From Data to Decision: Development and Open Multi-Center Validation of an AI-Enabled Decision Support System for Pediatric Epilepsy Diagnosis and Treatment

Abstract:
Background/Purpose: Decision support systems (DSS) support clinical reasoning with evidence-based recommendations. In pediatric epileptology, integrating multimodal data-EEG, seizure semiology, therapy, imaging, etc.-can improve diagnostic consistency and guide therapy selection beyond EEG-only approaches. Advances in attention-based deep learning enable the detection of complex patterns in real-world data while preserving clinician oversight and alignment with ILAE standards. Our goal is to develop and validate an AI-enabled DSS that supports standardized diagnosis and treatment decision-making in children and adolescents with epilepsy, and to establish an open, scalable framework for multi-center external validation.

Methods:
Development uses an anonymized retrospective cohort of 300 pediatric patients from a tertiary epilepsy unit (KJK, KUK Linz), analyzing age at onset, seizure type, development, lateralization, family history, EEG, imaging, therapies, and reference diagnoses. A neural network is trained for multimodal diagnostic support and therapy-informed outputs. External validation compares DSS-derived diagnoses with clinician reference diagnoses using Cohen’s kappa, sensitivity, specificity, and accuracy. Secondary analyses summarize cohort characteristics and examine concordance with treatment choices and outcomes. Stratified analyses and public EEG datasets are used for bias mitigation.

Results:
Generalizability across centers, equipment, and populations will be evaluated; discrepancy analyses will inform iterative model refinement. To strengthen external validity, neuropediatric centers are invited to contribute harmonized, de-identified datasets under local ethics approvals and data-sharing agreements. Two centers are currently participating in external validation, and additional sites are being recruited.

Conclusion:
Anticipated outcomes include improved diagnostic standardization, support for treatment selection, and reduced time-to-therapy, particularly in complex cases and resource-limited settings.