Session Item

Surveillance is universally recommended for NSCLC patients treated with curative-intent radiotherapy. High-quality evidence to inform optimal surveillance strategies is lacking. Machine learning (ML) demonstrates promise in accurate outcome prediction for a variety of health conditions. The purpose of this study was to utilise readily available patient, tumour, and treatment data to develop, validate and externally test machine learning models for predicting overall survival (OS). Such models may allow for personalised follow-up resulting in potentially earlier detection of recurrence for high-risk patients or avoidance of unnecessary hospital visits for low-risk patients. This would have implications for patient care and healthcare resource use globally. 

A retrospective, multicentre study of patients receiving curative-intent radiotherapy for NSCLC was undertaken. A total of 657 patients from 5 hospitals were eligible for inclusion. Data pre-processing derived 34 features (variables) for predictive modelling. Combinations of 8 feature reduction methods and 10 ML classification algorithms were compared, producing a risk-stratification model for predicting recurrence OS at 2 years from the first fraction of radiotherapy (Figure 1). Models were compared with 10-fold cross validation and an external test set and benchmarked against TNM-stage and performance status. Youden Index was derived from validation set ROC curves to distinguish high and low risk groups and Kaplan-Meier analyses performed.

Median follow-up time was 852 days. Parameters were well matched across train-validation and external test sets: Mean age was 73 and 71 respectively. OS rate at 2 years was 54% vs 47% across train-validation and external test sets respectively. The best feature reduction and ML combination was Kendall’s rank correlation followed by an ensemble of Mixture Discriminant Analysis, XG Boost and a single hidden layer neural net. The respective validation and test set AUCs are shown in Table 1. Our model AUC values were superior to TNM stage and performance status in predicting 2-year OS. Kaplan-Meier curves show good separation with significant log-rank test in the external test set.

We demonstrate that our model is superior to TNM or Performance status and believe that our methodology can be replicated across health systems using local clinical datasets, without complex imaging and computational requirements, to benefit surveillance stratification for patients following radical radiotherapy for NSCLC globally.

Our models are built on routinely available clinical data and set the stage for future clinical trials entailing quantitative personalised risk-stratification and surveillance following curative-intent radiotherapy for NSCLC.