Discrimination of prediction models between cold-heat and deficiency-excess patterns

Abstract Objective The purpose of this study was to extract important patient questionnaire items by creating random forest models for predicting pattern diagnosis considering an interaction between deficiency-excess and cold-heat patterns. Design A multi-centre prospective observational study. Setting Participants visiting six Kampo speciality clinics in Japan from 2012 to 2015. Main outcome measure Deficiency–excess pattern diagnosis made by board-certified Kampo experts. Methods We used 153 items as independent variables including, age, sex, body mass index, systolic and diastolic blood pressures, and 148 subjective symptoms recorded through a questionnaire. We sampled training data with an equal number of the different patterns from a 2 × 2 factorial combination of deficiency–excess and cold–heat patterns. We constructed the prediction models of deficiency–excess and cold–heat patterns using the random forest algorithm, extracted the top 10 essential items, and calculated the discriminant ratio using this prediction model. Results BMI and blood pressure, and subjective symptoms of cold or heat sensations were the most important items in the prediction models of deficiency–excess pattern and of cold–heat patterns, respectively. The discriminant ratio was not inferior compared with the result ignoring the interaction between the diagnoses. Conclusions We revised deficiency–excess and cold–heat pattern prediction models, based on balanced training sample data obtained from six Kampo speciality clinics in Japan. The revised important items for diagnosing a deficiency–excess pattern and cold–heat pattern were compatible with the definition in the 11th version of international classification of diseases.