Predicting lung cancer recurrence in patients within two years of curative radiotherapy via a trained machine learning algorithm.

Objectives: Develop a trained machine learning algorithm to predict recurrent lung cancer within two years of curative radiotherapy.

Methods: 571 patients diagnosed with T1a-T3N0M0 lung cancer were treated with radiotherapy between January 6, 2001 and August 8, 2014 at the Ottawa Hospital, with patient outcomes (local, regional, or distant lung cancer recurrences) tracked until October 25, 2016. 280 patients were excluded from analysis; Exclusion criteria includes patients who received radiotherapy prior to 2003, patients who did not receive a PET scan within four months of initiation of radiotherapy, and patients who died of causes other than lung cancer during the follow-up period. Of the 291 patients, segmentation was performed on 161 PET scans including the primary pulmonary lesion, up to two additional pulmonary nodules, and any hilar or mediastinal lymph nodes. Background liver, splenic and blood pool values were assessed with spherical regions of interest (ROI). Uptake within pleural or pericardial fluid was measured as an SUVmax. On the CT scan, variables included the distance of the primary lung tumour from the mediastinum/chest wall, and the length of the interface between the primary lung tumour and the mediastinum/chest wall. Binary variables included subpleural or subpericardial fat blurring, pericardial effusions, pneumothoraxes and lesions simultaneously contacting the chest wall/mediastinum. All computations were performed using R version 3.3.2¹. Multivariate imputation was performed using the “mice” package² to predict missing variables, including predicted FEV1 (25.7% missing), pack years of smoking (10% missing), and weight loss (2.5% missing). All categorical variables were converted to binary dummy variables using the “caret” package³. Right-skewed variables were log-transformed. All numeric variables were normalized to a fraction between 0 and 1. Relevant predictors were selected by random forest feature selection using the “randomForest” package⁴; Clinical factors included in the machine learning algorithm include age, pack years of smoking, Eastern cooperative Oncology Group (ECOG) status, gender, Charlson Comorbidity Index (CCMI) and weight loss. Treatment parameters entered include biological equivalent dose (BED10), equivalent dose (EQD2) and the time interval between the PET/CT and initiation of radiotherapy. Of the 161 patients analyzed, 21% demonstrated recurrent cancer within two years. The 161 patients were randomly divided into a group of 106 patients to train the neural network, and 51 patients on which to test the predictions of the neural network. The neural network was created from the predictor variables within the training patient dataset using the “neuralnet” package⁵.

Results: The trained neural network was able to predict which patients would experience a recurrence within two years with a sensitivity of 60% and a positive predictive value of 69%. The neural network was also able to predict which patients would not recur within two years with a specificity of 88% and a negative predictive value of 82% for an overall accuracy of 79% (Table 1).

$$graphic_50683EC5-99FF-40D7-8B29-20D2254133FA$$

Conclusion: With an accuracy of 79% in predicting two-year lung cancer recurrence post-SBRT, the neural network is effective in predicting which patient's cancer will recur within two years post-SBRT. The accuracy of the neural network would likely benefit from analyzing even greater numbers of patients in the training dataset. Furthermore, adding more predictors, such as mutation status, radiomic nodule assessment, growth rates, or more detailed radiotherapy treatment volumes may improve accuracy. Machine learning strategies will likely aid in creating patient-specific treatment plans, as accuracy continues to improve. Further research is required to determine whether machine-learning guided treatment plans could improve patient outcomes, and/or reduce health-care expenditures.⁶ Research Support: None.