Table 1 Overview of Model Performances according to 80% sensitivity.

Model	Pre/Post	Features	AUC	Threshold	MCC	Sensitivity	Specificity	PPV	NPV
SSI models
Naïve Gradient Boosting	Pre	8	0.759	0.046	0.236	0.805	0.540	0.106	0.976
Naïve Gradient Boosting	Post	56	0.857	0.041	0.308	0.805	0.761	0.185	0.983
Literature-based Gradient Boosting	Pre	12	0.762	0.044	0.208	0.805	0.615	0.124	0.979
Literature-based Gradient Boosting	Post	15	0.853	0.049	0.317	0.805	0.770	0.191	0.983
Leak models
Naïve Gradient Boosting	Pre	60	0.775	0.033	0.134	0.795	0.582	0.059	0.988
Naïve Gradient Boosting	Post	82	0.855	0.031	0.222	0.795	0.758	0.098	0.981
Literature-based Gradient Boosting	Pre	8	0.793	0.031	0.160	0.795	0.644	0.069	0.990
Literature-based Gradient Boosting	Post	10	0.861	0.034	0.233	0.795	0.774	0.104	0.991

Table depicts model performance of all pre- and postoperative SSI and leak models developed as part of this study. Naïve Gradient boosting refers to gradient boosting models where all pre-calculated features were considered, and the type and number of features were selected based on recursive feature elimination. Literature-based Gradient Boosting are gradient boosting models developed based on features reported in the literature combined with recursive feature elimination to reduce the number of features included in the models. Table contains AUC (area under the ROC curve) computed for the test set as well as specificity, PPV (positive predicted value), NPV (negative predicted value), and MCC (Matthew Correlation Coefficient) with respect to the probability threshold chosen to achieve 80% sensitivity.

Quick links

Search