Abstract
KDIGO stage-3 acute kidney injury (AKI), a life-threatening complication in critically ill patients with traumatic cervicothoracic spinal cord injury (TCTSCI), was associated with a 49.3% 60-day mortality and a median survival of 20 days in a combined MIMIC-IV/eICU analysis, underscoring its severe clinical consequences and the need for early identification and prediction. To address this need, MedFusion-GP-AKI was developed as a multimodal deep learning framework trained on the MIMIC-IV/eICU cohort and externally validated in 188 patients from four tertiary Chinese centers. Missing data were imputed with a GAN-based method, and key predictors were derived from the original dataset using NOTEARS, variational bottleneck, and adversarial analysis, yielding eleven variables led by lactate, mean arterial pressure, temperature, potassium, and TCTSCI level, with the dataset subsequently balanced using an SMOTified-GAN. Fifteen baseline models were benchmarked under uniform protocols, and the best-performing architectures were integrated into an ensemble that achieved AUCs of 0.938, 0.909, 0.969, 0.945, and 0.921 with APs of 0.841, 0.884, 0.992, 0.927, and 0.878 across pre- and post-SMOTE training, validation, and external cohorts, demonstrating reliable discrimination and calibration, stable clinical net benefit across thresholds, and balanced overall classification performance with strong generalizability across independent institutions. SHAP analysis confirmed that model attributions aligned with known clinical and physiological patterns, and a web-based calculator was developed for practical use. Overall, this study connects artificial intelligence, nephrology, and critical care by using multimodal deep learning and causal inference to predict severe AKI occurrence from early clinical data in critically ill TCTSCI patients.