Evaluation of inflammatory serum parameters as a diagnostic tool in patients with endometriosis: a case-control study.

Even though non-invasive prediction of endometriosis may seem technically feasible using sophisticated machine learning algorithms, a standard clinical use case for non-surgical diagnosis of endometriosis has not yet been established. In the present paper, we assess the potential of the inflammatory serum markers hepcidin, soluble urokinase-type plasminogen activator receptor (suPar), and interleukin-6 (IL-6) in a cohort of 87 patients. Hereby, 59 patients were histologically diagnosed with endometriosis, whereas other 28 patients served as our non-endometriosis control group. An initial exploratory univariate statistical analysis (Mann-Whitney test) revealed the diagnostic potential of different serum levels of suPar (p = 0.024) and IL-6 (p < 0.001) between both groups; the formation of a distinct training data set (n = 77) subsequently allowed to train a supervised machine learning analysis (tree classifier) employing serum levels of suPar, hepcidin, and IL-6 as predictor variables. Based on an internal 5-fold cross validation, the classifier performance was initially assessed using standard metrics such as sensitivity, positive predictive value, and AUROC curve. Additionally, the algorithm was tested on an external validation (holdout) data set (n = 10), showing sufficient overall accuracy of 80% without tendencies of overfitting. In conclusion, our data demonstrates the diagnostic potential of IL-6 and suPar as pro-inflammatory serum biomarkers in endometriosis. Using a decision tree-based supervised learning approach, we additionally present a straight-forward way of a potential clinical employment, aiming at less invasive (non-surgical) diagnosis.