(72) Machine Learning-Based Prediction Models for Outpatient Prescription of Japanese Kampo Formulations: Development and Validation Using a Health Insurance Claims Database
Background: Kampo is a traditional Japanese medicine that uses formulae of natural agents. Although Kampo products are widely used in Japan, characteristics of patients receiving different Kampo formulations have not been documented in detail.
Objectives: To develop and internally validate machine learning-based models that predict outpatient prescription of different Kampo formulations using health insurance claims database, thereby identifying factors associated with Kampo formulation use.
Methods: Ten-percent sample of enrollees in 2018 and 2019 were extracted from the JMDC Claims Database to create the training and testing sets, respectively. Logistic regression with lasso regularization were performed in the training set to construct models that predict ≥1 prescription or ≥3 prescriptions of ten commonly used Kampo formulations in one year from data of the preceding year. Models were applied to the testing set to calculate the C-statistics. We also tested the performances of simplified prediction scores using ten variables.
Results: There were 338,924 and 399,174 enrollees included in the training and testing sets, respectively. Lasso models predicting ≥1 prescription and ≥3 prescriptions showed the C-statistics ranging from 0.641 (maoto) to 0.886 (tokishakuyakusan) and 0.799 (bakumondoto) to 0.930 (kamishoyosan), respectively. The models identified common predictors of different Kampo formulations, such as female sex and previous Kampo use. The models also identified specific characteristics associated with particular Kampo formulations, such as mental and behavioral disorders for kamishoyosan, hochuekkito, and hangekobokuto. The simplified prediction scores were slightly inferior to full models.
Conclusions: Lasso regression models showed good performances in predicting various Kampo formulation prescriptions from claims data. The models identified characteristics associated with Kampo formulations uses.
