Machine Learning Model to Differentiate Between Acute Kidney Injury and Functional Decline in Children with Urinary Tract Infection

eMethods described the machine learning approach using the Python XGBoost package. Five hyperparameters were noted: max_depth, learning_rate (eta), booster, reg_alpha, and rea_lambda. The booster used was gbtree. In experiments, only max_depth and learning_rate (eta) were actively tuned, as these provided the best performance across tests and were used to control overfitting. An imbalance issue was highlighted: XGBoost, like other standard machine learning methods, experiences performance decline when label ratios are skewed. Although XGBoost can still perform well on skewed datasets when combined with ensembling, the decay necessitates revising XGBoost or employing strategies to handle class imbalance to ensure strong performance. Subgroup modeling process: In a subgroup analysis, febrile controls (FC) with C-reactive protein (CRP) >30 mg/L (3 mg/dL; AHA guideline cut-off in evaluating incomplete Kawasaki disease, KD) were designated as New FC. From 73,499 original FC, 13,256 with CRP >30 mg/L and all 1,142 KD were included. Two new prediction models were built using the same machine-learning strategy as the original model: (1) a model using only the top-5 important features (CRP, ALT, urinalysis, and eosinophil count) and (2) a model using all features.

- Misclassification in the testing set: Among 14,929 FC, 17 of 228 KD (7.5%) were misclassified as FC; 400 of 14,701 FC (2.7%) were misclassified as KD. - Characteristics of misclassified groups (means with 95% CI for mean difference or odds ratios; groups: KD misclassified as FC [N=17] vs FC misclassified as KD [N=400]): - Age (years): 1.34 vs 1.18; mean diff 1.21 (0.72 to 2.03) - Male, %: 52.9% vs 53.8%; OR 0.91 (0.33 to 2.53) - Pyuria, %: 29.4% vs 33.5%; OR 0.84 (0.27 to 2.61) - WBC (10^3/µL): 9.91 vs 13.80; diff 3.90 (0.90 to 6.89) - RBC (10^6/µL): 4.43 vs 4.42; diff -0.01 (-0.28 to 0.26) - Hemoglobin (g/dL): 11.77 vs 11.26; diff -0.51 (-1.07 to -0.06) - Hematocrit (%): 34.92 vs 33.82; diff -1.10 (-2.68 to -0.49) - MCH (pg/cell): 26.69 vs 25.68; diff -1.01 (-2.30 to -0.29) - MCHC (g/dL): 33.71 vs 33.30; diff -0.41 (-0.93 to -0.11) - RDW: 13.54 vs 13.63; diff -0.09 (-0.84 to 1.02) - Platelets (10^3/µL): 316.06 vs 322.46; diff 6.40 (-58.79 to 71.59) - Segment (%): 52.55 vs 50.36; diff -2.19 (-10.86 to 6.49) - Band (%): 0.53 vs 0.87; diff 0.33 (-0.56 to 1.23) - Lymphocyte (%): 36.34 vs 38.03; diff 1.68 (-6.36 to 9.73) - Monocyte (%): 7.62 vs 7.94; diff 0.32 (-1.56 to 2.20) - Eosinophil (%): 1.49 vs 1.47; diff -0.03 (-1.05 to 1.00) - Basophil (%): 0.12 vs 0.23; diff 0.10 (-0.08 to 0.29) - AST (U/L): 54.11 vs 55.27; diff 1.15 (-44.38 to 46.68) - ALT (U/L): 46.06 vs 40.81; diff -5.25 (-42.33 to 31.83) - CRP (mg/L): 25.94 vs 48.54; diff 22.60 (-4.19 to 49.39) - Urine WBC (count/hpf): 63.59 vs 44.21; diff -19.38 (-71.54 to 32.79) - Subgroup analysis performance (New FC: CRP >30 mg/L) at targeted sensitivities using models with top-5 features vs all features: - Sensitivity target >80%: - Top-5: Sens 81.6%, Spec 95.2%, PPV 59.2%, NPV 98.4%, +LR 16.9, -LR 0.20 - All features: Sens 81.6%, Spec 99.1%, PPV 88.2%, NPV 98.4%, +LR 86.5, -LR 0.19 - Sensitivity target >85%: - Top-5: Sens 85.1%, Spec 92.8%, PPV 50.3%, NPV 98.6%, +LR 11.8, -LR 0.20 - All features: Sens 85.1%, Spec 98.3%, PPV 81.5%, NPV 98.7%, +LR 51.3, -LR 0.15 - Sensitivity target >90%: - Top-5: Sens 90.4%, Spec 90.6%, PPV 45.3%, NPV 99.1%, +LR 9.6, -LR 0.10 - All features: Sens 90.4%, Spec 97.0%, PPV 72.0%, NPV 99.2%, +LR 30.0, -LR 0.10 - Sensitivity target >95%: - Top-5: Sens 95.2%, Spec 87.6%, PPV 39.8%, NPV 99.5%, +LR 7.7, -LR 0.05 - All features: Sens 95.2%, Spec 94.6%, PPV 60.4%, NPV 99.6%, +LR 17.8, -LR 0.05

The subgroup analysis indicates that excluding FC children with low CRP (retaining only CRP >30 mg/L) maintains strong model performance. Using all features improves specificity and PPV across sensitivity targets compared with using only the top-5 features, suggesting that incorporating the full feature set enhances performance in this clinical discrimination task.