Session Item

Recurrence occurs in up to 36% of patients treated with radiotherapy for NSCLC. High-quality evidence to provide specific recommendations on the nature of post-treatment surveillance is lacking. Risk-stratification models are required to determine optimal follow-up. We present a radiomics model to predict recurrence and recurrence-free survival (RFS) 2 years post-treatment, using routinely available radiotherapy planning CTs and the gross tumour volume (GTV), contoured for radiotherapy purposes as the region of interest for feature extraction.

A retrospective multi-centre study of patients receiving stereotactic or conventional (chemo)radiotherapy for stage I-III NSCLC was undertaken. Cases with a GTV encompassing the primary tumour were included from 5 hospitals. Cases from 4 hospitals were divided into training and validation sets, stratified by outcome, with the 5^th hospital providing an external test set. Radiotherapy planning CTs were pre-processed and features extracted from GTVs using TexLAB 2.0. Time to recurrence/RFS data were binarized (yes/no) at 2 years from the first fraction of radiotherapy for classification purposes. 

We explored a combination of 9 feature reduction techniques with 11 machine learning classifiers, producing risk-stratification models for recurrence and RFS. The model with the highest validation set AUC was selected and deployed on the external test set. Models were compared with 10-fold cross validation and bench-marked against TNM stage. Youden index, calculated from validation set ROC curves, was used to define high and low risk groups. Kaplan-Meier curves were produced.

509 patients were included, training = 302, validation = 75, test = 132. Median follow-up was 762 days. Train-validation and external test set mean age was 74 and 71 respectively. Recurrence and RFS rates at 2 years were 36.3% vs 30.3%, and 48% vs 43.9% respectively.

 For recurrence Principal Component Analysis followed by an ensemble of Partial Least Squares (PLS), K-Nearest Neighbours and Elastic-net was the best performing model. For RFS, Pearson correlation with PLS was the best performing model.

 The respective validation and test set AUCs and 95% confidence intervals are shown in Table 1. Our model had superior AUC to the TNM model in both validation and test sets for predicting RFS. For recurrence, AUC was superior to the TNM stage model in the validation set and similar in the test set. Kaplan Meier curves showed marked separation with significant log-rank tests (Figure 1).

We present validated and externally tested recurrence risk-stratification models that use routinely available radiotherapy data. Such models could be integrated into the radiotherapy workflow to inform personalised surveillance at the point of treatment for patients with NSCLC.