Explainable Deep-Learning Models to Predict Diaphragmatic Dysfunction and Cognitive Stress in ICU Patients Under Mechanical Ventilation

YONGHUA WANGYulingBai baiGE JIN^*

• First Affiliated Hospital of Zhengzhou University, Zhengzhou, China

Background: Diaphragmatic dysfunction and acute cognitive stress/delirium are major complications of mechanical ventilation, prolonging ICU stay and increasing mortality. Evidence supports a lung–brain axis linking respiratory muscle impairment to adverse neurocognitive trajectories, but early risk stratification is limited by multimodal, temporally dynamic, partially observed clinical signals. We aimed to develop an interpretable multimodal deep learning model to predict diaphragmatic dysfunction and high cognitive stress/delirium and to identify shared predictors reflecting lung–brain crosstalk. Methods: We conducted a multicenter retrospective study including 25,751 mechanically ventilated ICU patients. A multimodal LSTM network was trained using continuous clinical time-series variables (vital signs, ventilator parameters, and medication dosing trajectories) integrated with diaphragm ultrasound videos from a sub-cohort (n = 4,783). Discrimination was evaluated using AUC and AUPRC. Calibration was assessed using the Brier score. Post-hoc SHAP quantified feature importance and cross-modal interactions. Results: In the independent test set, the multimodal model outperformed clinical-only (AUC = 0.811) and video-only (AUC = 0.749) baselines for diaphragmatic dysfunction, achieving AUC = 0.902 (95% CI, 0.88–0.92) and AUPRC = 0.594 (P < 0.001 vs. baselines). For high cognitive stress/delirium, the model achieved AUC = 0.792 with acceptable calibration (Brier score = 0.12). SHAP identified ultrasound-derived DTF and cumulative neuromuscular blockade exposure as top predictors for diaphragmatic dysfunction. Diaphragm excursion and heart rate variability were most influential for cognitive stress/delirium. These two features also emerged as shared predictors across both tasks, supporting the lung–brain crosstalk hypothesis and suggesting physiologic coupling between respiratory mechanics, autonomic stress, ventilator dependence, and neurocognitive vulnerability. Longitudinal medication exposure patterns aided temporal risk differentiation. Conclusion: We present an interpretable multimodal deep learning framework for early identification of mechanically ventilated patients at elevated risk of diaphragmatic dysfunction and high cognitive stress/delirium. Integration of diaphragm ultrasound with longitudinal physiologic and treatment data improves prediction beyond single-modality models. SHAP explainability highlights candidate shared predictors relevant to lung–brain axis interactions and supports proactive ventilator management and neuroprotective ICU care.