Analysis and Mortality Prediction using Multiclass Classification for Older Adults with Type 2 Diabetes

Designing proper treatment plans to manage diabetes requires health practitioners to pay heed to the individuals remaining life along with the comorbidities affecting them. Older adults with Type 2 Diabetes Mellitus (T2DM) are prone to experience premature death or even hypoglycaemia. The structured dataset utilized has 68 potential mortality predictors for 275,190 diabetic U.S. military Veterans aged 65 years or older. A new target variable is invented by combining the two original target variables. Outliers are handled by discretizing the continuous variables. Categorical variables have been dummy encoded. Class balancing is achieved by random under-sampling. A benchmark regression model is built using Multinomial Logistic Regression with LASSO. Chi-Squared and Information Gain are the filter-based feature selection techniques utilized. Classifiers such as Multinomial Logistic Regression, Random Forest, Extreme Gradient Boosting (XGBoost), and One-vs-Rest classifier are employed to build various models. Contrary to expectations, all the models have constantly underperformed. XGBoost has given the highest accuracy of 53.03 percent with Chi-Squared feature selection. All the models have consistently shown an acceptable performance for Class 3 (remaining life is more than 10 years), significantly low for Class 1 (remaining life is up to 5 years), and the worst for Class 2 (remaining life is more than 5 but up to 10 years). Features analysis has deduced that almost all input variables are associated with multiple target classes. The high dimensionality of the input data after dummy encoding seems to have confused the models, leading to misclassifications. The approach taken in this study is ineffective in producing a high-performing predictive model but lays a foundation as this problem has never been viewed from a multiclass classification perspective.