EPID-05. A novel, clinically-relevant classification of pediatric CNS tumors for cancer registries using a clustering analysis

Abstract

Abstract To accurately evaluate the burden of pediatric central nervous system (CNS) tumors, estimate resources for cancer control, and monitor outcomes, a classification system that segregates tumors into clinically relevant groups is essential. The current classification of CNS tumors included in the third revision of the International Childhood Cancer Classification does not identify key clinical groups, such as low- and high-grade gliomas. To address this need, a novel classification was embarked upon using ICD-O-3 codes, CBTRUS grouping, incidence, survival, and treatment modalities as inputs. For each ICD-O-3 code with >50 new cases/year in CBTRUS from 2000 to 2016, 2 clinicians reached consensus defining the efficacy of three treatment modalities: surgical resection, radiotherapy, and chemotherapy. Then, patient level 5-year overall survival (OS) times were simulated based on total incidence and 5-year OS for each code. Subsequently, 5 factors were included as potential classifiers: tumor behavior, CBTRUS sub-group, and efficacy of the three treatment modalities. A “survival tree” was developed by using partitioning. Starting with the patient cohort (root), univariate cox proportional hazards model was used to identify statistically significant (P < 0.05) factors. The factors with the largest hazard ratio were selected manually to create child nodes. Within each child node, the partitioning process was repeated on remaining factors until no statistically significant factor remained. This clustering yielded 4 main groups (low-, intermediate-, high-, and very high-risk tumors) and 11 subgroups, including “embryonal tumors” and “low-risk glial and glioneuronal tumors”. Further validation of the classification will be sought through a structured consensus process using multidisciplinary experts. This systematic method to develop a classification for pediatric CNS tumors will allow for more relevant estimations of outcomes and better estimation of resource utilization. Furthermore, this strategy could be replicated for other disease groups.