Interpretable Modelling of a Streptococcal Exposure-Associated Neuropsychiatric Profile in a Pediatric Cohort

Pediatric neuropsychiatric presentations in the context of suspected previous streptococcal exposure remain clinically heterogeneous. This study examined whether a predefined clinical-serological profile could be reconstructed within a naturalistic psychiatric cohort using interpretable supervised models and whether discrimination persisted after removing endpoint-proximal predictors.

This secondary exploratory analysis included 96 children and adolescents evaluated in Romanian psychiatric inpatient and outpatient settings. The operational endpoint required PANS-31 > 36, antistreptolysin O (ASO) > 200 IU/mL, abrupt symptom onset or worsening, and a reported streptococcal episode within six months; 24 participants met all four conditions. Infection history was derived from caregiver reports and available medical documentation. Logistic Regression, Random Forest, and XGBoost were evaluated using stratified five-fold cross-validation and a 70/30 training-test split. Full-feature models included 27 predictors, whereas restricted-feature models were refitted after excluding continuous ASO, infection recency, and abrupt onset. SHAP was used to examine feature attribution.

Full-feature models achieved held-out ROC-AUC values of 0.909-0.961, compared with 0.630-0.792 after feature restriction. The held-out sample comprised 29 participants, including seven endpoint-positive cases. Full-feature attribution emphasized serological, temporal, and onset-related variables, whereas restricted-feature attribution shifted toward symptom-domain, age-related, and contextual variables.

High full-feature discrimination primarily reflected internal reconstruction of the rule-defined profile rather than independent diagnostic prediction. The attenuation following feature restriction indicates substantial dependence on information related to endpoint construction. These exploratory findings do not establish streptococcal causality, disease specificity, or generalizable diagnostic performance.