Table 2.
Model performance across immobilization methods (Isolation Forest, N = 7,539).
| Immobilization method | N | AUC (95% CI) | FDR-adjusted P | ΔAUC vs Overall | Remark |
|---|---|---|---|---|---|
| Neck-Thorax Mask + Vacuum Cushion | 41 | 1.000 (1.000 – 1.000) | – | 0.04 | Best performance; ideal stability |
| Neck-Thorax Mask + Thermoplastic Pad | 1,131 | 0.995 (0.990 – 1.000) | 0.3 | 0.03 | Excellent stability; consistent with SRT dual-layer mask |
| SRT Dual-Layer Mask | 388 | 0.994 (0.987 – 1.000) | 0.32 | 0.03 | Excellent stability and reproducibility |
| Head-Neck-Shoulder Mask + Shaping Pad | 2,431 | 0.971 (0.968 – 0.974) | 0.41 | 0.01 | Mainstream fixation; consistent results |
| Neck-Thorax Mask + Shaping Pad | 237 | 0.971 (0.963 – 0.979) | 0.45 | 0.01 | Comparable to head-neck-shoulder pad |
| SBRT Fixation Frame | 51 | 0.952 (0.931 – 0.972) | 0.18 | –0.01 | Small sample; high precision |
| Head-Neck-Shoulder Mask | 293 | 0.944 (0.933 – 0.956) | 0.22 | –0.02 | Stable but less constrained |
| Head-Neck-Shoulder Mask + Vacuum Cushion | 369 | 0.943 (0.934 – 0.952) | 0.2 | –0.02 | Stable hybrid fixation |
| Head-Neck-Shoulder Mask + Thermoplastic Pad | 164 | 0.933 (0.921 – 0.945) | 0.26 | –0.03 | Moderate variability |
| Vacuum Cushion | 2,225 | 0.930 (0.927 – 0.933) | 0.33 | –0.03 | Less restrictive; robust overall |
| U-Mask | 173 | 0.925 (0.912 – 0.938) | 0.29 | –0.04 | Acceptable performance; higher variability |
| Abdominal Mask + Shaping Pad | 30 | 0.918 (0.900 – 0.936) | 0.37 | –0.05 | Limited sample; larger motion range |
| Abdominal Mask + Vacuum Cushion | 6 | 0.912 (0.872 – 0.952) | 0.44 | –0.05 | Minimal sample; highest variability |
| Total | 7 539 | 0.962 (0.959 – 0.965) | – | – | Overall average performance |
Overall mean AUC (Isolation Forest, full dataset), 0.962 (0.959 – 0.965). FDR-adjusted P values derived from Kruskal–Wallis tests with Benjamini–Hochberg correction. Positive ΔAUC values indicate higher discrimination relative to the overall model; negative values indicate marginally lower performance. All fixation methods maintained AUC ≥ 0.92, indicating robust generalization across clinical conditions.Subgroups with N < 50 were retained for completeness but should be interpreted as underpowered, and their near-perfect AUC estimates are likely influenced by sampling variability rather than true performance.