Modeling of infectious prediction in Human Brucellosis via metrology-driven machine learning based on routine laboratory results: a study from North China

Background:Human brucellosis shows high morbidity, severe economic losses and public health problems globally. Because of traditional cultural method shortcomings, a novel tool for assisting clinical decision to identify high-risk infectious patients is urgently required. Methods: The data of total 2283 clinically confirmed brucellosis patients (including acute phase 816/chronic phase 989) and 13093 patients with characterized healthy outcomes was collected. Models with 3 different case groups, different sizes of variables, 7 different machine learning algorithms were tested and compared for model optimization. Metrological means combined with Shapley additive explanation (SHAP) was used for model explanation. Results: The gradient boosting machine with acute phrase patients as case group achieved the highest accuracy (AUC=0.997, 95%CI 0.994-0.999), specificity/sensitivity of 89.6%/99.8% and positive predictive value/negative predictive value of 99.4%/96.7%. Finally, 16 variables based on Pearson's correlation coefficient scores and recursive feature reduction using random forest algorithm was selected for this model. The measurement uncertainty (MU) of percent basophil, direct count eosinophil, percent eosinophil in complete blood count accounted for a large proportion in all variables. Thus, the influence of each input feature for the accuracy and the generalization of our model was quantitative and visualized by MU together with SHAP. Conclusions The proposed metrology-driven artificial intelligence-basedmodel, exclusive using regular laboratory results offers a promising tool to preliminarily identify high-risk brucellosis infection patients and risk stratify patients in different population, thereby promoting the health of the patients while protecting the health of the public and overcoming financial or supply constraints, especially in rural areas.