In this study, attempts were made to design a diagnostic model that was as objective as possible based on an analysis of quantitative data from various Oriental medical clinics. We tried to extract common criteria for SC diagnosis from a great variety of data, including practitioners’ varying points of view. Four individual diagnostic models for face, body shape, voice, and questionnaire information were obtained. Then, an integrated model was proposed that was inspired by SCM practitioners’ holistic diagnostic processes, which provided more sufficient information and which could be a more desirable approach for SC diagnosis.
In the individual training set, the diagnostic accuracy for body shape in the group of male patients was as high as 61.3%. The accuracies for face and questionnaire responses were almost the same, whereas that of voice was slightly lower. In the group of female patients, the diagnostic accuracy for face was 55.6%, followed by body shape, questionnaire, and voice, respectively, with almost an identical accuracy except for voice, as in the case of the male patients. The relatively low accuracy of the model for voice implies that it is difficult to find a stable diagnostic rule for voice. The individual diagnostic accuracies in the test set were slightly lower than those of the training set; however, the decrease in accuracy was within an acceptable range, except for the model for voice.
The diagnostic accuracies in the training set, after the four individual models had been integrated into a single model, improved to 69.3% and 63.5% in the male and female patient groups, respectively; however, when the model was applied to the test set, the accuracies decreased to 61.4% in the male patients and 52.2% in the female patients. The poor individual performance for voice was blamed, which led us to test an alternative integrated model by assigning a weight of half for voice. Although the new model showed almost the same diagnostic accuracy for the training set (68.9% in male patients and 62.3% in female patients), it clearly produced a greater accuracy in the test set (64.0% in male patients and 55.2% in female patients).
Although the resulting diagnostic power was not great, it was better than that of QSCCII
, which is 51% and has been widely used for the diagnosis of SC type
. It may be difficult to directly compare our new method to QSCCII because QSCCII was developed using only questionnaire information and was tested on limited data collected from a single site; however, an individual’s response to qualitative questions on the four diagnostic components in the QSCCII could be subjective depending on his/her own point of view. It should be noted that the integrated model using quantitative data is superior in terms of validity and objectiveness.
A more desirable result was that a higher integrated SC score corresponded to a higher diagnostic accuracy, as shown in Figure
2. Using a cut-off value for the integrated SC score, such as 1.6, the accuracies increased by 14.7% in male patients and by 4.6% in female patients in the test set.
The insufficient diagnostic accuracies can be explained as follows. First, we collected data from 23 different sites, which incurred the possibility that the data might still contain too many practitioners' subjective opinions. This fact could have affected the SC diagnostic accuracy, although the characteristics of the practitioners and the subjects were kept very strict. Evidence of there being different opinions among practitioners can be found in other research
Second, some of the variables that were extracted from face, body shape, voice, and questionnaire responses did not show clear differences among the SC types (see Additional files
20). This finding reveals that there still exists limitation to fully describing constitutional characteristics, as listed quantitatively in the SCM literature.
Third, because the Confucian culture of Korea has an influence on Korean women to modify their characteristics and their voices in talking, it is more difficult to obtain a diagnostic model in female patients than in male patients. The relatively low accuracies for questionnaire responses and voice partly explain this influence.
The proposed model was implemented in the form of a web-based prototype and is currently being tested in several clinics to get feedback from the practitioners. In the future, it will be necessary to collect more data on the TY type to complete the SC diagnostic model. Despite 2,973 samples being collected, only 1,075 samples were used, mainly due to a lack of featured extraction techniques and a lack of data quality control. It is necessary to develop an advanced technique for automatic feature extraction and to explore new feature variables for an improved diagnostic method. A different weighting method might be considered to properly reflect the importance of each diagnostic component. Furthermore, a future study may place emphasis on improving the performance for the groups which have poor diagnostic accuracies and sensitivities. At this point, we might need to analyze the typical subjects chosen by consensus from among SCM practitioners.
This study represents the first trial of integrating the objectification of SC diagnosis based on quantitative data and SCM practitioners’ holistic diagnostic processes. Although the diagnostic accuracy was not great, it is noted that the proposed diagnostic model represents common rules among practitioners who have various points of view. Our results are expected to contribute as a desirable research guide for objective diagnosis in traditional medicine, as well as to contribute to the precise diagnosis of SC types in an objective manner in clinical practice.