Lung cancer prediction using machine learning on data from a symptom e-questionnaire for never smokers, formers smokers and current smokers

Abstract

Purpose The aim of the present study was to investigate the predictive ability for lung cancer of symptoms reported in an adaptive e-questionnaire, separately for never smokers, former smokers, and current smokers. Patients and methods Consecutive patients referred for suspected lung cancer were recruited between September 2014 and November 2015 from the lung clinic at the Karolinska University Hospital, Stockholm, Sweden. A total of 504 patients were later diagnosed with lung cancer (n = 310) or no cancer (n = 194). All participants answered an adaptive e-questionnaire with a maximum of 342 items, covering background variables and symptoms/sensations suspected to be associated with lung cancer. Stochastic gradient boosting, stratified on smoking status, was used to train and test a model for predicting the presence of lung cancer. Results Among never smokers, 17 predictors contributed to predicting lung cancer with 82% of the patients being correctly classified, compared with 26 predictors with an accuracy of 77% among current smokers and 36 predictors with an accuracy of 63% among former smokers. Age, sex, and education level were the most important predictors in all models. Conclusion Methods or tools to assess the likelihood of lung cancer based on smoking status and to prioritize investigative and treatment measures among all patients seeking care with diffuse symptoms are much needed. Our study presents risk assessment models for patients with different smoking status that may be developed into clinical risk assessment tools that can help clinicians in assessing a patient’s risk of having lung cancer.