Abstract
Early detection of positive pulmonary nodules (PPNs) is a critical step in reducing lung cancer mortality. This study presents a machine learning framework that utilizes targeted plasma metabolomics—specifically focusing on central carbon metabolism—to identify PPNs in high-risk populations. By combining metabolic signatures with clinical data, the researchers developed predictive models, with the Random Forest algorithm achieving high diagnostic accuracy (AUC up to 0.87). This approach offers a promising non-invasive tool for early clinical intervention.
Introduction
Lung cancer remains a leading cause of global disease burden, largely due to late-stage diagnosis. While pulmonary nodules (PNs) are common early indicators, distinguishing between "positive" nodules (those with high malignant potential) and benign nodules is a significant clinical challenge. Current screening relies heavily on Low-Dose Computed Tomography (LDCT), but there is a pressing need for molecular biomarkers to increase diagnostic precision. This research explores whether central carbon metabolism—the core pathways of cellular energy production frequently hijacked by cancer cells—can provide a "metabolic fingerprint" to accurately predict PPN risk.
Methodology
The study utilized data from the Ma'anshan Community Lung Cancer Screening Cohort (MALSC), a prospective cohort in China involving 2,248 high-risk individuals. The analytical pipeline included:
- Targeted Metabolomics: High-Performance Ion Chromatography–Mass Spectrometry (HPIC-MS) was used to measure 18 specific central carbon-related metabolites in plasma.
- Feature Selection: The Least Absolute Shrinkage and Selection Operator (LASSO) and Recursive Feature Elimination (RFE) were employed to identify the 10 most relevant features, including metabolites and clinical indicators like monocyte ratio and triglycerides.
- Machine Learning Modeling: Five algorithms were tested: Random Forest (RF), XGBoost, Support Vector Machine (SVM), Naive Bayes (NB), and Decision Trees (DT).
- Interpretability: SHapley Additive exPlanations (SHAP) were used to quantify the contribution of each feature to the model's final output, providing clinical transparency.
Key Results
The Random Forest (RF) and XGBoost models emerged as the top performers. In the primary task of PPN risk prediction within the high-risk population, the RF model achieved an AUC of 0.87 and an accuracy of 0.93.
For the more difficult task of distinguishing PPNs from negative pulmonary nodules (NPNs), the RF model maintained a strong AUC of 0.83 and an accuracy of 0.86. Notably, models trained on metabolites alone performed nearly as well as those including demographic data, suggesting that metabolic states are potent indicators of nodule status.
Interpretability and Biological Insights
To move beyond "black box" predictions, the team used SHAP analysis to reveal the biological drivers of the model. The study found that PPNs are associated with significantly higher concentrations of metabolites like succinic acid, gluconic acid, and hippuric acid, while concentrations of homogentisic acid and maleic acid were notably lower.
These findings are biologically plausible; succinic acid and fumaric acid are key components of the TCA cycle, and their importance likely reflects the cellular hypoxia and altered energy needs of developing nodules.
Conclusion and Future Work
This study demonstrates that central carbon metabolism provides a valuable metabolic signature for early lung cancer risk assessment. While the results are promising, the authors acknowledge the need for external validation in independent cohorts and long-term follow-up to confirm if predicted PPNs transition into malignant cancer. Nonetheless, these machine learning models represent a significant step toward non-invasive, high-precision screening for at-risk individuals.
Summary updated