Using Historical Information of Patients for Prior Authorization Learning

Prior authorization (or preauthorization) is a control mechanism used by a Health Maintenance Organization (HMO) to minimize the waste of resources by analyzing each medical request. One of the strategies used to optimize this task is the use of a system that automates part of this process through a machine learning approach. This paper presents an approach that introduces the historical request from beneficiaries to the learning process in preauthorization in order to improve performance of the classification. For this, different approaches to data pre-processing and classification were tested. In the proposed methodology three types of historical representation were tested: binary, term frequency (TF) and term frequency-inverse document frequency (TF-IDF). For each of these representations, feature selection and transformation algorithms were applied: consistency subset eval, wrapper subset eval and Kernel PCA (KPCA). In the classification were tested three classic algorithms of machine learning: Random Forest (RF), Support Vector Machine (SVM) and K-neareast neighbors (KNN). In order to evaluate the results obtained, we used the Precision, Recall and Kappa index metrics for each of the problem classes. A comparison based on the hypothesis test Z is also performed to evaluate the improvement that the addition of the historical information has brought about to the learning process. The results point to a significant improvement in the performance of classifiers, where the representation in term frequency combined with KPCA brought the most positive results.