Fig. 5

Model prediction results of TSP-based methods vs LASSO vs random forests. Box plots of model prediction performance for DKD stage: 100 iterations of fivefold cross-validated a overall accuracy, b sensitivity, c specificity, d balanced accuracy, e positive predictive value, and f negative predictive value. Cutoff value for prediction in LASSO and random forests is 0.5. Boxplot for LASSO using residualized metabolite ions not displayed in (e) because method did not predict advanced-stage DKD even once for any patient in all iterations. Model type: (K-)TSP: (K) Top-Scoring Pair(s). LASSO: Least Absolute Shrinkage and Selection Operator. RF: Random Forests