Earl John T. Conag, Marc Lloyd S. Quisel, Glenn B. Pepito
Medical coding is a tedious and complex process that may lead to problematic scenarios such as upcoding which forces insurance companies to pay more than what they had to. As discussed by Lopez [1] in an article, this issue was notably exemplified in the Philippines in 2020 when PhilHealth, a government health insurance corporation, was overcharged by three private healthcare providers in Zamboanga del Norte, Koronadal City, and Davao del Norte due to upcoding. Medical data with its complexity and varied sample sizes possess a challenge for machine learning models. With this the researchers aimed to create an ensemble model composed of convolutional neural networks in classifying international classification of diseases (ICD) codes with the application of stratified k-fold, to address instances where single CNN can perform poorly due to biases in the dataset. Discharge summaries under 3 different recurring codes from the MIMIC-III database were used as the dataset. Due to data imbalances between classes, Stratified K Fold strategy was then used to create 10 folds from the dataset to ensure robustness. These 10 folds allowed the researchers to construct an ensemble model comprising 10 convolutional neural networks. The researchers found out that the ensemble model performs slightly better in classifying ICD codes than a single convolutional neural network with an F1 score of 0.99 whereas a single CNN model with a fine tuned hyperparameter had an accuracy of just 96%. The findings in this study can be used as additional knowledge in the continued use of ensemble models in ICD code classification. © 2023 IEEE.
University of San Carlos, Department of Computer, Information Sciences, and Mathematics, Cebu City, Philippines
