Machine Learning for Early Lung Cancer Identification Using Routine Clinical and Laboratory Data.

Area of Science:

• Oncology
• Medical Informatics
• Machine Learning

Background:

• Most lung cancers are diagnosed at advanced stages, limiting treatment efficacy.
• Early detection through presymptomatic identification of high-risk individuals is crucial for improving patient outcomes.
• Routine clinical and laboratory data hold potential for predicting future lung cancer diagnoses.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for predicting future non-small cell lung cancer (NSCLC) diagnosis.
• To compare the predictive accuracy of the novel ML model against a modified Prostate, Lung, Colorectal and Ovarian Cancer Screening Trial risk model (mPLCOm2012).

Main Methods:

• Assembled data from 6,505 NSCLC cases and 189,597 control subjects.
• Developed a novel machine learning model using routine clinical and laboratory data.
• Compared model performance using Area Under the Receiver Operating Characteristic Curve (AUC), sensitivity, and diagnostic odds ratio (OR) at 95% specificity.

Main Results:

• The ML model demonstrated superior accuracy in predicting NSCLC 9-12 months pre-diagnosis among ever-smokers compared to mPLCOm2012 (AUC 0.86 vs. 0.79).
• At 95% specificity, the ML model achieved 40.1% sensitivity and an OR of 12.3, outperforming mPLCOm2012 (27.9% sensitivity, OR 7.4).
• Key predictors included known risk factors and novel variables like white blood cell and platelet counts.

Conclusions:

• The developed machine learning model significantly outperforms standard lung cancer screening criteria and the mPLCOm2012 for early NSCLC detection.
• This ML model shows promise for enhancing early lung cancer diagnosis, potentially reducing mortality through timely intervention.
• Incorporating novel predictors like blood cell counts can improve predictive accuracy for lung cancer.