Table 3.
Performance in TKR prediction of OA pretrained models for radiographs and MRI, stratified by severity of OA. Pretraining strategy yields useful information to both models, but performance at no OA in particular leaves room for improvement, justifying subsequent model fine-tuning. Standard errors used to calculate confidence intervals.
| OA status | Model type | Accuracy (95% CI) | Sensitivity (95% CI) | Specificity (95% CI) | Non-TKR cases | TKR cases |
|---|---|---|---|---|---|---|
| None | Radiograph | 92.1 ± 0.083 | 25.2 ± 2.16 | 92.4 ± 0.081 | 2,892 | 12 |
| MRI | 94.3 ± 0.070 | 48.7 ± 2.48 | 94.4 ± 0.070 | |||
| Moderate | Radiograph | 29.3 ± 0.151 | 93.8 ± 0.439 | 26.7 ± 0.156 | 2,056 | 83 |
| MRI | 65.4 ± 0.154 | 65.5 ± 0.848 | 65.4 ± 0.158 | |||
| Severe | Radiograph | 29.7 ± 0.488 | 100.0 ± 0.000 | 1.4 ± 0.180 | 141 | 57 |
| MRI | 33.4 ± 0.523 | 82.2 ± 0.824 | 14.0 ± 0.441 | |||
| All | Radiograph | 64.2 ± 0.124 | 90.7 ± 0.378 | 63.4 ± 0.126 | 5,089 | 152 |
| MRI | 80.2 ± 0.079 | 70.4 ± 0.595 | 80.5 ± 0.082 |