Establishment and validation of multiclassification prediction models for pulmonary nodules based on machine learning

Area of Science:

• Oncology
• Medical Informatics
• Radiology

Background:

• Lung cancer is a leading global cause of cancer mortality.
• Accurate risk stratification of pulmonary nodules (PNs) is crucial for timely diagnosis and treatment.
• Existing models for PN malignancy prediction require further validation and improvement.

Purpose of the Study:

• To develop and evaluate novel machine learning (ML) based multiclassification prediction models for PNs.
• To compare the performance of developed ML models against three established prediction models (Mayo, PKUPH, and Brock).

Main Methods:

• A dataset of 914 patients with PNs from four institutions was utilized.
• Clinical, radiologic, and laboratory features were collected and categorized into benign (BL), precursor (PL), and malignant lesion (ML) groups.
• Models including logistic regression (LR), decision tree (DT), random forest (RF), and support vector machine (SVM) were trained and validated.

Main Results:

• The random forest (RF) model achieved high AUC values for predicting MLs (0.80), PLs (0.90), and BLs (0.75) in the internal test set.
• For external validation, the RF model showed a weighted average AUC of 0.71, outperforming the Mayo (0.68), PKUPH (0.64), and Brock (0.57) models.
• The RF model demonstrated strong predictive performance across different lesion types in external validation.

Conclusions:

• The developed RF model exhibits superior predictive performance for PNs compared to the three validated published models.
• This ML-based approach offers a potential new noninvasive method for the risk assessment of pulmonary nodules.
• Further research and clinical implementation of the RF model could improve lung cancer diagnosis and patient outcomes.