Fig. 3

Machine learning algorithms can predict lupus endotype membership with high accuracy

A Eight final endotypes of lupus were determined by k-means clustering of the 3166 patients’ concatenated GSVA enrichment scores of 26/32 features. Multi-class classification by ML categorized 3166 lupus patients from 17 datasets into eight patient endotypes. Area under the ROC curve (AUC), performance metrics, and confusion matrices for each of 4 classifiers on the testing cohort data (983 samples) are summarized: B random forest and C support vector machine. Each model was trained on 1746 lupus samples, validated with 437 lupus samples, and tested on the remaining 983 samples for a total n = 3166 from 17 datasets. D RF classification of the 983 samples into the eight endotypes. Heatmaps in A and D were generated in R with the ComplexHeatmap package. Plots in B, C were generated in Python using the scikit-learn and matplotlib libraries