Model predictions by sample grade—confusion matrix for the random forest model trained on 4 fixed variablesa
| Predicted | Reference | |||
|---|---|---|---|---|
| Normal | Lower | Higher | Total | |
| Conventional plus advanced | ||||
| Normal | 53 | 1 | 0 | 54 |
| Lower | 2 | 41 | 1 | 43 |
| Higher | 0 | 0 | 6 | 7 |
| Total | n = 55 | n = 42 | n = 7 | 104 |
| Conventional only | ||||
| Normal | 52 | 5 | 0 | 57 |
| Lower | 3 | 36 | 3 | 42 |
| Higher | 0 | 1 | 4 | 5 |
| Total | n = 55 | n = 42 | n = 7 | 104 |
↵a A perfect classifier would have all entries along the main diagonal. The random forest using conventional-plus-advanced imaging had 96.2% (100/104) overall accuracy with κ = 0.930, whereas the random forest using only conventional imaging had 88.5% (92/104) accuracy with κ = 0.788. The conventional-only classifier also misclassified 43% (3/7) of high-grade samples.