Machine Learning to Support Treatment Modality or Early-Stage NSCLC: A Framework for Clinical Application

Surgical resection (SR) and radiotherapy (RT) are the current standard treatment modalities for stage I non- small cell lung cancer (NSCLC). We aimed to develop a machine learning model to support treatment modality selection during multidisciplinary tumor board (MTB) discussions. The baseline variables of 475 subjects suspected of NSCLC and discussed during a MTB at the University Hospitals of Leuven (2015-2019) were used to develop a random forest model that predicted the probability for SR or RT. It was trained using 332 subjects (70%) and validated on the remaining 143 subjects (30%). We defined a gray zone on the probabilities to reject uncertain predictions that require further human review. ECOG, FEV1, DLCO, Charlson comorbidity index, age at diagnosis, number of comorbidities, presence of biopsy and tumor size were the model-selected top input parameters. The model had an accuracy of 79.7%, sensitivity of 80.7% and specificity of 78.2% for predicting SR in the test set. When only considering subjects outside the gray rejection zone (40%), positive predictive value was 90% and negative predictive value was 100% . The model demonstrated high predictive performance when used with a reject option while still correctly assessing a significant proportion of the population. The model may be used as clinical decision support to reduce the caseload when selecting treatment modalities for NSCLC by MTBs. After a long investigation to make sure of the reliability of the pre-trained model, a software application was developed based on the RF pre-trained model in python for the easy access of the health care professionals and specialists for an everyday use in the future.