Table 3.
C-statistics and 95% confidence intervals (CI) (over 2000 iterations) for fivefold cross-validated Cox proportional hazard models built exclusively from imaging-based features decomposed with PCA (with no added clinical covariates) to predict survival, and log-rank p values for Kaplan–Meier curve separation. ComBat (Manufacturer) indicates data was harmonized by manufacturer with ComBat.
fivefold CV c-statistic | 95% CI | Log-rank p value | |
---|---|---|---|
Lung3/CAPTK | |||
Original | 0.59 | [0.53, 0.64] | 0.0004 |
ComBat (Manufacturer) | 0.62 | [0.56, 0.67] | 0.014 |
OPNested | 0.62 | [0.57, 0.67] | 0.0021 |
OPNested + GMM | 0.50 | [0.43, 0.56] | 0.027 |
OPNested-GMM | 0.59 | [0.53, 0.64] | 0.0011 |
Lung3/PyRadiomics | |||
Original | 0.62 | [0.57, 0.67] | 0.061 |
ComBat (Manufacturer) | 0.64 | [0.59, 0.68] | 0.0004 |
OPNested | 0.65 | [0.60, 0.60] | 0.0032 |
OPNested + GMM | 0.61 | [0.55, 0.66] | 0.0036 |
OPNested-GMM | 0.65 | [0.60, 0.69] | 0.0019 |
Radiogenomics/CAPTK | |||
Original | 0.58 | [0.52, 0.62] | 0.015 |
ComBat (Manufacturer) | 0.58 | [0.51, 0.64] | 0.036 |
OPNested | 0.54 | [0.46, 0.61] | 0.019 |
OPNested + GMM | 0.56 | [0.49, 0.61] | 0.012 |
OPNested-GMM | 0.55 | [0.47, 0.61] | 0.029 |
Radiogenomics/PyRadiomics | |||
Original | 0.63 | [0.59, 0.67] | 0.082 |
ComBat (Manufacturer) | 0.59 | [0.54, 0.63] | 0.13 |
OPNested | 0.59 | [0.54, 0.63] | 0.12 |
OPNested + GMM | 0.58 | [0.52, 0.63] | 0.41 |
OPNested-GMM | 0.59 | [0.54, 0.64] | 0.055 |