Objective

2

To construct a diagnostic model for coronary heart disease （CHD） with blood stasis syndrome by integrating objective multi-modal data from the four examinations of traditional Chinese medicine （TCM）， thereby enhancing the accuracy and objectivity of TCM syndrome differentiation.

Methods

2

A total of 426 patients with CHD were enrolled and divided into blood stasis syndrome group （221 cases） and non-blood stasis syndrome group （205 cases）. Different devices were used to collect tongue image， pulse manifestation， and auscultation information of CHD patients， so as to establish an objective TCM four-diagnostic database for CHD. Meanwhile， clinical questionnaires were used for manual information collection. In single-modal analysis， pulse diagnosis data were discriminated using principal component analysis （PCA）， least squares regression （LS）， and Lasso regression； tongue diagnosis data were classified using a convolutional neural network； manually collected information data were classified using a random forest. For multi-modal modeling， deep learning was used to splice multi-source information， such as pulse manifestations and tongue images， with inquiry data from CHD patients with blood stasis syndrome through a fully connected layer， thereby constructing the diagnostic model for CHD with blood stasis syndrome.

Results

2

The single-modal objective analysis showed that in the pulse analysis， there were significant differences in pulse diagnosis information between the blood stasis syndrome group and the non-blood stasis syndrome group. The PCA difference was mainly in the first principal component， with a discrimination accuracy of 99% （

P

=0.010）； the classification accuracy of LS and Lasso was 64% （

P

=0.360） and 67% （

P

=0.400）， respectively. Compared with the non-blood stasis syndrome group， the time-domain parameter differences in the blood stasis syndrome group were mainly manifested as a decrease in the left-hand

Cun

pulse force h1 （

P

=0.018）， and increases in the left-hand

Guan

pulse S1/S2 （

P

=0.008） and right-hand

Chi

pulse t1 （

P

=0.002）； the frequency-domain parameter differences were particularly significant in the left-hand

Chi

pulse and right-hand

Guan

pulse. In the tongue analysis， the average vascular width and variability of sublingual collaterals in the blood stasis syndrome group were significantly higher than those in the non-blood stasis syndrome group. In the analysis of manually collected information， the clinical features with the highest discriminative value for blood stasis syndrome included distended and tortuous sublingual collaterals， chest pain， chest tightness， rough pulse， palpitations， and shortness of breath. In addition， among the single-modal analysis in the blood stasis syndrome group， the tongue diagnosis achieved the highest discrimination accuracy of 83.33% （30/36）， while the accuracy of pulse， auscultation， and inquiry diagnosis was lower than that of tongue diagnosis. The optimal modal combination （tongue diagnosis + pulse diagnosis + inquiry + auscultation） achieved the highest accuracy of 86.11% （31/36）， which was superior to the single-modal model.

Conclusion

2

The diagnostic model based on multi-modal data fusion can effectively achieve objective differentiation of coronary heart disease with blood stasis syndrome， validating the advantage of multi-modal information fusion in improving TCM syndrome differentiation accuracy， and providing scientific evidence for objective research of TCM syndrome diagnosis.