Risk Prediction Models for Sentinel Node Positivity in Melanoma: A Systematic Review and Meta-Analysis.

Importance: There is a need to identify the best performing risk prediction model for sentinel lymph node biopsy (SLNB) positivity in melanoma.

Objective: To comprehensively review the characteristics and discriminative performance of existing risk prediction models for SLNB positivity in melanoma.

Data sources: Embase and MEDLINE were searched from inception to May 1, 2024, for English language articles.

Study selection: All studies that either developed or validated a risk prediction model (defined as any calculator that combined more than 1 variable to provide a patient estimate for probability of melanoma SLNB positivity) with a corresponding measure of model discrimination were considered for inclusion by 2 reviewers, with disagreements adjudicated by a third reviewer.

Data extraction and synthesis: Data were extracted in duplicate according to Data Extraction for Systematic Reviews of Prediction Modeling Studies, Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis, and Preferred Reporting Items for Systematic Reviews and Meta-Analyses reporting guidelines. Effects were pooled using random-effects meta-analysis.

Main outcome and measures: The primary outcome was the mean pooled C statistic. Heterogeneity was assessed using the I2 statistic.

Results: In total, 23 articles describing the development of 21 different risk prediction models for SLNB positivity, 20 external validations of 8 different risk prediction models, and 9 models that included sufficient information to obtain individualized patient risk estimates in routine preprocedural clinical practice were identified. Among all risk prediction models, the pooled weighted C statistic was 0.78 (95% CI, 0.74-0.81) with significant heterogeneity (I2 = 97.4%) that was not explained in meta-regression. The Memorial Sloan Kettering Cancer Center and Melanoma Institute of Australia models were most frequently externally validated with both having strong and comparable discriminative performance (pooled weighted C statistic, 0.73; 95% CI, 0.69-0.78 vs pooled weighted C statistic, 0.70; 95% CI, 0.66-0.74). Discrimination was not significantly different between models that included gene expression profiles (pooled C statistic, 0.83; 95% CI, 0.76-0.90) and those that only used clinicopathologic features (pooled C statistic, 0.77; 95% CI, 0.73-0.81) (P = .11).

Conclusions and relevance: This systematic review and meta-analysis found several risk prediction models that have been externally validated with strong discriminative performance. Further research is needed to evaluate the associations of their implementation with preprocedural care.