Font Size: 

Predictive analytics embraces an extensive range of techniques including but are not limited to statistical modeling, machine learning, Artificial Intelligence and data mining. It has profound usefulness in different application areas such as data-driven decision making, business intelligence, public health, disaster management and response, as well as many other fields. In this study, we design and implement a predictive analytics system that can be used to forecast the likelihood that a diabetic patient will be readmitted to the hospital.  Upon extensively cleaning the Diabetes 130-US hospitals dataset containing patient records spanning for over 9 years i.e. from 1999 till 2008, we modelled the relationship between the predictors and the response variable using the XGBoost classifier. Upon performing hyperparameter optimization for the XGBoost, we obtained a maximum AUC of 0.671. Our study reveals that attributes such as lab procedures, number of medication, time in hospital, discharge disposition and number of inpatient visit are strong predictors for the response variable (i.e. re-admission of patients). Findings from this study can help hospitals design suitable protocols to ensure that patients with a higher probability of re-admission are recovering well and possibly reducing the risk of future re-admission. In the long run, not only will our study improve the life quality of diabetic patients, it will also help in reducing the medical expenses associated with re-admission.

 

Bias-Variance Trade off, Sampling technioque, Hyperparameter tuning, Classification and Regression