Application of machine learning algorithms for the diagnosis of primary brain tumors.

Abstract

2533 Background: Primary lesions of the CNS refer to a heterogeneous group of benign or malignant tumors arising in different parts of the brain and spinal cord. According to the 2016 CNS WHO classification, the accurate diagnosis of primary brain tumors requires a layered approach of histologic, anatomic and molecular features to generate an integrated diagnosis with clinical and prognostic significance. However, in the US and worldwide, scarce resources are available to perform all the required tests routinely, so methods that improve pre-test probabilities and decrease false positive results have significant clinical and financial impact. Aims: 1) validate new diagnostic workflows with implementation of modern machine learning/artificial intelligence approaches; 2) design a reliable and interactive computational platform for primary CNS tumor diagnosis. Methods: To achieve these goals we have developed a population model in Rstudio, “La Tabla”, based on the articles from open resources of MEDLINE database and the latest version of WHO classification of CNS tumors. The data of “La Tabla” is comprised of more than 100,000 adult and pediatric cases, including rare brain tumor diagnoses, such as Gangliocytoma, Diffuse Midline Glioma and etc. Results: Boruta package and weights function in R have been used to distinguish the most important features for diagnosis prediction. To visualize correlation between these features (age, ki67 level, tumor location, presence of myxoid areas, calcifications, necrosis and etc.) and all diagnoses in two-dimensional space, we used a t-SNE algorithm. Models trained with decision tree algorithms (randomForest, XGBoost and C5.0) showed high overall accuracy in predicting diagnoses of “La Tabla” (95%, 94% and 92%) and 300 patients at OSUCCC-James (93%, 74% and 87%) in the absence of IHC and molecular data. Neural networks provided by keras and nnet packages predicted diagnoses using just clinical and histological findings with 94% and 88% accuracy on “La Tabla” and James patient databases respectively. Currently, we are building “Shiny” applications with R to deliver easily operated platform for pathologists and physicians. Conclusions: In conclusion, we managed to generate models that are able to diagnose primary brain lesions using basic clinical data (age, gender, tumor location), ki67 levels and distinct features of histological architecture. Most of the models distinguish brain tumors and associated molecular status with high accuracy and will serve as a reliable tool for second opinion in clinical neuro-oncology.