Determining the most effective machine learning model on segmented vs. unsegmented patient data to assign Drug Classification

Abstract

Throughout history and in the modern world, accurate drug prescription has been one of the most important tasks that a medical professional does in their everyday tasks. Through recent innovations in the machine learning environment, algorithms and models can more accurately predict the physiological activity of drugs and further classify drugs based on their physiological properties. This project focuses on the latter, and leverages a sample patient dataset and runs logistic regression, k-neighbors, support vector machine, naïve Bayes, decision trees, and the random forest models to determine accuracy between models. Afterwards, the data is segmented by sex, and the models are implemented on each dataset, and the accuracies are compared. The accuracy for the models that were applied to the entire dataset are the following (greatest to least accurate): decision tree (100%), random forest (100%), SVM (98%), naive Bayes (83%), logistic regression (83%), and k-Neighbors (66.67%). Overall, segmentation had the smallest effect on the Random Forest and Decision Tree models as both produced a 0% difference in accuracy between male and female datasets, and had the biggest effect on the k-neighbors model with a 38.03% between male and female datasets.