Comparative Analysis of Machine Learning Algorithms for Prognostic Prediction in Pulmonary Large Cell Carcinoma

Area of Science:

• Oncology
• Data Science
• Bioinformatics

Background:

• Large Cell Carcinoma (LCC) is an aggressive non-small cell lung cancer with poor prognostic assessment.
• Conventional staging models inadequately capture LCC's biological complexity.
• Machine learning offers potential for improved prognostic accuracy and personalized treatment.

Purpose of the Study:

• To develop and validate machine learning models for predicting Overall Survival (OS) in LCC patients.
• To compare the performance of machine learning models against traditional methods.
• To identify key prognostic factors influencing OS in LCC.

Main Methods:

• Retrospective study using the SEER database (1,867 LCC patients).
• Feature selection using Lasso-Cox regression and Boruta algorithm.
• Model validation with time-dependent AUC, calibration plots, decision curve analysis, and Brier scores.
• Interpretability analysis using SHAP, PDP, and RCS.

Main Results:

• Random Survival Forest (RSF) outperformed Cox regression and other ML models.
• RSF achieved high AUC values for 3-year (0.858) and 5-year (0.863) OS prediction.
• Key prognostic factors identified: tumor size, metastatic status, and treatment interventions.
• Non-linear relationships between tumor size and survival were observed.

Conclusions:

• The developed machine learning framework, especially RSF, shows strong predictive performance for LCC patient OS.
• Identified prognostic factors provide insights into LCC behavior.
• A web-based platform enhances clinical utility for risk stratification and personalized treatment planning.