Using machine learning algorithms based on laboratory indicators to establish a diagnostic model for lung cancer.

Area of Science:

• Medical Informatics
• Machine Learning in Oncology
• Diagnostic Modeling

Background:

• Lung cancer is a leading cause of mortality, with current diagnostics being invasive.
• There's a need for non-invasive methods for early lung cancer detection and monitoring.
• Pulmonary nodules require accurate differentiation between benign and malignant conditions.

Purpose of the Study:

• To compare machine learning algorithms for distinguishing benign pulmonary nodules from lung cancer.
• To identify the optimal model for early and advanced-stage lung cancer detection.
• To develop a non-invasive predictive tool for lung cancer diagnosis.

Main Methods:

• Collected clinical data from 1,238 patients with pulmonary nodules and 250 healthy controls.
• Developed and evaluated five machine learning algorithms, including eXtreme Gradient Boosting (XGBoost).
• Assessed model performance using AUC, accuracy, sensitivity, precision, and SHAP for interpretability.

Main Results:

• XGBoost demonstrated superior performance in distinguishing between healthy, benign, and lung cancer groups.
• Achieved high AUC values, including 0.999 (healthy vs. benign) and 0.970 (healthy vs. advanced lung cancer).
• SHAP analysis identified key clinical features influencing model predictions.

Conclusions:

• XGBoost models show strong potential for non-invasive lung cancer diagnosis and staging.
• This approach can aid clinicians in early identification and risk stratification of pulmonary nodules.
• The predictive models may contribute to reducing lung cancer-related mortality through timely intervention.