Comparative Analysis of Data Preparation Methods Employed in Prediction of Diabetes Mellitus Diagnosis using eICU-CRD Dataset

In any classification system it’s necessary to preprocess the source data in order to draw useful insights. In this paper, two different preprocessing approaches on WiDS Datathon 2021 Data were compared and analyzed deeply to account for the difference between resultant prediction accuracies.A comparative analysis was conducted on the preprocessing steps of both the approaches The steps applied are but not limited to - handling missing values, Feature scaling, Standardization, etc. in the process of determining whether a patient has been diagnosed with Diabetes Mellitus before. The readings in the first 24 hours from the patient’s stay in ICU were used here and all the readings were acquired through invasive and noninvasive methods. LightGBM Classifier was used for the purpose of classifying the target variable.An attempt to determine the probability that a patient who is admitted to an ICU has been diagnosed with Diabetes Mellitus was made using two approaches with different machine learning pipelines. The first pipeline resulted in an accuracy score of 0.87242 and the second one gave a score of 0.86471. The boosting algorithms used in both the methods were the same but due to the differences in the missing values handling techniques and feature engineering steps, their prediction accuracies were different.