An Interpretable Model for Predicting Occult Lymph Node Metastasis in Clinical T1N0 Non-Small Cell Lung Cancer with a Consolidation-to-Tumour Ratio > 0.5: A Multicentre Study

Background Occult lymph node metastasis (oLNM) remains a concern in clinical T1N0 non-small cell lung cancer (NSCLC). In this study, we developed and externally validated a preoperative prediction model for patients with a consolidation-to-tumour ratio (CTR) of > 0.5. Methods This retrospective study included 1,784 patients from the development centre and 525 from an independent external validation centre. All patients had clinical T1N0 NSCLC, a tumour diameter ≤ 3 cm, and a CTR > 0.5, and they all underwent pulmonary resection with systematic lymph node dissection. We compared 12 machine-learning algorithms using 46 candidate preoperative features and assessed their discrimination, calibration, and potential clinical utility. SHapley Additive exPlanations (SHAP) was used to quantify feature contributions. An exploratory analysis was conducted to examine patients with tumours > 2 to ≤ 3 cm. Results oLNM was detected in 295 patients (16.5%) in the development cohort and 97 (18.5%) in the external cohort. The final 10-feature CatBoost model had an area under the receiver operating characteristic curve (AUC) of 0.870 (95% CI, 0.827–0.906) in the internal validation cohort and 0.855 (95% CI, 0.814–0.895) in the external validation cohort. Among 296 patients with tumours > 2 to ≤ 3 cm in the external validation cohort, 68 had oLNM; the AUC was 0.852 (95% CI, 0.802–0.898). External calibration revealed some risk overestimation at intermediate-to-high predicted probabilities. SHAP was applied to determine feature contributions across the cohort and for individual patients. Conclusions The model showed good discrimination of preoperative oLNM risk assessment in this population. Prospective validation of the developed model is required before its predictions can be used to guide the extent of nodal assessment or dissection.