The recurrence problem Most melanoma deaths occur in patients whose cancer was originally caught at early stage (I or II) but later recurred. Recurrence often goes undetected until symptomatic metastatic disease appears, and with immunotherapy now approved for stage IIB and IIC melanoma, clinicians urgently need reliable tools to identify which patients are at highest risk.
Study approach Researchers from Harvard Medical School assembled 1,720 early-stage melanomas from two major Boston cancer centers (Mass General Brigham and Dana-Farber Cancer Institute), extracted 36 clinical and histopathologic features from electronic health records, and applied multiple supervised machine-learning algorithms to predict recurrence.
Main findings The best-performing model (gradient boosting) achieved AUC of 0.845 internally and 0.812 on external validation for recurrence classification, and time-dependent AUC of 0.853 and 0.820 respectively for time-to-event prediction. Breslow tumor thickness and mitotic rate were the top two predictive features across all models.
Patient cohort The study enrolled 1,720 stage I/II cutaneous melanomas diagnosed between 2000 and 2020 - 1,172 from MGB (used for training) and 548 from DFCI (used as independent external test set). Median follow-up was 7.2 years and 310 (18%) melanomas recurred, of which 48.7% were distant recurrences.
Feature categories A total of 36 features were extracted covering demographics (age, sex, race, insurance type, median income), medical history (comorbidities, prior skin cancers, autoimmune disease), and 22 tumor characteristics (histologic type, site, Breslow thickness, mitotic rate, ulceration, lymphovascular invasion, and others).
Machine learning algorithms Five algorithms were tested for binary classification: gradient boosting (GB), random forest (RF), logistic regression (LR), multi-layer perceptron (MLP), and support vector machine (SVM). For time-to-event prediction, gradient boosting survival analysis, random survival forest, Coxnet, and CoxPH were applied.
Recurrence classification performance Gradient boosting achieved the highest performance in both internal (AUC: 0.845, PPV: 0.803) and external (AUC: 0.812, PPV: 0.785) validations. SVM performed worst. Tumor characteristics dominated - adding them to demographics and medical history significantly improved AUC for all models.
Time-to-event prediction The gradient boosting time-to-event model (GB-T) achieved internal time-dependent AUC of 0.853 and concordance index 0.820, and external time-dependent AUC of 0.820 and concordance index 0.809 - consistently the best of all survival models. Random survival forest performed close behind.
Sensitivity analyses confirm robustness When restricted to patients with complete data for core tumor features (known mitotic rate, ulceration, and negative lymph nodes), performance remained similar (external AUC: 0.809 vs. 0.812). Extending the minimum follow-up for non-recurrent melanomas to 7 years actually slightly improved external validation AUC to 0.827.
Breslow thickness and mitotic rate dominate Across all gradient boosting and random forest models for both binary classification and time-to-event prediction, Breslow thickness and mitotic rate were consistently the top two most important features. This held true regardless of whether other collinear features like AJCC stage or anatomic level were included.
Mitotic rate's underappreciated importance Mitotic rate was recently removed from the AJCC 8th edition T1 staging criteria due to poor interobserver reliability. Yet in this study it remains among the top two features across all well-performing models, suggesting it carries independent prognostic value that current staging guidelines may underutilize.
Socioeconomic factors as risk signals Insurance type (particularly self-pay) and median income consistently ranked in the top 10 features. Linear regression analysis confirmed that self-pay and Medicaid patients presented with significantly thicker tumors at diagnosis, reflecting documented healthcare access disparities that directly affect prognosis.
Adjuvant therapy decision support The most immediate clinical application is identifying stage I/II patients at high recurrence risk who might benefit from adjuvant immunotherapy, recently approved for stage IIB and IIC disease. The current models can differentiate high-risk from lower-risk patients better than AJCC staging alone.
Individualized surveillance intensity Time-to-event recurrence probability curves generated by the GB-T model for individual patients could guide surveillance intervals - patients predicted to have recurrence probability greater than 0.8 at 10 years warrant more aggressive follow-up than those below 0.1.
Balancing immunotherapy benefits against risks Adjuvant immunotherapy carries immune-related adverse events in up to 40% of patients, some potentially fatal. By identifying who is truly at high recurrence risk, the ML models could help clinicians select patients most likely to benefit, reducing unnecessary exposure to treatment toxicity.
Performance plateau and missing data types Despite incorporating 36 features, the models plateau around AUC 0.81-0.85. The authors explicitly note that incorporating digital histopathology images, genomic data, and novel tumor biomarkers could significantly improve performance beyond this ceiling.
Retrospective design and measurement variability The study is retrospective, and mitotic rate was extracted from pathology reports - a metric known for high interobserver variability. Future studies using automated mitotic figure detection from whole slide images via deep learning could provide a more objective and reproducible measurement.
Future integration goals Planned enhancements include deep learning analysis of histopathology whole-slide images, integration of genomic and proteomic data, and larger prospective cohort validation. The code is publicly available at GitHub (SemenovLab/early-stage-melanoma-recurrence-prediction) to enable external replication and extension.