Table 3 Comparison of performance of determining fasting status by XGBoost, CatBoost, H₂O Ensemble and logistic regression models in the testing dataset (n = 70,644).

Algorithm/modeling strategy	Feature	Sensitivity	Specificity	Precision	F1-score	Accuracy	AUC
Parsimonious modeling
Logistic regression	Model 2*	0.7608	0.8084	0.8081	0.7804	0.7845	0.868 (0.865–0.870)
XGBoost	Model 2*	0.8261	0.7700	0.7844	0.8047	0.7982	0.887 (0.885–0.890)
CatBoost	Model 2*	0.8415	0.7614	0.7813	0.8103	0.8017	0.889 (0.887–0.892)
H2O Ensemble	Model 2*	0.8823	0.7093	0.7546	0.8135	0.7964	0.886 (0.884–0.889)
Full modeling
XGBoost	67	0.8394	0.7785	0.7934	0.8158	0.8092	0.896 (0.894–0.898)
CatBoost	67	0.8511	0.7574	0.7805	0.8142	0.8046	0.892 (0.890–0.894)
H2O Ensemble	67	0.8770	0.7399	0.7735	0.8220	0.8089	0.897 (0.894–0.899)
Feature selection modeling
XGBoost	Top 45	0.8369	0.7789	0.7932	0.8145	0.8081	0.895 (0.892–0.897)
XGBoost	Top 35	0.8413	0.7735	0.7901	0.8149	0.8076	0.894 (0.892–0.897)
XGBoost	Top 25	0.8414	0.7706	0.7880	0.8138	0.8062	0.893 (0.891–0.896)
XGBoost	Top 10	0.8502	0.7496	0.7748	0.8108	0.8002	0.887 (0.885–0.890)

*Model 2 involves the features including glucose, age, male, timing of the day, interval between request and sampling, No. of outpatient visits, distance from home to hospital, division, hypertension, diabetes, coronary artery disease, stroke, statin use, and concomitant lipid testing as in Table 2.

Quick links

Search