Table 1.
Balanced Accuracy (BACC), Sensitivity (SEN) and Specificity (SPE) of the proposed and comparative methods and Students t-test (99% confidence) results. The CMR and Echocardiography column headings indicate which data were used to apply the model. The first group corresponds to the results of the MMDL method for CRT response prediction, the second group corresponds to the results of the base-line method, the third group corresponds to the comparative methods and the last group corresponds to the results on the GSTFT CRT echocardiography database. An asterisk indicates a statistically significant improvement in accuracy over the base-line comparative approach. Bold-italic text indicates the method with the overall best classification accuracy, bold indicates the indicates the method with the highest classification accuracy for the la4Ch data, and dash indicates that this approach cannot be applied for this modality.
| CMR | Echocardiography | ||||||
|---|---|---|---|---|---|---|---|
| BACC(%)) | SEN(%) | SPE(%) | BACC(%) | SEN(%) | SPE(%) | ||
| 1. Multimodal deep learning approach | |||||||
| MMDL | 81.19* | 86.21 | 76.19 | 77.38* | 83.33 | 71.43 | |
| 2. Baseline approach (echocardiography-trained) | |||||||
| VGGEcho | - | - | - | 70.26 | 75.00 | 65.52 | |
| 3. Single modality (CMR-trained) approaches | |||||||
| VGGCMR | 71.98 | 68.97 | 75.00 | - | - | - | |
| VAESAX | 82.64 * | 87.50 | 77.78 | - | - | - | |
| VAEla4Ch | 78.30* | 84.38 | 72.22 | - | - | - | |
| GSTFT CRT echocardiography | |||||||
| BACC(%)) | SEN(%) | SPE(%) | BACC(%) | SEN(%) | SPE(%) | ||
| MMDL | - | - | - | 72.86 | 85.72 | 60.00 | |
| VGGEcho | - | - | - | 55.71 | 71.43 | 40.00 | |