Melanoma Intelligence: Explainable AI Reveals Histopathologic Aggressiveness as the Dominant Axis of Lymph-Node Metastasis

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Background/Objectives: Lymph-node metastasis remains central to staging, prognosis, surveillance, and treatment planning in malignant melanoma. At present, most statistical models assessing nodal metastatic risk in malignant melanoma consider pathological descriptors, inflammatory markers, metabolic alterations, clinical data, and related variables independently of each other. Therefore, we developed a transparent artificial intelligence (AI)-based approach to assess whether the propensity for nodal metastasis is determined by a single layer of local histopathological aggressiveness or by the integration of different biological levels, including local histopathological aggressiveness, systemic inflammatory–metabolic dysregulation, biological heterogeneity, or a clinicobiological pattern. Methods: In this retrospective study, we assessed 73 adult patients undergoing surgical removal of malignant melanoma. The primary endpoint was histopathologically confirmed lymph-node metastasis. Routinely collected patient-related data, including clinical, anatomical, operative, histopathological, nodal, comorbidity, biological, clinical course, and available staging data, were structured into interpretable constructs. These included the Histopathologic Aggressiveness Index (HAI), the Inflammatory–Metabolic Dysregulation Index (IMDI), the Biological–Histological Discordance Score (BHDS), model-estimated nodal metastatic probability, integrated clinicobiological risk, and explanation stability. The AI-based framework was evaluated by applying bias-reduced and penalized logistic regression, machine learning benchmarking, leave-one-out cross-validation, bootstrap estimation, permutation testing, decision curve analysis, rule extraction, feature stability evaluation, network analysis, similarity-based retrieval, conformal uncertainty estimation, and unsupervised phenomapping. Results: For 72 out of 73 patients, nodal histopathology results were available. Among these patients, 22 had positive nodal status. Positive nodal status was associated with a higher Breslow thickness, an increased mitotic rate, ulceration, lymphovascular invasion, a nodular subtype, and palpable adenopathy. The HAI demonstrated the strongest discriminative signal between node-positive and node-negative patients (median values of 67.8 vs. 45.3; p < 0.001) and retained an independent association with nodal metastasis within the bias-reduced logistic model (odds ratio [OR] per 10-point increase: 2.74; 95% confidence interval [CI]: 1.58–4.75; p < 0.001). The IMDI showed a weak exploratory relationship and did not retain an independent association after adjustment. Similarly, the BHDS did not show significant differences in separating the two endpoint groups. The penalized logistic model including only the HAI showed good performance under leave-one-out cross-validation, with ROC AUC = 0.889, PR-AUC = 0.706, and Brier score = 0.137. Through rule extraction, we found a cohort-specific HAI threshold value > 58.6, above which all node-positive cases were located. With respect to explainability, feature stability, network analysis, similarity retrieval, conformal prediction, and phenomapping, there was convergence toward a dominant high-risk phenotype defined primarily by histopathological criteria. Conclusions: Routine melanoma registries may be transformed into internally evaluated melanoma intelligence frameworks. Histopathologically confirmed lymph-node metastasis among patients with malignant melanoma was organized primarily along an axis of local histopathological aggressiveness, while systemic inflammatory–metabolic dysregulation provided subordinate contextual biological information.