Application of Shapley Additive Explanation Towards Determining Personalized Triage from Health Checkup Data

Abstract

Machine learning has become a powerful tool to assist humans making decisions. In most cases, machine learning models act like a black box, a user can only view the outcome without knowing the decision-making process or the deciding factors. Explainable AI has shown good performance in interpreting prediction models and identifying the influential parameters behind the prediction/decision. Our previous works have been analyzing health checkup data collected by a digital healthcare system, called Portable Health Clinic (PHC), developed by us. The system uses a standard logic set based on WHO recommendations to triage the health status of a patient. The triage used in PHC is almost a static standard logic set that works for any patient at any age. We argue that the triage logic should vary from person to person. This paper attempts to use explainable AI to check whether triage could be personalized. An experiment has been carried out over a health check-up data set (N = 44,460), by applying XGBoost, a popular machine learning algorithm to predict a patient’s health status (risky or not risky). An eXplainable AI (XAI) technique called SHAP is used to explain the prediction results. The SHAP value clearly indicates that each health parameter (BMI, Blood Pressure, hemoglobin, etc.) has different cut-off points for different age groups, which suggests that the threshold to determine one’s health status is different and can be obtained. The results will be useful to improve the existing triage static logic. This paper demonstrates cut-off points for BMI and Blood Pressure (Systolic) for two age groups which is an indication of group triage. Our future work will search for the individual cut-off point for developing personalized triage. The obtained cut-off points need to be verified by health professionals.