Age-stratified machine learning using de-identified clinical and transcriptomic data for pediatric appendicitis

Pediatric appendicitis triage remains clinically challenging, as standard single-model diagnostic scores yield moderately discriminative performance.

We developed an innovative age-stratified machine learning routing architecture integrating clinical biomarkers with transcriptomic data for three-category severity triage (complicated, uncomplicated, and negative), accounting for age-dependent developmental immunological differences.

We analyzed 200 pediatric patients (aged 1-17 years; 103 males, 97 females) from de-identified secondary datasets. Clinical features (n=18, including age, C-reactive protein [CRP], leukocyte differentials, and appendiceal diameter) were integrated with 56,666 transcriptomic features via temporal cohort alignment without individual identifiers. Following a 160/40 train-test split, a baseline unstratified model and an age-stratified multi-model system (AgeAwarePredictor) were trained using Bayesian-optimized gradient boosting with training-set BorderlineSMOTE balancing.

On the held-out test set (N=40), the baseline global model achieved an overall accuracy of 0.700 (95% confidence interval [CI]: 0.548-0.825) and a macro-averaged area under the precision-recall curve (Macro-AUPRC) of 0.753 (95% CI: 0.620-0.850), a Matthews correlation coefficient of 0.555, and a mean Brier score of 0.190. Incorporating the exploratory age-stratified routing framework improved overall accuracy to 0.875 and Macro-AUPRC to 0.865. Pathway sub-analyses showed descriptive numerical superiority in adolescents (13-17 years, n=22: accuracy 0.909, AUPRC 0.882) over younger children (7-12 years, n=18: accuracy 0.833, AUPRC 0.847; Δ+7.6% accuracy). SHapley Additive exPlanations analysis identified neutrophils, immature granulocytes, and CRP as key predictors.

Compared to global unstratified modeling, an age-stratified multi-modal framework significantly improves pediatric appendicitis classification. These exploratory subgroup findings require external multicenter validation.