Seventy-one patients were included in the study. Patient demographics are presented in figure 1.
Most patients were stage IIIa (46%) or stage IIIb (51%). All completed radiotherapy, 97% receiving 66-60Gy in 30-33 fractions and 97% concurrent carboplatin and taxol.
Fifty-two patients (73.2%) patients developed recurrence at a median 6.1 months (IQR 3.3-11.5). Sites of first failure were local in 15 (21%), nodal or distant in 37 (52%). Kaplan Meier estimates for failure free survival (FFS) are given in figure 2. The 1 year FFS was 45% (95%CI 34.9-58.3) and 3 year FFS 24% (95% CI 15.2 -38).
On univariable analysis clinical stage (stage IIIc vs IIIa) was the only factor significantly associated with a higher risk of failure as shown in figure 3, though there were only 2 patients in this group. Of all radiomic features kurtosis, maximum and range had a significant effect on FFS as shown in figure 4.
The final model obtained in the multivariable analysis for clinical and radiomic features associated with any failure is given in figure 5. Kurtosis and Stage IIIc significantly influence the risk of failure. The results for stage need validation with a larger cohort as there were only 2 patients who developed failure.
Four radiomic machine learning models were developed using all radiomic features and those significant for any failure from the univariate analysis. Comparison of models for prediction of any failure is displayed in figure 6. Random forest and SVC models had the higher predictive values using all features (66.7%). Models using only 3 limited radiomic features made a minimal change in prediction accuracy across all models.
For local recurrence, non-adenocarcinoma histology was the only significant factor found on univariable analysis. There were no significantly associated radiomic features with local failure on univariable analysis so no further analysis was undertaken.
At a median follow-up of 35.5 months (range 7 – 58.3), there were 38 (53.5%) deaths. The 1-year OS was 74.5% (95%CI 64.9-85.4), 3-year OS was 42.7% (95%CI 31.5- 57.8). On univariate Cox proportional hazard analysis, increasing age, non-adenocarcinoma histology and Stage (stage IIIb and IIIc vs Stage IIIa) had a significantly negative effect on survival, as shown in figure 7.
Analysis of radiomic features correlating with survival are given in figure 8. Features significantly related to overall survival included kurtosis, maximum, range and GLCM autocorrelation. Results of the final model obtained after stepwise multivariable Cox regression are given in figure 9. Overall a model using age, Stage (stage IIIc vs stage IIIa) and radiomic features of range and GLCM autocorrelation can significantly influence risk of death.
The machine learning models using all radiomic features and 9 limited highly correlated features from univariate analysis were compared for their predictive power for survival at 3 years, as shown in figure 10.
Decision tree and logistic regression models showed the highest predictive values (84.8%) for 3-year survival when using all radiomic features. The predictive power of all models was increased using limited radiomic features, apart from logistic regression which remained high.