Length of Stay Prediction for Diabetes Mellitus Patients on Indonesia Health Insurance Data using Logistic Regression and Process Mining

Diabetes Mellitus has a substantial impact on Indonesians, affecting 88.51% of those aged 44-94, as revealed by the Indonesia Health Insurance program. The urgency of this health concern is underscored by the International Diabetes Federations alarming statistic: one person succumbs to Diabetes Mellitus every five seconds, demanding immediate attention to diagnosis and risk reduction. This study focuses on predicting the Length of Stay (LOS) for Diabetes Mellitus patients, utilizing Logistic Regression and various machine learning models, particularly Random Forest. Concurrently, Process mining, encompassing Process Discovery and Conformance Checking, elevates model quality. Key factors influencing patient LOS, including the healthcare facilitys location, patient age, and arrival time, are unveiled. Logistic Regression exhibits commendable accuracy at 0.76 with rapid prediction times of 0.00082 seconds, while Random Forest excels with an impressive accuracy of 0.88 and a prediction time of 0.07 seconds. Specific medical procedures, such as those involving 'Diabetes and mild Nutritional/Metabolic Disorders,' reveal prolonged sojourn times, averaging 41.2 days. Best conformance-checking results emerge with a dependency threshold of 0.6 and a loop-two threshold of 32, yielding a Precision of 0.97, Trace Fitness of 0.66, and Generalization of 0.60. The finding of this study underscores the pivotal role of model selection and class distribution in LOS prediction. Despite Random Forest superior performance, Logistic Regression demonstrates notable accuracy and efficiency. Analysis of sojourn times underscores the importance of efficient healthcare services, particularly in cases involving 'Diabetes and Mild Nutritional/Metabolic Disorders.' Adjusting threshold values significantly impacts model accuracy, with lower threshold values potentially leading to overfitting.