Background The purpose of this study is to evaluate the predictive accuracy of a machine learning-based herbal medicine response prediction algorithm in patients with functional dyspepsia (FD). In a preliminary clinical study, the algorithm was developed using the XGBoost regressor framework to predict the relative effect sizes of three commonly prescribed herbal formulations— Yijung-tang ( Lizhong-tang ), Pyeongwi-san ( Pingwei-san ), and Shihosogan-tang ( Chaihu Shugan-tang ). The prediction system recommends the formulation expected to yield the greatest therapeutic benefit for each individual. Methods This study is a randomized, assessor-blinded, parallel-group, open-label, multicenter clinical trial. A total of 100 patients with FD will be recruited from two Korean medical hospitals and randomly assigned to either the ACCORD group ( n = 50), which will receive treatment guided by the machine learning algorithm, or the DISCORD group ( n = 50), which will receive one of the two treatments not recommended by the algorithm. Patients will take the assigned herbal medicine for 8 weeks, three times daily, between meals. Outcomes The primary outcome will be gastrointestinal symptom score. Secondary outcomes will include total dyspepsia symptom score, adequate relief of dyspepsia, overall treatment effect, visual analog scale score, functional dyspepsia–related quality of life, and pattern identification questionnaire results. Exploratory outcomes will include blood and fecal metabolome analysis, fecal and salivary microbiota profiling, and measurements obtained using Korean medicine diagnostic devices (heart rate variability, tongue, pulse, and abdominal diagnosis). Conclusion Integrating a machine learning-based prediction system into treatment strategies for FD may enhance clinical practice and support the broader adoption of artificial intelligence-driven approaches in personalized medicine. Clinical trial registration Clinical Research information Service (registration number: KCT0010587) and Open Science Framework ( https://osf.io/2ecz8 ).
Park et al. (Wed,) studied this question.