Integrated biomarkers and renal Doppler for early acute kidney injury diagnosis in hepatitis C virus cirrhosis

Table 6 Internal validation performance of machine learning models for acute kidney injury subtyping

Model features	Area under the curve from repeated nested cross-validation (95%CI)	Accuracy	F1-score	Brier score	Calibration slope	Sensitivity (95%CI)	Specificity (95%CI)
Neutrophil gelatinase-associated lipocalin + RRI (XGBoost)	0.92 (0.89-0.95)	86%	0.85	0.12	0.96	89% (83-93)	84% (77%-89%)
Creatinine + cystatin C	0.81 (0.75-0.86)	74%	0.73	0.18	0.91	76% (69-82)	71% (63%-78%)
RRI alone	0.78 (0.72-0.84)	72%	0.70	0.20	0.89	73% (66-80)	69% (61%-76%)

Area under the curve from repeated nested cross-validation: (1) 5-fold inner; (2) 10-fold outer; and (3) 10 repeats. Brier score: Lower is better (perfect = 0). Calibration slope: 1 = perfect calibration. Confidence intervals for area under the curve derived from bootstrapping (1000 replicates). RRI: Renal resistive index.