Table 2.
Predictive performance of EGFR status and EGFR mutated subtypes using three methods in the training and testing cohorts.
| Methods | Cohorts | AUC(95%CI) | Accuracy(%) | Sensitivity(%) | Specificity(%) |
|---|---|---|---|---|---|
| EGDFR Mutation Status | |||||
| DL | Training | 0.880(0.871-0.892) | 0.805(0.797-0.815) | 0.832(0.820-0.845) | 0.783(0.769-0.798) |
| Testing | 0.842(0.825-0.855) | 0.763(0.750-0.777) | 0.797(0.777-0.818) | 0.769(0.746-0.789) | |
| Radiomics | Training | 0.838(0.827-0.850) | 0.769(0.758-0.779) | 0.794(0.780-0.812) | 0.760(0.743-0.771) |
| Testing | 0.805(0.789-0.827) | 0.735(0.720-0.755) | 0.768(0.748-0.793) | 0.716(0.696-0.734) | |
| Joint | Training | 0.919(0.914-0.924) | 0.840(0.831-0.850) | 0.839(0.829-0.852) | 0.831(0.820-0.844) |
| Testing | 0.895(0.883-0.907) | 0.819(0.803-0.835) | 0.791(0.765-0.816) | 0.850(0.834-0.870) | |
| EGFR Subtypes | |||||
| DL | Training | 0.842(0.828-0.855) | 0.753(0.740-0.769) | 0.716(0.696-0.739) | 0.853(0.836-0.873) |
| Testing | 0.805(0.779-0.829) | 0.732(0.707-0.755) | 0.707(0.676-0.742) | 0.815(0.787-0.849) | |
| Radiomics | Training | 0.809(0.791-0.829) | 0.725(0.708-0.743) | 0.672(0.647-0.702) | 0.848(0.830-0.870) |
| Testing | 0.767(0.735-0.791) | 0.697(0.663-0.728) | 0.705(0.670-0.745) | 0.742(0.712-0.773) | |
| Joint | Training | 0.873(0.860-0.884) | 0.790(0.776-0.804) | 0.758(0.739-0.778) | 0.862(0.844-0.881) |
| Testing | 0.841(0.818-0.864) | 0.767(0.746-0.790) | 0.767(0.732-0.798) | 0.827(0.803-0.858) | |
EGFR, epidermal growth factor receptor; PD-L1, programmed death ligand-1; DL model, deep learning model.