Table 3 External validation performance of multiview versus single-view DNNs on MHI data

From: Multiview deep learning improves detection of major cardiac conditions from echocardiography

DNN class

AUC

Sensitivity

Specificity

PPV

NPV

F1 score

P value

LV/RV abnormality cohort, test dataset (N = 1,650)

   

Multiview (A2c, A4c, PLAX)

0.909 (0.896–0.922)

0.837 (0.808–0.866)

0.799 (0.780–0.818)

0.614 (0.582–0.644)

0.928 (0.914–0.940)

0.708 (0.683–0.733)

ref.

Single-view A2c

0.787 (0.777–0.816)

0.604 (0.569–0.641)

0.812 (0.793–0.830)

0.550 (0.515–0.586)

0.843 (0.826–0.861)

0.576 (0.546–0.608)

<0.001

Single-view A4c

0.870 (0.854–0.885)

0.741 (0.705–0.775)

0.807 (0.788–0.826)

0.593 (0.561–0.627)

0.891 (0.874–0.906)

0.659 (0.632–0.686)

<0.001

Single-view PLAX

0.861 (0.843–0.876)

0.725 (0.689–0.761)

0.828 (0.809–0.845)

0.616 (0.583–0.650)

0.888 (0.871–0.903)

0.666 (0.638–0.694)

<0.001

Average of three single-view DNNs

0.892 (0.877–0.905)

0.802 (0.770–0.829)

0.800 (0.781–0.818)

0.604 (0.570–0.635)

0.914 (0.900–0.926)

0.689 (0.661–0.714)

0.031

Diastolic dysfunction cohort, test dataset (N = 766)

   

Multiview (A2c, A4c, PLAX)

0.791 (0.765–0.817)

0.502 (0.454–0.551)

0.852 (0.824–0.878)

0.694 (0.642–0.743)

0.719 (0.686–0.751)

0.582 (0.538–0.627)

ref.

Single-view A2c

0.647 (0.615–0.678)

0.603 (0.558–0.647)

0.599 (0.563–0.636)

0.501 (0.458–0.544)

0.693 (0.655–0.730)

0.547 (0.509–0.582)

<0.001

Single-view A4c

0.713 (0.683–0.743)

0.599 (0.551–0.646)

0.712 (0.679–0.747)

0.582 (0.536–0.629)

0.727 (0.692–0.763)

0.591 (0.553–0.630)

<0.001

Single-view PLAX

0.698 (0.666–0.728)

0.824 (0.789–0.860)

0.488 (0.450–0.523)

0.518 (0.481–0.555)

0.806 (0.764–0.846)

0.636 (0.602–0.668)

<0.001

Average of three single-view DNNs

0.743 (0.714–0.772)

0.404 (0.358–0.453)

0.852 (0.824–0.880)

0.646 (0.588–0.699)

0.681 (0.652–0.714)

0.497 (0.448–0.543)

0.024

Valve regurgitation cohort, test dataset (N = 303)

   

Multiview (A5c, A4c, PLAX)

0.924 (0.890–0.954)

0.750 (0.647–0.857)

0.886 (0.853–0.918)

0.500 (0.400–0.607)

0.959 (0.937–0.976)

0.600 (0.506–0.689)

ref.

Single-view A5c

0.878 (0.838–0.914)

0.725 (0.606–0.838)

0.840 (0.802–0.877)

0.408 (0.319–0.500)

0.953 (0.929–0.974)

0.523 (0.430–0.610)

0.658

Single-view A4c

0.881 (0.834–0.920)

0.725 (0.609–0.833)

0.817 (0.778–0.855)

0.377 (0.292–0.464)

0.951 (0.927–0.974)

0.496 (0.404–0.580)

0.100

Single-view PLAX

0.810 (0.746–0.870)

0.600 (0.486–0.725)

0.905 (0.875–0.935)

0.490 (0.375–0.612)

0.937 (0.913–0.960)

0.539 (0.432–0.641)

0.009

Average of three single-view DNNs

0.915 (0.875–0.948)

0.725 (0.615–0.833)

0.886 (0.856–0.919)

0.492 (0.390–0.606)

0.955 (0.932–0.975)

0.586 (0.488–0.678)

1.000

  1. All values are point estimate (95% CI, derived via bootstrap). We present P values comparing the AUC of multiview DNNs against each single-view DNN and the average of three single-view DNNs using the two-sided DeLong’s test with Bonferroni correction. P values lower than 1 × 10−3 are displayed as <0.001.
Back to article page