What question did this study set out to answer?

February 28, 2026Open Access

Interpretable machine learning models using peripheral blood biomarkers for diagnosis and prognosis of glottic laryngeal squamous cell carcinoma

Key Points

The study aims to create interpretable machine learning models using peripheral blood biomarkers for diagnosing and predicting outcomes in glottic LSCC.
Enrolled 124 LSCC patients, 124 benign vocal cord lesion patients, and 124 healthy controls.
Developed Random Forest and XGBoost models using preoperative blood parameters.
Assessed clinical utility with Decision Curve Analysis (DCA) and interpretability with SHAP.
Conducted multivariate Cox analysis for prognostic evaluation.
Performed sensitivity analyses to adjust for demographic confounding factors.
XGBoost model achieved AUC of 0.93 for distinguishing malignant from benign lesions.
SHAP identified INR, fibrinogen, thrombin time, NMR, and LMR as key predictive features.
NPR correlated with perineural invasion (r=0.625, P<0.001).
NLR was an independent predictor of disease-free survival (HR=2.11, P<0.001).
SII was an independent overall survival predictor (HR=1.005, P<0.001).

Abstract

To develop and validate interpretable machine learning models using routine peripheral blood biomarkers for the differential diagnosis and prognostic stratification of glottic laryngeal squamous cell carcinoma (LSCC). In this retrospective cohort study, we enrolled 124 patients with glottic LSCC, 124 patients with benign vocal cord lesions, and 124 healthy male controls. Preoperative peripheral blood parameters were utilized to develop Random Forest and XGBoost models. The clinical utility and interpretability of the models were assessed using Decision Curve Analysis (DCA) and SHapley Additive exPlanations (SHAP), respectively. The prognostic value of biomarkers was evaluated using multivariate Cox proportional hazards regression. Stratified sensitivity analyses were performed to control for demographic confounders. The XGBoost model demonstrated excellent discrimination in distinguishing malignant from benign lesions, achieving an area under the curve (AUC) of 0.93 (95% CI: 0.86–0.99) on the independent test set. SHAP analysis identified the international normalized ratio (INR), fibrinogen, thrombin time, neutrophil-to-monocyte ratio (NMR), and lymphocyte-to-monocyte ratio (LMR) as the most influential predictive features. Sensitivity analysis confirmed that these markers remained significantly different between groups. The neutrophil-to-platelet ratio (NPR) was significantly correlated with perineural invasion (r = 0.625, P < 0.001). In multivariate Cox analysis, the neutrophil-to-lymphocyte ratio (NLR) was an independent predictor for disease-free survival (HR = 2.11, P < 0.001), and the systemic immune-inflammation index (SII) was an independent predictor for overall survival (HR = 1.005, P < 0.001). Interpretable machine learning models integrating peripheral blood biomarkers of inflammation, coagulation, and nutrition provide a robust, non-invasive tool for the preoperative diagnosis and prognostic assessment of glottic LSCC. These models have the potential to optimize clinical decision-making and personalize patient management.

Interpretable machine learning models using peripheral blood biomarkers for diagnosis and prognosis of glottic laryngeal squamous cell carcinoma

Key Points

Abstract

Cite This Study