Exploratory deep-learning-driven early risk stratification of significant neurological injury in pediatric extracorporeal membrane oxygenation

Abstract

Objective To develop and examine a multimodal deep-learning framework utilizing clinical variables with and without time-dependent electroencephalogram (EEG) spectral features for predicting significant neurological injury (SNI) in pediatric patients with extracorporeal membrane oxygenation (ECMO). Methods Data were collected from 73 pediatric ECMO patients. After excluding patients who had missing or discontinuous temporal data within 72 h after EEG initiation, 43 patients with usable EEG time-series data were analyzed. Neurological injury severity was quantified using a neuroimaging score (NIS) derived from post-cannulation imaging, with SNI defined as NIS ≥ 8. Model inputs included six clinical variables and six EEG features representing relative delta (1–4 Hz) and theta (4–8 Hz) power from the left, right, and bilateral hemispheres. Down-sampled EEG power signals were segmented into multiple 30-minute windows. A deep-learning fusion framework with modality-specific encoders was developed for SNI prediction, followed by SHapley Additive exPlanations (SHAP)-based feature attribution analysis. Results At the patient level, the proposed multilayer perceptron model used only for clinical variables achieved the highest overall performance with an area under the curve (AUC) of 74.29%. The Wilcoxon signed-rank test was performed and showed no statistical significance in model performance between the clinical-only and fusion model. SHAP analysis indicated that clinical variables were the primary contributors, not the EEG inputs, to SNI prediction. Conclusion A multilayer perceptron neural network used with six clinical inputs demonstrated moderate performance in identifying SNI among pediatric ECMO patients; integration with EEG-derived spectral features did not significantly enhance the performance. However, this conclusion is highly exploratory and dependent on selected EEG inputs, thus needing to be valid with a larger patient cohort, appropriate EEG feature selections, and assess cross-site generalizability.