Answers are grounded in this paper and cite their sources.
Background: Routine laboratory blood tests are central in evaluating children with suspected appendicitis. Machine learning (ML) offers promise for improving diagnostic accuracies by predicting from variables that are non-linearly associated. We explored and internally validated ML models in their prediction of appendicitis with routine laboratory parameters in comparison with logistic regression and established clinical algorithms. Methods: We carried out a secondary analysis of a public dataset of 780 children admitted with suspected appendicitis of which 463 (59.4%) were diagnosed with appendicitis with ten laboratory predictors. A 70/30 data split was done before multiple imputations, and imputation models provided estimates using the development set only. Seven algorithms, including logistic regression and Bayesian additive regression trees (BART), were tuned by 10-fold cross-validation. Discrimination, calibration and net benefit were assessed in the held-out test set (n = 234) and compared with C-reactive protein (CRP) alone, the Pediatric Appendicitis Score (PAS) and the Alvarado score, alone and combined with the laboratory model. We examined the model explainability using SHAP and permutation importance analysis, quantified uncertainly using BART, and carried out several exploratory subgroup analyses. Results: Amongst the ML models, XGBoost achieved a test AUC of 0.785 (95% CI 0.723–0.844) with good calibration (Brier 0.182) and was comparable with logistic regression (AUC: 0.765; p = 0.13). On a common test set (n = 213), the laboratory model (AUC 0.771) outperformed CRP alone (AUC: 0.692; p < 0.001) and was not statistically significantly different from PAS (AUC: 0.721; p = 0.14) and Alvarado score (0.770; p = 0.95). However, adding it to PAS increased the AUC to 0.782 (p = 0.019 vs. PAS). The key model predictors were CRP, WBC count and neutrophil percentage. Restricting to histologically confirmed appendicitis increased the AUC for the primary model to 0.833, for the PAS to 0.741 and for the Alvarado score to 0.814. Conclusions: Routine laboratory parameters provided moderate, well-calibrated discrimination for pediatric appendicitis, were not significantly different from the Alvarado score and were complementary to the PAS. ML performance was like logistic regression. External validation in emergency department populations is required.
No takes yet. Share an insight, caveat, or question.
Sridharan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: