Identifying brain tumor patients' subtypes based on pre-diagnostic history and clinical characteristics: a pilot hierarchical clustering and association analysis

Introduction: Central nervous system (CNS) tumors are severe health conditions with increasing incidence in the last years. Different biological, environmental and clinical factors are thought to have an important role in their epidemiology, which however remains unclear. Objective: The aim of this pilot study was to identify CNS tumor patients' subtypes based on this information and to test associations with tumor malignancy. Methods: 90 patients with suspected diagnosis of CNS tumor were recruited by the Neurosurgery Unit of IRCCS Neuromed. Patients underwent anamnestic and clinical assessment, to ascertain known or suspected risk factors including lifestyle, socioeconomic, clinical and psychometric characteristics. We applied a hierarchical clustering analysis to these exposures to identify potential groups of patients with a similar risk pattern and tested whether these clusters associated with brain tumor malignancy. Results: Out of 67 patients with a confirmed CNS tumor diagnosis, we identified 28 non-malignant and 39 malignant tumor cases. These subtypes showed significant differences in terms of gender (with men more frequently presenting a diagnosis of cancer; p = 6.0 ×10-3) and yearly household income (with non-malignant tumor patients more frequently earning ≥25k Euros/year; p = 3.4×10-3). Cluster analysis revealed the presence of two clusters of patients: one (N=41) with more professionally active, educated, wealthier and healthier patients, and the other one with mostly retired and less healthy men, with a higher frequency of smokers, personal history of cardiovascular disease and cancer familiarity, a mostly sedentary lifestyle and generally lower income, education and cognitive performance. The former cluster showed a protective association with the malignancy of the disease, with a 74 (14-93) % reduction in the prevalent risk of CNS malignant tumors, compared to the other cluster (p=0.026). Discussion: These preliminary data suggest that patients' profiling through unsupervised machine learning approaches may somehow help predicting the risk of being affected by a malignant form. If confirmed by further analyses in larger independent cohorts, these findings may be useful to create potential intelligent ranking systems for treatment priority, overcoming the lack of histopathological information and molecular diagnosis of the tumor, which are typically not available until the time of surgery.