Decoding Medical Diagnosis with Machine Learning Classifiers

Abstract

Clinical Decision Support Systems (CDSS) are gaining popularity in disease screening and grading in the current era of digital healthcare. This paper attempts to model how a computer learns to grade an Infectious Disease (ID), e.g., Typhoid fever using the Machine Learning (ML)-based approach mimicking how a novice doctor learns to diagnose a case with the help of senior doctors. To achieve the goal, ten Virtual Junior Clinicians (VJCs) are developed using ten Machine Learning Classifiers (MLC)-based CDSS, which are then trained with 'weighted' [0, 1] sign-symptoms and the corresponding 'labeled'grade of synthetic Typhoid fever cases (N = 198). Weights and labels are assigned by ten Senior Clinicians (SC) providing their rich Clinical Knowledge Base (CKB). The performance of each VJC is then measured in terms of their diagnostic Accuracy, Precision, Recall, and F-score. Results show that Random Forest (RF, i.e., VJC9) and Decision Tree (DT, i.e., VJC4)-based CDSS can grade with an average of 87% accuracy, which is even higher than human clinicians' accuracy. The reason behind RF and DT's appreciable performance is that clinicians use tree-search-based methods with probabilistic 'yes' and 'no' logic to learn the disease patterns alike the working principles of DT and RF for diagnosing and grading any ID. Apart from modeling, the paper provides insight into how to select the right Machine Learning Classifier (MLC) algorithm in the field of ID diagnosis. It also throws light on various hardships and challenges with MLC-based CDSS implementations in the real-world scenario.