Abstract
Diaphragm dysfunction represents a significant complication in elderly patients undergoing mechanical ventilation, often resulting in extended intensive care stays, unsuccessful weaning attempts, and increased healthcare expenditures. To address the deficiency of precise, real-time decision support in this context, a novel artificial intelligence framework is proposed, integrating imaging, physiological signals, and ventilator parameters. Initially, a hierarchical Transformer encoder is employed to extract modality-specific embeddings, followed by an attention-guided cross-modal fusion module and a temporal network for dynamic trend prediction. The framework was assessed using three public datasets, which are, the MIMIC-IV, eICU, and Chest X-ray. The proposed model achieved the highest accuracy (92.3% on MIMIC-IV, 91.8% on eICU, 92.0% on Chest X-ray) and surpassed all baselines in precision, recall, F1-score, and Matthews correlation coefficient. Additionally, the model's probability estimates were well-calibrated, and its SHAP-based explainability analysis identified ventilator volume and key imaging features as primary predictors. The clinical implications of this study are significant. By providing precise and interpretable predictions, the proposed model has the potential to transform critical care practices by offering a pathway to more effective and personalized interventions for high-risk patients.