External validation and nomogram for risk factors of CNS metastasis in patients with clinically localized melanoma.

Abstract

2012 Background: Surveillance for CNS mets is not routinely performed in CLM due to lack of evidence-based guidelines. We previously reported factors associated with CNS mets in CLM in a single institution study (MD Anderson Cancer Center; MDA). However, external validation of predictors is needed to develop evidence-based surveillance strategies. Methods: Demographics, primary tumor characteristics, and clinical events were collected for pts diagnosed with AJCC 8th edition Stage I-II melanoma in 1998-2014 in two institutions: MDA and Melanoma Institute Australia (MIA). Cumulative incidence of CNS mets was determined by competing risks method, including death; pts without CNS mets and alive at last follow-up (FU) were censored. The external validation of the MDA prognostic model was determined by calibration and discrimination using data from MIA. A ratio of observed and predicted outcomes (O/P ratio) was used to summarize calibration. An O/P ratio of 1 indicates perfect calibration, < 1 too high, and > 1 too low. The prediction model’s discriminative ability was obtained by area under operating characteristic curve (AUC), ranging from 0.5 (none) to 1.0 (perfect) discrimination. Nomograms for predicting Cumulative Incidence of CNS mets were produced. Results: MDA and MIA cohorts included 4,332 and 9,610 pts, respectively. Pts and clinical characteristics were similar in the cohorts. Median FU time was longer for MDA than MIA (88.9 vs. 26.3 months). The 2-, 5-, 10-year (yr), and final cumulative incidence of CNS mets were 1.5%, 4.9%, 7.8%, and 11.0% for MDA and 1.4%, 4.7%, 7.9%, and 10.8% for MIA pts, respectively. The initial MDA predictive model considered for validation included gender, primary tumor site, melanoma subtype, Breslow thickness (BT), ulceration, mitotic rate (MR), lymphovascular invasion and perineural invasion. External validation of the predictive model was performed at 2, 5, and 10 yrs. The original model’s validation properties were not ideal with the predicted estimates overestimating the observed values; thus the model was refined. The final reduced model evaluated included primary tumor site, melanoma subtype, BT, and MR. For calibration, O/P ratio (95% CI) for this model was 0.96 (0.75, 1.17) at 2 yrs, 0.97 (0.84, 1.11) at 5 yrs, and 1.03 (0.91, 1.15) at 10 years. The AUCs (95% CI) for 2, 5, and 10 yrs were 0.75 (0.70, 0.81), 0.71 (0.68, 0.75), and 0.69 (0.65, 0.72), respectively. The reduced model had similar AUCs with, but better calibration results than the original model. Nomograms of pts with high- and low-risk of CNS mets will be presented. A risk calculator will be available. Conclusions: A validated nomogram including primary tumor site, melanoma subtype, BT, MR can predict risk of CNS mets in CLM. This nomogram/risk calculator will help stratify CLM pts for CNS mets risk and facilitate development of personalized CNS surveillance strategies.