.802 [0.765, 0.839] ↑

0.026 [0.001,0.0507]