Machine learning model for predicting regional lymph node metastasis in acral melanoma

Acral melanoma (AM) is an aggressive subtype of cutaneous melanoma characterized by delayed diagnosis and poor prognosis. Accurately and non-invasively assessing regional lymph node metastasis (RLNM) remains a significant clinical challenge. This study aimed to develop an interpretable machine learning (ML) model to preoperatively predict RLNM in AM patients. Clinical, pathological and imaging data of 189 patients with histologically confirmed AM were analysed retrospectively. Patients were randomly assigned in a 7:3 ratio to a training cohort (n = 132) and an internal validation cohort (n = 57). Boruta and least absolute shrinkage and selection operator (LASSO) regression were used to screen candidate variables within the training cohort. Seven algorithms were compared by nested cross-validation performed exclusively inside the training cohort; the final algorithm was chosen by a pre-specified composite of out-of-fold discrimination and calibration. The decision threshold was fixed on the training out-of-fold predictions. The internal validation cohort was evaluated once, with the model and the threshold locked in advance. SHAP (Shapley Additive Explanations) analysis was used to interpret the locked model, and calibration, decision curve analysis (DCA) and DeLong tests against conventional Breslow/TNM-based models were performed on the untouched validation cohort. Five predictors were retained: Breslow thickness, lymph-node short-axis diameter, positive contrast-enhanced ultrasound (CEU) finding, vascular invasion and positive positron emission tomography/computed tomography (PET-CT) finding. In the training cohort, the pooled out-of-fold area under the curve (AUC) ranged from 0.850 to 0.877 across the seven algorithms. On the pre-specified composite rule, the GradientBoost model was selected and locked; it achieved a single locked AUC of 0.868 (95% CI 0.769–0.967) in the internal validation cohort, with a sensitivity of 0.889 (95% CI 0.762–1.000) and a specificity of 0.733 (95% CI 0.565–0.885). The locked model significantly outperformed the Breslow-only logistic model (AUC 0.610, DeLong p = 0.0002) and a categorical T-stage (TNM-T) logistic model (AUC 0.562, DeLong p < 0.001). Calibration was adequate (slope 1.02, intercept − 0.63, Brier score 0.156) and DCA demonstrated a positive net benefit over the conventional models across clinically relevant threshold probabilities. This study developed a machine learning-based non-invasive diagnostic model that effectively assesses the risk of regional lymph node metastasis in patients with AM. The locked gradient boosting model retained good discrimination and calibration and outperformed conventional Breslow/TNM-based models. This non-invasive model may assist in assessing the risk of RLNM in patients with AM.