A predictive modeling approach for Taiwanese diagnosis-related groups medical costs: A focus on laparoscopic appendectomy

Abstract

Abstract Background: In the Taiwanese diagnosis-related groups (Tw-DRGs) system, effective management of medical resources is vital to ensure the sustainability of hospital operations. Objectives: The aim is to create a predictive model to estimate the medical costs linked to a specific Tw-DRGs item, utilizing laparoscopic appendectomy without complications or comorbidities (DRG16701) as a representative case. Materials and Methods: We employed a dataset comprising 248 surgical cases performed at a regional teaching hospital between January 2017 and December 2019. These cases were classified based on the difference between the Tw-DRGs payment standard and actual medical costs. Two experiments were conducted: one without feature selection and one with feature selection. We utilized random forest (RF) and principal component analysis in each experiment. Each experiment applied the following four predictive models: decision tree, RF, logistic regression, and backpropagation neural network. The models were evaluated by measuring the accuracy, F1-score, and area under the receiver operating characteristic curve (AUROC). Results: The RF model demonstrated satisfactory performance, achieving an accuracy and F1-score of 0.920 on the testing set, with an AUROC ranging from 0.92 to 0.95. Feature selection methods enhanced model performance, particularly for the RF model. Critical features included premeal glucose levels, age, body mass index, weight, potassium, activated partial thromboplastin time, C-reactive protein level, and height. Conclusion: On average, each laparoscopic appendectomy case resulted in a deficit of NTD 3173.6. Cost prediction proved feasible using routine blood test data obtained upon admission or before surgery. The RF model and feature selection emerged as the most suitable predictive model for this specific purpose.