Publication

Abstract : A brain tumor is a mass of cells growing abnormally in the brain. The lesions formed in the suprasellar region of the brain, called suprasellar lesions, affect common anatomical locations causing an array of symptoms, including headache and blurred or low vision. These symptoms lead to a misdiagnosis of the tumor as common issues like refractive index problems, and the tumor gets diagnosed very late. This study focuses on these suprasellar lesions (namely Pituitary adenoma, Craniopharyngioma, and Meningioma), which have not been explored much using machine learning. We have collected 422 discharge summaries of patients admitted to the neurosurgery department of the National Institute of Mental Health and Neuroscience (NIMHANS), Bangalore, India, during 2014-2019. This work aims to build a model for classifying lesions into three categories. Features are the clinical concepts identified from the discharge summary using Natural Language Processing (NLP) and regular expression-based rules. The features and corresponding values thus extracted are represented as Analytical Base Table and fed to the classification model after the processing. The model utilizes XGBoost, Local Cascade Ensemble, Histogram-based gradient boosting, LightGBM, and CatBoost classifiers, which have the ability to inherently handle the missing data. Though the machine learning models perform well in classification, the interpretability and generalizability is often questioned especially in critical domains such as medical and healthcare. Hence model performance has been analyzed using the ELI5 tool, a python package for explainable AI. This tool identifies the critical features in the data on a patient basis, providing a more interpretable model for clinicians.