Table 2.
Internal validation of trained AI model with endoscopic images and videos
| Prediction target | Image–based mean, (95% CI) |
Video-based mean, (95% CI) |
P value |
|---|---|---|---|
| Undifferentiated histology | |||
| Accuracy (%) | 91.9 (91.0–92.9) | 89.7 (86.5–92.9) | 0.209 |
| Sensitivity (%) | 85.3 (82.6–88.0) | 83.4 (72.3–94.5) | 0.841 |
| Specificity (%) | 94.1 (92.1–96.0) | 92.1 (88.7–95.5) | 0.222 |
| PPV (%) | 81.1 (72.5–89.8) | 79.4 (62.4–96.5) | 0.690 |
| NPV (%) | 95.3 (94.0–96.7) | 86.9 (76.8–97.0) | 0.151 |
| Submucosal invasion | |||
| Accuracy (%) | 88.4 (87.8–89.0) | 88.0 (83.8–92.3) | 0.834 |
| Sensitivity (%) | 82.4 (76.6–88.1) | 80.4 (64.7–96.0) | 0.996 |
| Specificity (%) | 90.2 (87.8–92.6) | 91.1 (86.4–95.8) | 0.996 |
| PPV (%) | 72.6 (67.6–78.1) | 71.7 (67.9–85.5) | 0.841 |
| NPV (%) | 94.3 (92.7–95.8) | 93.1 (89.2–97.1) | 0.841 |
| Lymphovascular invasion | |||
| Accuracy (%) | 84.7 (79.7–89.8) | 87.9 (80.7–95.0) | 0.310 |
| Sensitivity (%) | 24.2 (15.6–32.8) | 20.0 (11.5–28.5) | 0.203 |
| Specificity (%) | 96.2 (94.9–97.5) | 97.0 (94.3–99.7) | 0.590 |
| PPV (%) | 53.6 (44.3–62.9) | 44.3 (32.5–56.1) | 0.537 |
| NPV (%) | 87.0 (81.9–92.1) | 90.2 (86.4–94.0) | 0.398 |
| Lymph node metastasis | |||
| Accuracy (%) | 86.8 (85.3–88.2) | 92.7 (87.7—97.7) | 0.008 |
| Sensitivity (%) | 27.4 (20.1–34.7) | 16.7 (4.1–22.6) | 0.085 |
| Specificity (%) | 94.0 (90.2–97.9) | 96.5 (89.3–100) | 0.151 |
| PPV (%) | 37.3 (20.5–54.2) | 27.0 (16.3–37.7) | 0.672 |
| NPV (%) | 91.4 (88.8–94.1) | 95.7 (91.1–100) | 0.151 |
AI, artificial intelligence; PPV, positive prediction value; NPV, negative prediction value; CI, confidence interval