Fig 1.
Diagram of the workflow followed in the multivariable model development phase.
After a test-retest and inter-observer study, 77 features remained for further analysis, based on a cut-off of 0.85 for the ICC analysis. Further identification of comparable features extracted from the structure merging all metastatic lymph nodes (LNmerged) to the largest (LNvolume) or most active node (LNmax), by means of an intraclass correlation (ICC) over 0.85 and ±10% limits of agreement (LoA) between measurements, was performed (further details in S1 File. Feature pre selection). In summary, 77 features of the primary tumor and 16 from the metastatic lymph nodes were entered in the model development phase.
Table 1.
Demographics and clinical information of development and validation cohorts.
Table 2.
Univariable Cox regression of clinical variables in development cohort.
Fig 2.
Pearson correlation plot for metabolic descriptors and volume of primary tumor and metastatic lymph nodes in the development dataset.
Table 3.
Distribution of common PET descriptors (maximum, peak and mean) and volume of the primary tumor and LNs.
Fig 3.
Log-linear and proportional hazards assumptions verification.
Graphically, log-linearity was verified by fitting a penalised smoothing spline on the univariable effect of each variable included in models (left graph), while proportional hazards were analysed by plotting Schoenfeld residuals versus log (time) (right graph). These included variables for LN, the (A) volume, (B) GLRLM grey level non-uniformity, (C) GLRLM short run high grey level emphasis, (D) GLCM entropy, (E) surface to volume ratio, and (F) uniformity, and (G) GLRLM short run emphasis of tumour. All variables were log (linear), except LN volume (A left), for which a logarithmic transformation was performed (A middle). All variables satisfied the proportional hazards assumption. Automatic feature selection for model 1 (based solely on primary tumor imaging features) converged to a single metric of the GLRLM group—short run emphasis, with a C-index of 0.53 (95% confidence interval [CI] = 0.49–0.58) and an external validation of 0.54. Model 2 (based on imaging features from LN) included total volume and the surface to volume ratio (shape), histogram uniformity (first order statistics), grey level non-uniformity and short run high grey level emphasis (GLRLM of the textural group), reaching a C-index of 0.62 (95% CI = 0.57–0.66) with an external validation of 0.56. Important to note that LN volume is an independent prognostic metric, with an univariable performance of 0.60 (95% CI = 0.51–0.68). Finally, model 3 selected the same feature as model 1 and four features from the LN, replacing short run high grey level emphasis–GLRLM, by entropy–GLCM, and reached a performance of 0.62 (95% CI = 0.58–0.67), and 0.59 in the external cohort. No metrics from the IVH sub-category were selected from any of the analyzed structures for the derived models. Based on an AIC test, model 3 (1854.5) was shown to be a better fit than model 2 (1857.4), which itself was already a more precise fit compared to model 1 (1876.4). In summary, the addition of nodal imaging information resulted in a better model fit, compared to a model based exclusively on features derived from the primary tumor.
Table 4.
Distribution of features included in the Cox regression model for FDG-PET-CT-based features extracted from pre-radiotherapy scans of NSCLC patients.