Confusion matrices for the physicians (survey responses from 61 physicians classifying lung ultrasound clips into their respective causes, numbers in parenthesis reflect classifications from the aggregated approach used to calculate area under the receiver operating characteristic curve), model performance on the test-2 holdback set at the frame and the encounter level
| Physicians | Predicted | Total | |||
| COVID | NCOVID | HPE | |||
| Actual | COVID | 173 (3) | 162 (3) | 34 (2) | 369 (8) |
| NCOVID | 177 (4) | 163 (1) | 30 (2) | 370 (7) | |
| HPE | 138 (0) | 102 (0) | 302 (6) | 542 (6) | |
| Total | 488 (7) | 427 (4) | 366 (10) | ||
| CNN-Frames | Predicted | Total | |||
| COVID | NCOVID | HPE | |||
| Actual | COVID | 3188 | 256 | 7 | 3451 |
| NCOVID | 1176 | 3741 | 3 | 4920 | |
| HPE | 109 | 1119 | 2771 | 3999 | |
| Total | 4473 | 5116 | 2781 | ||
| CNN-Encounters | Predicted | Total | |||
| COVID | NCOVID | HPE | |||
| Actual | COVID | 6 | 0 | 0 | 6 |
| NCOVID | 1 | 6 | 0 | 7 | |
| HPE | 0 | 3 | 4 | 7 | |
| Total | 7 | 9 | 4 | ||
‘Predicted’ represents the model or physicians’ opinions; ‘actual’ is the true label of the clip.
CNN, convolutional neural network; HPE, hydrostatic pulmonary edema.