Machine learning and discriminant analysis model for predicting benign and malignant pulmonary nodules.

Area of Science:

• Medical Imaging
• Machine Learning
• Oncology

Background:

• Pulmonary nodules (PNs) are early indicators of lung cancer, but distinguishing benign PNs (BPNs) from malignant PNs (MPNs) remains challenging.
• Current screening methods yield a high false positive rate, with over 90% of detected PNs being benign.

Purpose of the Study:

• To develop and validate a predictive model for accurately differentiating between benign and malignant pulmonary nodules.
• To improve the diagnostic accuracy in distinguishing MPNs from BPNs.

Main Methods:

• A case-control study included 5197 patients, with data randomly assigned to training, validation, and test sets.
• Three machine learning algorithms (Random Forests, Gradient Boosting Machine, XGBoost) were employed to screen metrics and construct predictive models.
• The best-performing model was selected, internally validated using 10-fold cross-validation, and compared against existing PKUPH and Block models.

Main Results:

• The detection rate of PNs in the physical examination population was 21.57%, showing an upward trend.
• The developed GMU_D model demonstrated excellent discriminative performance (AUC = 0.866), outperforming the PKUPH (AUC = 0.559) and Block (AUC = 0.823) models.
• Internal validation confirmed the model's excellent performance (AUC = 0.866).

Conclusions:

• A novel prediction tool, the GMU_D model, was developed and validated for distinguishing between benign and malignant pulmonary nodules.
• This model offers excellent predictive performance and differentiation capabilities, addressing current challenges in PN classification.
• The study highlights the potential of machine learning in improving the accuracy of lung cancer screening and diagnosis.