NutriSighT: Interpretable Transformer Model for Dynamic Prediction of Underfeeding Enteral Nutrition in Mechanically...

Achieving adequate enteral nutrition among mechanically ventilated patients is challenging, yet critical. We develop NutriSighT, a transformer model using learnable positional encodings to predict which patients would be underfed (receive less than 70% daily caloric requirements) between days 3-7 of mechanical ventilation and compared its performance against XGBoost. Using retrospective data from two ICU databases (3284 patients from AmsterdamUMCdb for development and 6456 from MIMIC-IV for external validation), we included adults mechanically ventilated for at least 72 h. NutriSighT achieved AUROC of 0.81 (95% CI: 0.81 - 0.82) and AUPRC of 0.70 (95% CI: 0.70 - 0.72) internally. External validation yielded AUROC of 0.76 (95% CI: 0.75 - 0.76) and an AUPRC of 0.70 (95% CI: 0.69 - 0.70). In comparison, XGBoost achieved AUROC of 0.58 (95% CI: 0.58 - 0.59) and AUPRC of 0.48 (95% CI: 0.46 - 0.50). This approach may help clinicians personalize nutritional therapy in critical care.