Development, deployment, and feature interpretability of a three-class prediction model for pulmonary diseases

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning for Medical Imaging
• Pulmonary Disease Diagnostics

Background:

• Accurate classification of pulmonary diseases is crucial for effective treatment.
• Distinguishing between non-small cell lung cancer (NSCLC), granulomatous inflammation, and benign tumors presents a diagnostic challenge.
• Machine learning offers potential for improving diagnostic accuracy in pulmonary medicine.

Purpose of the Study:

• To develop and evaluate a high-performance machine learning model for the classification of pulmonary diseases.
• To compare the efficacy of XGBoost and Random Forest (RF) algorithms in this classification task.
• To interpret the predictive features identified by the models.

Main Methods:

• Retrospective analysis of multicenter clinical and imaging data from 3030 patients.
• Feature selection using the Boruta algorithm.
• Development and validation of Random Forest (RF) and XGBoost models.
• Performance assessment using receiver operating characteristic (ROC) analysis, Obuchowski indices, calibration curves, and decision curve analysis (DCA).

Main Results:

• XGBoost demonstrated superior performance over RF in both the internal test set (Obuchowski index 0.8282 vs. 0.7193) and the external validation set (0.8074 vs. 0.7932).
• XGBoost achieved higher accuracy (0.81 in test set, 0.79 in validation set).
• Decision Curve Analysis indicated XGBoost provided a greater net clinical benefit.

Conclusions:

• The XGBoost model significantly outperforms the Random Forest model for the three-class classification of lung diseases (NSCLC, granulomatous inflammation, benign tumors).
• XGBoost exhibits strong potential for clinical application in diagnosing pulmonary conditions.
• The developed XGBoost model can be deployed on a web-based platform for clinician use.