ABSTRACT
We have previously reported on a predictive model for deficiency-excess pattern diagnosis that was unable to predict the medium pattern. In this study, we aimed to develop predictive models for deficiency, medium,and excess pattern diagnosis, and to confirm whether cutoff values for diagnosis differed between the clinics. We collected data from patients' first visit to one of six Kampo clinics in Japan from January 2012 to February 2015. Exclusion criteria included unwillingness to participate in the study, missing data, duplicate data, under 20 years old, 20 or less subjective symptoms, and irrelevant patterns. In total, 1,068 participants were included. Participants were surveyed using a 153-item questionnaire. We constructed a predictive model for deficiency, medium, and excess pattern diagnosis using a random forest algorithm from training data, and extracted the most important items. We calculated predictive values for each participant by applying their data to the predictive model, and created receiver operating characteristic (ROC) curves with excess-medium and medium-deficiency patterns. Furthermore, we calculated the cutoff value for these patterns in each clinic using ROC curves, and compared them. Body mass index and blood pressure were the most important items. In all clinics, the cutoff values for diagnosis of excess-medium and medium-deficiency patterns was > 0.5 and < 0.5, respectively. We created a predictive model for deficiency, medium, and excess pattern diagnosis from the data of six Kampo clinics in Japan. The cutoff values for these patterns fell within a narrow range in the six clinics.