Machine Learning–Based Lung Cancer Risk Prediction for Smokers and Non-Smokers Using National N3C Data

Technical Summary: Machine Learning–Based Lung Cancer Risk Prediction for Smokers and Non-Smokers Using National N3C Data

Problem Statement
Lung cancer remains the leading cause of cancer mortality in the United States, where early detection significantly improves survival rates. Current screening guidelines from the USPSTF and ACS rely heavily on self-reported smoking history (pack-years) to determine eligibility. However, these guidelines face two critical limitations: (1) accurate smoking history is often difficult to obtain in primary care, particularly in rural settings with limited resources or incomplete records; and (2) a growing proportion of lung cancer cases occur in individuals with minimal or no smoking history, a population currently excluded from standard screening criteria. Existing prediction models often lack validation for non-smokers or fail to utilize routinely available clinical data that could support more inclusive risk stratification without relying solely on detailed smoking histories.

Methodology
The study utilized the National COVID Cohort Collaborative (N3C) Data Enclave, a harmonized electronic medical record (EMR) repository aggregating data from over 75 health systems across the U.S. The study population was derived from patients with at least two health records between January 1, 2018, and December 31, 2024. Individuals with suspected/confirmed COVID-19 or those who died during the study period were excluded to minimize confounding.

• Cohort Construction: Lung cancer cases were identified using OMOP-based concept sets. A 1:5 matching strategy was employed to select controls based on sex, race/ethnicity, age (±3 years), smoking status, and contributing site. The final analytical cohort included 97,036 lung cancer cases and 451,409 matched controls (Total N = 548,445).
• Stratification: The cohort was stratified into two distinct groups: smokers (n = 203,262) and non-smokers (n = 345,112).
• Predictors: The models utilized 85 routinely available clinical indicators (e.g., diagnoses, procedures, lab abnormalities) derived from the EMR. Indicators with >80% missing values were excluded.
• Modeling Approach: Separate Extreme Gradient Boosting (XGBoost) models were developed for smokers and non-smokers. Data were split 80/20 for training and testing.
• Evaluation: Performance was assessed using precision, recall, F1-score, and the Area Under the Receiver Operating Characteristic Curve (AUC). Model interpretability was achieved using SHAP (SHapley Additive exPlanations) values to rank feature importance and visualize risk contributions.

Key Results
Both models demonstrated good discriminatory performance on the testing sets:

• Smokers: Achieved an AUC of 0.78. The most influential predictors were pulmonary and smoking-related factors, specifically Chronic Obstructive Pulmonary Disease (COPD), hemoptysis, platelet abnormalities, and family history of lung cancer.
• Non-Smokers: Achieved a higher AUC of 0.83. The risk profile for this group reflected a broader interplay of systemic, metabolic, and demographic factors. The top predictors included urine protein abnormalities, substance use disorder, age, and various laboratory abnormalities (e.g., alkaline, platelet, LDL/HDL cholesterol).

SHAP analysis revealed distinct risk mechanisms: lung cancer risk in smokers was driven primarily by direct pulmonary injury, whereas risk in non-smokers was shaped by systemic health status, metabolic conditions, and demographic variables.

Key Contributions

1. Stratified Modeling: The study developed separate, tailored risk prediction models for smokers and non-smokers using a standardized national EMR cohort, addressing the heterogeneity in risk profiles between these populations.
2. Clinical Feasibility: The models rely exclusively on routinely available clinical information captured in EMRs, making them potentially integrable into primary care workflows without requiring complex genetic testing or detailed smoking histories that are often missing in rural settings.
3. Resource Utilization: This work represents one of the first applications of the N3C national EMR dataset to build lung cancer risk prediction models, demonstrating the utility of this harmonized resource for population-level early cancer detection research.

Significance and Claims
The authors position this work as a supplementary risk-stratification tool designed to assist primary care clinicians. The models aim to identify higher-risk individuals, particularly non-smokers who are currently excluded from existing screening guidelines, thereby facilitating earlier detection opportunities. The study emphasizes that these models are not intended for direct clinical decision-making or treatment guidance but rather to support risk assessment and inform future updates to screening strategies. The authors acknowledge limitations, including the lack of longitudinal biomarker data (e.g., FEV1, CA19-9) due to missingness in the N3C dataset and the absence of chest imaging data, which precluded comparison with deep learning models or existing clinical risk scores like PLCOm2012. The study concludes that while the models show promise, further refinement of calibration and the establishment of clinical workflows are necessary before implementation.