Local Interpretable Model-Agnostic Explanations of Predictive Models for Hypertension

Abstract

Hypertension is one of the chronic medical conditions and the major risk factor for multiple diseases and strokes. Prevention of hypertension is significant to delay the incidence of disease progression and decrease the severe health complications. In recent years, predictive models have been developed to recognize hypertensive and normotensive individuals. However, most accurate machine learning classifiers are problematic to understand the rationale behind their predictions and utilize in medical decision support systems due to the lack of interpretation of black boxes. In this study, we suggest the model-agnostic explanation approach to enhance the interpretability for the end-user of the system. Our proposed framework consists of the main two steps. In the first step, significant features were selected harnessing the t-test and chi-square test followed by a dissimilarity feature selection using correlation analysis. Thereafter, we construct the prediction models using several machine learning classifiers in order to determine the best hypertension predictive model among the Korean population. In the second step, Local Interpretable Model-Agnostic Explanations (LIME) are utilized to give interpretation at the individual level for each prediction model. In our experimental result, the eXtreme Gradient Boosting (XGBoost) and neural networks (NN) achieved the best prediction performances. In addition, utilization of the LIME technique is presented the prediction probability in each risk factor at the individual level. This may support to healthcare experts to make personalized decisions as well as public healthcare concerns.