Multiplex analysis of 40 cytokines do not allow separation between endometriosis patients and controls

Abstract

Endometriosis is a common gynaecological condition characterized by severe pelvic pain and/or infertility. The combination of nonspecific symptoms and invasive laparoscopic diagnostics have prompted researchers to evaluate potential biomarkers that would enable a non-invasive diagnosis of endometriosis. Endometriosis is an inflammatory disease thus different cytokines represent potential diagnostic biomarkers. As panels of biomarkers are expected to enable better separation between patients and controls we evaluated 40 different cytokines in plasma samples of 210 patients (116 patients with endometriosis; 94 controls) from two medical centres (Slovenian, Austrian). Results of the univariate statistical analysis showed no differences in concentrations of the measured cytokines between patients and controls, confirmed by principal component analysis showing no clear separation amongst these two groups. In order to validate the hypothesis of a more profound (non-linear) differentiating dependency between features, machine learning methods were used. We trained four common machine learning algorithms (decision tree, linear model, k-nearest neighbour, random forest) on data from plasma levels of proteins and patients' clinical data. The constructed models, however, did not separate patients with endometriosis from the controls with sufficient sensitivity and specificity. This study thus indicates that plasma levels of the selected cytokines have limited potential for diagnosis of endometriosis.

Figure 1

Flowchart of patient recruitment.

Figure 2

Box plots comparing plasma levels of the three cytokines that differ between the control group of patients and patients with peritoneal endometriosis in the univariate analysis. Plasma levels of cytokines are presented as Tukey box-and-whiskers plots with median, the box from the 25^th to 75^th percentiles, and whiskers correspond to the 25^th percentile minus 1.5 times IQR (interquartile range) and to the 75^th percentile plus 1.5 IQR. After correction for multiple testing no statistical difference (ns) was observed. Plasma concentrations of the cytokines are represented on a logarithmic scale. C, controls; PE, peritoneal endometriosis.

Figure 3

Principal component analysis plot. Data from the protein concentrations were scaled and normalized. The PCA plot is based on the whole protein set and coloured according to the disease status (red circles - patients with endometriosis; blue squares - controls). Transformed data show no meaningful grouping between patients with endometriosis and controls.

Figure 4

Averaged classification performance of four classifiers and box plots for the selected features that were used for training a random forest model. (A) Four different classifiers were used based on the data from the training set with the highest average classification performance (i.e. accuracy) achieved with random forest (balanced accuracy of ~59%). (B) Box plots of the six most important features that were used for training a RandomForest model based on the training set and were the most differential between patients with endometriosis and controls. Red color designate patients with endometriosis and blue color controls. Machine Learning models used: glmnet, elastic-net regularized generalized models; kknn, Weighted k-Nearest Neighbors; rpart, Recursive Partitioning and Regression Trees; rf, RandomForest. Dashed red line indicates expected balanced accuracy of a random chance.

Figure 5

Modelling results from the recursive feature elimination method. (A) Each dot that forms curves was chosen automatically by the random forest algorithm trained on the number of protein features specified by x-axis. The best performance was achieved by random forest that was trained on all 1444 protein concentrations or ratios of protein concentrations remaining after pre-processing which achieved AUC of 0.585 for all samples with a sensitivity of 34% and specificity of over 70%. (B) ROC curve based on the highest values of sensitivity, specificity and AUC. ROC, receiver operating characteristic; AUC, area under the curve.

Figure 6

Modelling results after inclusion of metadata. (A) Random forest achieved the highest balanced accuracy on average (~0.55). (B) Both protein and metadata features ranked by their relative importance for RandomForest predictive performance. The last step was to evaluate if there is any clear separation between patients within individual types of endometriosis and controls with the inclusion of metadata. Results also showed that there is no improvement of discriminating performance of the classifiers if we look into individual type of endometriosis.

Figure 7

Modelling results after inclusion of metadata for individual types of endometriosis and controls. Nested Cross-validation results for three machine learning methods on patients with ovarian (A,B) peritoneal (C,D) and deep infiltrating endometriosis (E,F) and control samples. 5-fold CV was repeated 10 times without any parameter learning or sharing allowed between the folds to ensure generalisation and robustness of the obtained models. Results suggest that machine learning models cannot differentiate between different types of endometriosis and controls with an accuracy that exceeds the one of random chance.