Table 3.
Results of subgroup analysis of sensitivity and specificity by fracture site and machine learning type
| Subgroup | Training dataset | Validation dataset | ||||
| N | Sensitivity (95% CI) | Specificity (95% CI) | N | Sensitivity (95% CI) | Specificity (95% CI) | |
| Fracture site | ||||||
| Vertebral fracture | 10 | 0.73 (0.61, 0.82) | 0.91 (0.86, 0.95) | 3 | 0.87 (0.70, 0.95) | 0.97 (0.94, 0.98) |
| Hip fracture | 13 | 0.90 (0.82, 0.94） | 0.82 (0.75, 0.88) | 5 | 0.84 (0.77, 0.89) | 0.85 (0.80, 0.89) |
| Multi-site fracture | 18 | 0.71 (0.59, 0.81) | 0.72 (0.60, 0.81) | 8 | 0.66 (0.61, 0.70) | 0.69 (0.53, 0.81) |
| Model type | ||||||
| LR | 17 | 0.70 (0.63, 0.77) | 0.73 (0.67, 0.79) | 4 | 0.66 (0.55, 0.75) | 0.65 (0.50, 0.77) |
| ANN | 4 | 0.91 (0.70, 0.98) | 0.93 (0.75, 0.98) | 3 | 0.78 (0.71, 0.83) | 0.85 (0.71, 0.93) |
| CNN | 3 | 0.83 (0.81, 0.84) | 0.91 (0.79, 0.96) | 1 | 0.98 | 0.95 |
| RF | 1 | 0.84 | 0.91 | 1 | 0.70 | 0.46 |
| SVM | 6 | 0.81 (0.63, 0.92) | 0.63 (0.13, 0.95) | 3 | 0.79 (0.72, 0.85) | 0.89 (0.79, 0.94) |
| DT | 2 | 0.97 (0.53, 1.00) | 0.70 (0.67, 0.73) | – | ||
| NB | 2 | 0.63 (0.13, 0.95) | 0.76 (0.70, 0.81） | – | ||
| kNN | 2 | 0.95 (0.39, 1.00） | 0.80 (0.77, 0.83） | 1 | 0.81 | 0.79 |
| Survival model | 1 | 0.81 | 0.52 | – | ||
| Boosted tree | 1 | 0.59 | 0.67 | 1 | 0.70 | 0.95 |
| Other DL | 2 | 0.81 (0.72, 0.87) | 0.96 (0.93, 0.98) | 2 | 0.83 (0.74, 0.90) | 0.95 (0.92, 0.97) |
| Overall | 41 | 0.79 (0.72, 0.84) | 0.81 (0.75, 0.86) | 16 | 0.76 (0.80, 0.81) | 0.83 (0.72, 0.90) |
ANN, artificial neural network; CNN, convolutional neural network; DL, deep learnimg model; DT, decision tree; kNN, k-nearest neighbour; LR, logistic regression; NB, Naive Bayes; RF, random forests; SVM, support vector machine.