Machine Learning-based Models for Outpatient Prescription of Kampo Formulations: An Analysis of a Health Insurance Claims Database

抄録

Background: Despite the widespread practice of Japanese traditional Kampo medicine, the characteristics of patients receiving various Kampo formulations have not been documented in detail. We applied a machine learning model to a health insurance claims database to identify the factors associated with the use of Kampo formulations.

Methods: A 10% sample of enrollees of the JMDC Claims Database in 2018 and 2019 was used to create the training and testing sets, respectively. Logistic regression analyses with lasso regularization were performed in the training set to construct models with prescriptions of 10 commonly used Kampo formulations in 1 year as the dependent variable and data of the preceding year as independent variables. Models were applied to the testing set to calculate the C-statistics. Additionally, the performance of simplified scores using 10 or 5 variables were evaluated.

Results: There were 338,924 and 399,174 enrollees in the training and testing sets, respectively. The commonly prescribed Kampo formulations included kakkonto, bakumondoto, and shoseityuto. Based on the lasso models, the C-statistics ranged from 0.643 (maoto) to 0.888 (tokishakuyakusan). The models identified both the common determinants of different Kampo formulations and the specific characteristics associated with particular Kampo formulations. The simplified scores were slightly inferior to full models.

Conclusion: Lasso regression models showed good performance for explaining various Kampo prescriptions from claims data. The models identified the characteristics associated with Kampo formulation use.