Predicting sentinel node positivity in patients with primary cutaneous melanoma: an international multicentre study validating and refining the MIA risk calculator

Background: The Melanoma Institute Australia (MIA) sentinel node (SN) metastasis risk-prediction online calculator[Lo2020] is now widely used around the world. The tool comprises eight models that variously use between three and six input parameters. However, the full (six-parameter) model has only been validated in the US population, the model with missing mitoses was validated for the Dutch and Swedish populations. Furthermore, some confidence intervals (CIs) of the tool are large due to uncommon input parameter values. This study further validated the tool in other populations and improved the precision of the risk estimates. Methods: Validation data were pooled from the Danish national Melanoma Database and eight cancer centres: UK(3), US(2), New Zealand(1), Sweden(1), Brazil(1). CI refinement data were pooled from this and previous validation and development cohorts. All patients had the SN result and the minimum data required for the tool (age, Breslow thickness and melanoma subtype), while the presence of ulceration, lymphovascular invasion or mitoses were included where available. The performance of the tool was assessed using C-statistics for discrimination and via a calibration plot. Re-calculation of CIs for the estimated risks was performed using the combined data from all original risk calculator development and validation cohorts. Results: The validation cohort consisted of 15,371 patients, 4,989 of whom had all six input parameters for the full model. The C-statistics were 73.0% (95% CI 70.6–75.3%) in the subset with all six parameters available, Receiver operating characteristic (ROC) curves showing the accuracy (AUC) and confidence intervals (CI) of the Melanoma Institute Australia (MIA) risk calculator to predict sentinel node positivity in the development cohort (black), n=3,477, and the international validation cohort with all six input parameters available (red), n=4,989. and 70.8%, 71.5% and 70.1% when 1, 2 or 3 optional parameters were missing. Calibration was excellent, with an intercept and calibration slope of 0.01 (95% CI -0.02–0.03) and 1.03 (95% CI 0.90–1.16), respectively. The revised CIs were substantially smaller than in the original tool, with a median reduction of over 75%. Conclusions: The results demonstrated that the MIA sentinel node risk-prediction tool performance was robust across a wide geographical range of populations. Furthermore, the precision of the models has been substantially improved with updated CIs based on a larger population sample. This study will therefore give users greater confidence in the tool’s reliability in predicting the risk of SN positivity. References [Lo2020] Lo, S. N., Ma, J., Scolyer, R. A., Haydu, L. E., Stretch, J. R., Saw, R. P. M., Nieweg, O. E., Shannon, K. F., Spillane, A. J., Ch'ng, S., Mann, G. J., Gershenwald, J. E., Thompson, J. F., Varey, A. H. R., (2020), Improved Risk Prediction Calculator for Sentinel Node Positivity in Patients With Melanoma: The Melanoma Institute Australia Nomogram, Journal of Clinical Oncology, 2719–2727, 38/24, doi: 10.1200/JCO.19.02362