TABLE 2.

Cross-Validated Area Under the Receiver Operating Curve (cv-AUC) for Incident Sexually Transmitted Infections (STIs)^* With ≥1 Repeat STI Diagnoses or ≥2 Repeat STI Diagnoses Stratified by Duration of Follow-Up, Atrius Health, 2008 to 2015

	730 d of Follow-Up		365 d of Follow-Up
Algorithm	Incident STIs With ≥2 Repeat STI, cv-AUC	Incident STIs With ≥1 Repeat STI, cv-AUC	Incident STIs With ≥2 Repeat STI, cv-AUC	Incident STIs With ≥1 Repeat STI, cv-AUC
SL	0.76	0.66	0.70	0.64
BART	0.75	0.67	0.72	0.65
LASSO	0.73	0.67	0.67	0.64
XGB2,50	0.73	0.64	0.69	0.61
XGB4,50	0.73	0.62	0.69	0.59
XGB2,25	0.72	0.65	0.70	0.63
NN2,screen	0.72	0.65	0.56	0.63
XGB2,10	0.71	0.64	0.65	0.62
NN1,screen	0.71	0.65	0.51	0.63
XGB4,25	0.70	0.62	0.70	0.59
GLMscreen	0.70	0.65	0.58	0.63
RIDGE	0.70	0.65	0.63	0.63
XGB4,10	0.67	0.61	0.62	0.58
SVM1.5,screen	0.55	0.48	0.58	0.49
SVM1,screen	0.52	0.48	0.57	0.50
GLM	0.51	0.57	0.52	0.49

^*Defined as positive laboratory result for chlamydia, gonorrhea, or a syphilis diagnosis.

BART indicates Bayesian Additive Regression Trees (200 trees, 2-way interactions); GLM, generalized linear model logistic regression; LASSO, least absolute shrinkage and selection operator using AUC loss; NN, neural net with 1 or 2 nodes in the hidden layer; RIDGE, ridge regression using AUC loss; “screen,” covariates were prescreened to include only those whose correlation with the outcome had magnitude at least 0.2, although a minimum of 20 covariates were retained; SL, super learner; SVM, support vector machine (radial kernel with cost, 1 or 1.5); XGB, eXtreme Gradient Boosting (1000 trees; shrinkage, 0.1; maxdepth, 2 or 4 [interaction depth], 10, 25, or 50 minimum number of observations in terminal node).