Explainable Machine Learning to Predict Treatment Response in Advanced Non-Small Cell Lung Cancer

Area of Science:

• Oncology
• Biomarker Discovery
• Machine Learning in Medicine

Background:

• Immune checkpoint inhibitors (ICIs) are effective cancer treatments.
• PD-L1 expression is a standard biomarker for immuno-oncology (IO) monotherapy in advanced non-small cell lung cancer (NSCLC).
• Predictive biomarkers are crucial for optimizing cancer therapy selection.

Purpose of the Study:

• To evaluate if a machine learning (ML) algorithm can outperform PD-L1 as a predictive biomarker for first-line therapy in advanced NSCLC.
• To assess the predictive performance of ML algorithms for 12-month progression-free survival (PFS) and overall survival (OS).

Main Methods:

• Utilized a deidentified electronic health record database of 38,048 advanced NSCLC patients.
• Trained binary prediction algorithms to predict 12-month PFS and 12-month OS.
• Evaluated algorithms using AUC on a test set and compared survival outcomes between low-risk and high-risk patient groups via Kaplan-Meier curves and Cox models.

Main Results:

• ML algorithms achieved AUCs of 0.701 for 12-month PFS and 0.718 for 12-month OS.
• Low-risk patients identified by ML had significantly lower 12-month disease progression (HR=0.47) and mortality (HR=0.31) compared to high-risk patients.
• ML-identified low-risk patients on IO monotherapy showed reduced progression (HR=0.53) and mortality (HR=0.30).

Conclusions:

• ML algorithms provide more accurate prediction of first-line therapy response in advanced NSCLC than PD-L1 alone.
• ML holds potential to improve clinical decision-making in oncology beyond single biomarker assessments.