A Bayesian Belief Network-based Probabilistic Mechanism to Determine Patient No-Show Risk Categories

Abstract Patients who miss their appointments (no-shows) reduce revenues and impair the delivery of quality healthcare. Much research has been devoted to identifying “no-show patients”. In this study, we build a Tree Augmented Naive Bayes (TAN)-based, probabilistic data driven methodology that consists of five steps. After data acquisition and preparation in the first step, the second step is dedicated to selecting important variables through purely data-driven wrapper methods such as Extreme Gradient Boosting (XGB), Particle Swarm Optimization (PSO) and Genetic Algorithms (GA). Then, in the third step, the Synthetic Minority Oversampling Technique (SMOTE) and Random under Sampling (RUS) are employed to overcome the data imbalance issue that exists in the dataset. In the fourth step, the conditional interrelations among the predictors are obtained via TAN model, and Bayesian-belief based posterior probabilities are calculated. It should be noted that, the parameters of the TAN model are tuned in a way that there does not have to exist a conditional interrelation between every single predictor. The consistency and reliability of these posterior probabilities are then justified/investigates via the proposed algorithm. The patients are clustered into 5 risk groups, and a web-based decision support system graphical user interface is built as a proof of concept. Results show that interesting conditional inter-relations exist between the seven variables that are commonly selected through by GA and PSO. In addition, an overall area under the receiver operating characteristic curve (ROC) score of 0.828 was achieved with a sensitivity score of 0.785. The reliable, TAN-based posterior probabilities and conditional relationships among the predictors for such a parsimonious model that has a fairly high sensitivity in detecting the minority samples, can be adopted by primary care facilities to improve the decision-making process in managing the no-show problem.