Readmission Prediction of Diabetic Patients

This project is for GW hippo hacks.

PROBLEM DEFINITION AND MOTIVATION: Hospital readmissions for Diabetes patients are very expensive since hospitals face severe penalties if the readmission rate is higher than expected.Therefore, in order to ensure better value,more satisfied patients and reduced readmission rates an attempt has been made by the authors to identify the key factors that influence readmission of diabetes patients and predict the probability of diabetic patients’ readmission.The dataset under consideration represents 10 years of clinical care and includes over 50 features representing patient and hospital outcomes.This problem falls under the category of Classification problem and suitable Classification Algorithms like Random Forest and Decision Trees are expected to help identify if a patient needs to readmitted or not.The development of prediction models through three different machine learning approaches is expected to help develop a more lateral understanding of the problem at hand and arrive at suitable recommendations to tackle the problem.

PROPOSED METHOD: The specific steps involve preprocessing the data with imputation function included, identifying various modeling techniques suitable to this dataset. The authors propose that the entire dataset be divided into train and test groups based on 70:30 split and deploy the algorithms : Logistic Regression, Decision Tree and Random Forest Algorithms.The parameters for each algorithm will be chosen based on the training set and the performance will be evaluated on the test set.The performance of the machine learning models will be compared and contrasted in order to assess the superiority of these models with the help of Precision-Recall Curve.

Target column in the data set is readmission, which is a categorical column and has three values >30, <30, NO. NO means no readmission.

30 means readmission after 30 days of discharge.
<30 means readmission before 30 days of discharge.

Two use cases for the project: Use Case 1: Converted >30 and <30 to Yes and converted the readmission column to a binary column and used machine learning algorithms to predict yes or no for readmission and find columns affecting the readmission.

Use Case 2: Used only >30 and <30 . Tried to find if readmission is True then whether that readmission will be within a month or after a month and the factors affecting it.

Conclusion:- Use Case 1: This gave an accuracy of around 62 percent using Logistic regression. Use Case 2: This gave an accuracy of around 75.77 percent using Logistic regression.

This can be used to predict readmission for person diagnosed with diabetes and help them get proper care. In future, an application can be made with the trained model in the back-end. Data regarding the patient can be filled in the application interface by the doctor itself, then the data goes through the machine learning pipeline, to clean and preprocess it and then give a result for readmission. If the result it true, it can further classify whether the readmission will be within 30 days or after 30 days. These results can help both the doctor and patient.

Built With

• jupyter-notebook
• python