Table 3 Confusion matrices for the three proposed cut-offs for the model predictions regarding presence of definite radiographic sacroiliitis on the test dataset
From: Deep learning for detection of radiographic sacroiliitis: achieving expert-level performance
| nr-axSpA | r-axSpA | ||
|---|---|---|---|
| Cut-off 1, favouring sensitivity over specificity | |||
| Model predicts nr-axSpA | 36 | 4 | 40 |
| Model predicts r-axSpA | 93 | 219 | 312 |
| 129 | 223 | 352 | |
| Cohen’s kappa | 0.3 (95% CI 0.21–0.4) | Accuracy | n = 255/352 (72.4%) |
| Cut-off 2, favouring specificity over sensitivity | |||
| Model predicts nr-axSpA | 120 | 41 | 161 |
| Model predicts r-axSpA | 9 | 182 | 191 |
| 129 | 223 | 352 | |
| Cohen’s kappa | 0.7 (95% CI 0.63–0.77) | Accuracy | n = 302/352 (85.8%) |
| Cut-off 3, optimal relationship between sensitivity and specificity | |||
| Model predicts nr-axSpA | 104 | 19 | 123 |
| Model predicts r-axSpA | 25 | 204 | 229 |
| 129 | 223 | 352 | |
| Cohen’s kappa | 0.72 (95% CI 0.65–0.8) | Accuracy | n = 308/352 (87.5%) |