Explainable deep-learning models to predict diaphragmatic dysfunction and cognitive stress in ICU patients under mechanical ventilation.

Background: Diaphragmatic dysfunction and acute cognitive stress/delirium are serious complications of mechanical ventilation that prolong intensive care unit (ICU) stay and are associated with increased mortality. Although accumulating evidence suggests a potential lung-brain axis linking impaired respiratory muscle function to adverse neurocognitive trajectories, early risk stratification remains challenging because clinically relevant signals are multimodal, temporally dynamic, and often partially observed. We aimed to develop an interpretable multimodal deep learning model to predict diaphragmatic dysfunction and high cognitive stress/delirium in mechanically ventilated patients and to identify shared predictors that may reflect lung-brain crosstalk. Methods: We conducted a multicenter retrospective study including 25,751 mechanically ventilated ICU patients. A multimodal long short-term memory (LSTM) network was trained using continuous clinical time-series variables (vital signs, ventilator parameters, and medication dosing trajectories) and diaphragm ultrasound videos from a sub-cohort (n = 4,783). Discrimination was evaluated using the area under the receiver operating characteristic curve (AUC) and the area under the precision-recall curve (AUPRC). Post-hoc SHapley Additive exPlanations (SHAP) were applied to quantify feature contributions and explore cross-modal interactions. Results: In the independent test set, the multimodal model outperformed both clinical-only (AUC = 0.811) and video-only (AUC = 0.749) baselines for predicting diaphragmatic dysfunction, achieving an AUC of 0.902 (95% CI, 0.88-0.92) and an AUPRC of 0.594 (both P < 0.001 vs. baselines). For high cognitive stress/delirium prediction, the model achieved an AUC of 0.792. Calibration was acceptable, with close agreement between predicted and observed risks (Brier score = 0.12). SHAP analyses indicated that ultrasound-derived diaphragm thickening fraction (DTF) and cumulative neuromuscular blockade exposure were among the strongest predictors of diaphragmatic dysfunction. Notably, diaphragm excursion and heart rate variability emerged as shared influential predictors for cognitive stress/delirium, consistent with the lung-brain crosstalk hypothesis in mechanically ventilated patients. Conclusion: This study presents a multimodal deep learning framework for early identification of mechanically ventilated patients at elevated risk of diaphragmatic dysfunction and high cognitive stress/delirium. Integrating diaphragm ultrasound with longitudinal bedside physiologic and treatment data improves prediction beyond single-modality models, while post hoc explainability highlights candidate shared predictors relevant to a lung-brain axis. These findings suggest that integrated monitoring of respiratory mechanics and sedation-related physiology may support more proactive ventilator management and neuroprotective ICU care.