Table 4 Performance of multiview versus single-view DNNs for individual components of the composite labels

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

DNN class

AUC

Sensitivity

Specificity

PPV

NPV

F1 score

LV size (N = 6,256)

Multiview

0.902 (0.89 –0.914)

0.819 (0.789–0.848)

0.820 (0.812–0.829)

0.257 (0.238–0.277)

0.984 (0.98–0.986)

0.392 (0.368–0.415)

Single-view A2c (a)

0.826 (0.809–0.843)

0.67 (0.63–0.706)

0.812 (0.803–0.82)

0.213 (0.194–0.231)

0.97 (0.965–0.974)

0.323 (0.298–0.346)

Single-view A4c (a,b)

0.843 (0.827–0.858)

0.753 (0.72–0.787)

0.767 (0.758–0.776)

0.197 (0.182–0.213)

0.976 (0.972–0.98)

0.313 (0.291–0.335)

Single-view PLAX (b)

0.859 (0.843–0.874)

0.733 (0.697–0.769)

0.811 (0.803–0.82)

0.228 (0.21–0.246)

0.976 (0.972–0.979)

0.348 (0.324–0.371)

LVEF (N = 6,256)

Multiview

0.942 (0.935–0.948)

0.823 (0.802–0.842)

0.906 (0.899–0.912)

0.616 (0.592–0.638)

0.966 (0.961–0.97)

0.704 (0.685–0.722)

Single-view A2c

0.879 (0.869–0.888)

0.807 (0.785–0.826)

0.783 (0.773–0.792)

0.405 (0.385–0.422)

0.957 (0.952–0.961)

0.539 (0.519–0.556)

Single-view A4c (c)

0.900 (0.891–0.909)

0.84 (0.82–0.859)

0.791 (0.781–0.8)

0.423 (0.403–0.441)

0.964 (0.959–0.969)

0.563 (0.543–0.58)

Single-view PLAX (c)

0.898 (0.888–0.907)

0.776 (0.755–0.797)

0.851 (0.843–0.859)

0.488 (0.467–0.508)

0.954 (0.949–0.959)

0.599 (0.58–0.618)

RV function (N = 6,256)

Multiview

0.931 (0.920–0.940)

0.906 (0.878–0.931)

0.821 (0.813–0.829)

0.219 (0.201–0.237)

0.994 (0.992–0.995)

0.353 (0.328–0.376)

Single-view A2c

0.814 (0.796–0.832)

0.778 (0.739–0.817)

0.703 (0.693–0.713)

0.127 (0.114–0.139)

0.983 (0.979–0.986)

0.218 (0.199–0.237)

Single-view A4c (d)

0.896 (0.883–0.907)

0.872 (0.841–0.902)

0.733 (0.723–0.742)

0.154 (0.14–0.167)

0.99 (0.988–0.993)

0.261 (0.241–0.281)

Single-view PLAX (d)

0.881 (0.865–0.894)

0.802 (0.765–0.837)

0.811 (0.803–0.819)

0.191 (0.173–0.208)

0.987 (0.984–0.989)

0.308 (0.283–0.332)

RV size (N = 6,256)

Multiview

0.897 (0.885–0.910)

0.835 (0.809–0.863)

0.805 (0.796–0.813)

0.259 (0.241–0.277)

0.984 (0.981–0.987)

0.395 (0.373–0.417)

Single-view A2c

0.705 (0.685–0.724)

0.665 (0.63–0.702)

0.618 (0.607–0.628)

0.124 (0.114–0.135)

0.958 (0.952–0.963)

0.209 (0.194–0.226)

Single-view A4c (e)

0.861 (0.845–0.874)

0.79 (0.759–0.82)

0.764 (0.755–0.773)

0.215 (0.199–0.231)

0.978 (0.974–0.982)

0.338 (0.316–0.358)

Single-view PLAX (e)

0.845 (0.827–0.861)

0.752 (0.718–0.785)

0.793 (0.784–0.801)

0.229 (0.212–0.247)

0.975 (0.971–0.979)

0.351 (0.328–0.373)

Mitral regurgitation (N = 4,148)

Multiview

0.930 (0.915–0.943)

0.842 (0.798–0.879)

0.866 (0.858–0.875)

0.268 (0.241–0.295)

0.990 (0.986–0.992)

0.407 (0.373–0.439)

Single-view A5c (f)

0.868 (0.849–0.887)

0.794 (0.747–0.837)

0.767 (0.756–0.778)

0.165 (0.147–0.184)

0.985 (0.981–0.988)

0.274 (0.247–0.3)

Single-view A4c (f)

0.858 (0.837–0.876)

0.768 (0.72–0.811)

0.771 (0.759–0.783)

0.163 (0.145–0.181)

0.983 (0.979–0.986)

0.269 (0.242–0.295)

Single-view PLAX

0.900 (0.884–0.915)

0.829 (0.788–0.868)

0.806 (0.796–0.816)

0.199 (0.177–0.22)

0.988 (0.985–0.991)

0.321 (0.29–0.349)

Aortic regurgitation (N = 4,148)

Multiview

0.929 (0.895–0.959)

0.862 (0.791–0.931)

0.884 (0.875–0.892)

0.105 (0.086–0.128)

0.998 (0.996–0.999)

0.188 (0.155–0.223)

Single-view A5c (g)

0.873 (0.826–0.915)

0.661 (0.565–0.754)

0.913 (0.906–0.920)

0.108 (0.084–0.135)

0.994 (0.992–0.996)

0.186 (0.147–0.228)

Single-view A4c (h)

0.786 (0.733–0.836)

0.692 (0.597–0.787)

0.742 (0.730–0.753)

0.041 (0.032–0.052)

0.993 (0.991–0.996)

0.077 (0.061–0.096)

Single-view PLAX (g,h)

0.863 (0.821–0.902)

0.662 (0.561–0.758)

0.844 (0.834–0.853)

0.063 (0.048–0.079)

0.994 (0.991–0.996)

0.115 (0.088–0.142)

Tricuspid regurgitation (N = 4,148)

Multiview (i,j)

0.864 (0.845–0.881)

0.782 (0.738–0.822)

0.814 (0.804–0.825)

0.221 (0.200–0.243)

0.982 (0.979–0.986)

0.345 (0.316–0.373)

Single-view A5c (k)

0.813 (0.788–0.836)

0.733 (0.684–0.780)

0.766 (0.755–0.777)

0.175 (0.156–0.193)

0.977 (0.972–0.982)

0.282 (0.256–0.307)

Single-view A4c (i)

0.884 (0.867–0.899)

0.863 (0.824–0.896)

0.743 (0.732–0.755)

0.184 (0.166–0.202)

0.988 (0.984–0.991)

0.304 (0.278–0.329)

Single-view PLAX (j,k)

0.842 (0.819–0.864)

0.733 (0.688–0.775)

0.798 (0.788–0.809)

0.197 (0.176–0.217)

0.978 (0.974–0.982)

0.310 (0.283–0.337)

  1. Multiview and single-view DNNs were trained for example individual components of the composite tasks in Table 2. Single-view models were trained using videos from the same studies as multiview models, but from a single view. All values are point estimate (95% CI, derived via bootstrap). Superscript letters indicate model pairs within each task with non-significant differences in AUC (P > 0.05), as determined using the two-sided DeLong’s test with Bonferroni correction for multiple comparisons. All model pairs within each task without superscripts had AUCs that were statistically significantly different from each other.
Back to article page