Explainable AI to enable precision risk stratification for advanced melanoma in underrepresented Middle Eastern cohorts.

Abstract

e21526 Background: Early identification of advanced-stage melanoma using baseline clinical and histopathological data could enable risk-stratified surveillance and earlier intervention, potentially improving outcomes. However, most predictive models are developed from Western populations, limiting their applicability to underrepresented Middle Eastern cohorts. We developed explainable machine learning (ML) models to identify actionable predictors of advanced-stage disease and brain metastasis using pre-treatment data from a Lebanese tertiary center cohort. Methods: We retrospectively analyzed 327 melanoma patients diagnosed at the American University of Beirut Medical Center. After excluding 132 patients with incomplete staging data, 195 patients were included for predictive modeling. The primary outcome was advanced-stage disease at diagnosis (stage III or IV; n = 72), with early-stage disease (stage 0–II; n = 123) as the comparator. The secondary outcome was brain metastasis during follow-up (n = 143; 9 events). We selected 20 clinically relevant baseline variables spanning demographics, medical history, and tumor characteristics, ensuring no data leakage by excluding post-diagnosis variables. We evaluated five machine learning algorithms using 5-fold stratified cross-validation with 1000 bootstraps. Feature importance was assessed using SHapley Additive exPlanations (SHAP), and robust predictors were identified by convergent evidence from SHAP rankings and bootstrap-validated univariable logistic regression. Results: The Random Forest model achieved the highest discriminative performance for advanced-stage prediction (AU-ROC 0.814, 90% CI 0.743–0.846; AU-PRC 0.781, 90% CI 0.636–0.827; F1-score 0.621). Gradient Boosting followed closely (AU-ROC 0.793). Univariable analysis identified five robust predictors of advanced stage, including Clark level V (OR 7.67, 90% CI 4.91–12.43, p = 0.018), high mitotic count (≥7 mitoses; OR 6.12, p = 0.033), lymphovascular invasion (OR 4.27, p = 0.017), nodular histology (OR 3.07, p = 0.017), and Charlson Comorbidity Index (CCI) (OR 1.41 per unit increase, p < 0.001). Conversely, radial growth phase was protective (OR 0.25, p = 0.016). SHAP explainable ML analysis ranked CCI as the most influential predictor (mean |SHAP| = 1.06). Notably, CCI was also significantly associated with brain metastasis risk (OR 1.35, 90% CI 1.10–1.65, p = 0.014). Conclusions: Our explainable ML framework enables immediate risk stratification at diagnosis using five convergent predictors: comorbidity burden (CCI), Clark level, mitotic count, lymphovascular invasion, and histology subtype (nodular), with radial growth phase showing a protective effect. These findings support targeted surveillance and management for high-risk patients, advancing equitable AI-driven precision oncology for underrepresented populations.