Early laboratory-based differentiation of severe influenza pneumonia and influenza-associated encephalopathy in children using a multi-model, stability-aware machine-learning workflow: a single-center retrospective cohort study.

backgroundInfluenza-associated encephalopathy and acute necrotizing encephalopathy (ANE) are rare but devastating complications of pediatric influenza, and early differentiation from severe pneumonia without neurological involvement is challenging at admission. We aimed to develop and internally validate an early prediction model to distinguish severe influenza A pneumonia from influenza-associated central nervous system (CNS) involvement using routine admission laboratory tests. The intended clinical use is early triage: to prioritize neurological evaluation (e.g., EEG/neuroimaging when indicated), closer monitoring, and escalation of care for children predicted to be at higher risk of CNS involvement.

methodsIn this retrospective cohort at a tertiary pediatric center, consecutive children hospitalized with RT-PCR-confirmed influenza A from 1 January 2023 to 1 June 2025 were classified as mild pneumonia (Group 1), severe pneumonia without neurological involvement (Group 2), or CNS involvement (Group 3) using prespecified clinical, neurophysiological, and neuroimaging criteria. The primary prediction contrasted Group 2 versus Group 3 using routine laboratory tests obtained within 6 h of admission. Several machine-learning algorithms were trained and tuned using internal resampling and a held-out validation set. Bootstrap stability selection identified robust markers, and a parsimonious elastic-net logistic regression model was derived. We assessed discrimination, calibration, and clinical net benefit at prespecified risk thresholds (0.15, 0.25 and 0.35).

resultsAmong 227 eligible children (median age 4.2 years), 95 (41.9%) had mild pneumonia, 74 (32.6%) severe pneumonia without neurological involvement, and 58 (25.6%) influenza-associated CNS involvement; mortality was 0%, 4.1%, and 22.4%, respectively. In the primary task of distinguishing Group 2 from Group 3 using admission laboratory tests, the ten prespecified algorithms achieved out-of-fold AUCs of 0.77-0.90, and simple probability averaging achieved an AUC of 0.91. Stability selection identified 11 robust biochemical markers; a parsimonious elastic-net logistic regression model restricted to eight markers achieved a validation AUC of 0.89 with good calibration (Brier score 0.12) and consistently positive net benefit on decision-curve analysis across prespecified risk thresholds (0.15, 0.25, and 0.35).

conclusionsIn internal validation, an eight-marker model differentiated severe pneumonia without neurological involvement from influenza-associated CNS involvement with a validation AUC of 0.89 and good calibration. This stability-aware framework provides an implementable baseline for multicenter external validation and refinement, rather than a ready-to-use clinical tool.