Classifying Pediatric Central Nervous System Tumors through near Optimal Feature Selection and Mutual Information: A Single Center Cohort

Background: Labeling, gathering mutual information, clustering and classificationof central nervous system tumors may assist in predicting not only distinct diagnosesbased on tumor-specific features but also prognosis. This study evaluates the epidemi-ological features of central nervous system tumors in children who referred to Mahak’sPediatric Cancer Treatment and Research Center in Tehran, Iran.Methods: This cohort (convenience sample) study comprised 198 children (≤15years old) with central nervous system tumors who referred to Mahak's PediatricCancer Treatment and Research Center from 2007 to 2010. In addition to the descriptiveanalyses on epidemiological features and mutual information, we used the LeastSquares Support Vector Machines method in MATLAB software to propose apreliminary predictive model of pediatric central nervous system tumor feature-labelanalysis. Results:Of patients, there were 63.1% males and 36.9% females. Patients' mean±SDage was 6.11±3.65 years. Tumor location was as follows: supra-tentorial (30.3%), infra-tentorial (67.7%) and 2% (spinal). The most frequent tumors registered were: high-gradeglioma (supra-tentorial) in 36 (59.99%) patients and medulloblastoma (infra-tentorial)in 65 (48.51%) patients. The most prevalent clinical findings included vomiting,headache and impaired vision. Gender, age, ethnicity, tumor stage and the presence ofmetastasis were the features predictive of supra-tentorial tumor histology.Conclusion: Our data agreed with previous reports on the epidemiology of centralnervous system tumors. Our feature-label analysis has shown how presenting features maypartially predict diagnosis. Timely diagnosis and management of central nervous systemtumors can lead to decreased disease burden and improved survival. This may be furtherfacilitated through development of partitioning, risk prediction and prognostic models.