Enrico Martin

53 Treatment and survival in the Netherlands whenever diagnoses were mentioned as doubtful or the diagnosis changed after e.g. (metastasis) resection, cases were excluded. Covariates Covariates extracted for analysis were: year of diagnosis (1989-2005/2006-2017), sex, age, established diagnosis of NF1, tumor site, tumor stage (presence of metastasis/no metastasis), tumor size (≤5/>5cm), tumor depth (superficial/deep of the fascia), tumor morphology (Triton tumor/within neurofibroma), obtained surgical resection margin (R0/R1/R2), the use of other treatment modalities, and sequence of treatment. NF1 status was extracted from pathological reports and was concluded either when stated as such in the report or when a pathology report of previous plexiform neurofibroma resections or two or more neurofibromas was present. Tumor sites were categorized as: head & neck, extremities, trunk (including thorax, abdomen, and pelvis), retroperitoneal, and not otherwise specified (NOS). Resection margins were regarded as tumor-free (R0), microscopically positive (R1, <1mm margin), and macroscopically positive (R2). Tumor grade is not registered in the NCR and its reporting is inconsistent in pathological reports. Vital status and date of death are routinely obtained from municipal demographic registries in the Netherlands. Pediatric and synchronous metastatic cases were excluded from all statistical analyses as they are treated differently. Statistical analysis Overall, analyses were stratified between retroperitoneal and non-retroperitoneal localized MPNSTs as they are generally treated differently. Estimated median survival was calculated using the Kaplan-Meier method for several covariates of interest and differences were assessed with log-rank tests. A conditional inference tree was constructed for localized non-retroperitoneal MPNST using the R package “partykit” to evaluate the most important predictors for survival. 18 A conditional inference tree generates a decision tree that splits the population of interest into subpopulations by means of recursive partitioning. At each partition, the best predictor separates one node into two child nodes. The decision tree extends until it cannot find any predictor that can significantly divide a node. Two separate Cox proportional hazard models were constructed for localized non-retroperitoneal MPNSTs and retroperitoneal MPNSTs by backward selection. Adjusted survival curves were made for individual prognostic factors, based on the final model. 19 Statistical analyses and data visualization were conducted using R version 3.6.0 (R Core Team, 2019). 3