Transformer-based multimodal precision intervention model for enhancing diaphragm function in elderly patients.

Area of Science:

• Critical Care Medicine
• Artificial Intelligence in Healthcare
• Biomedical Signal Processing

Background:

• Diaphragm dysfunction is a major complication in elderly patients on mechanical ventilation.
• This leads to prolonged intensive care unit (ICU) stays, difficult weaning, and increased costs.
• Current decision support systems lack real-time precision for this condition.

Purpose of the Study:

• To develop and validate a novel artificial intelligence (AI) framework for precise, real-time prediction of diaphragm dysfunction.
• To integrate multi-modal data including imaging, physiological signals, and ventilator parameters.
• To provide interpretable decision support for critical care clinicians.

Main Methods:

• A hierarchical Transformer encoder was used for modality-specific embedding extraction.
• An attention-guided cross-modal fusion module integrated diverse data streams.
• A temporal network predicted dynamic trends, validated on MIMIC-IV, eICU, and Chest X-ray datasets.

Main Results:

• The AI framework achieved high accuracy (92.3% on MIMIC-IV, 91.8% on eICU, 92.0% on Chest X-ray).
• It outperformed baseline models in precision, recall, F1-score, and Matthews correlation coefficient.
• SHAP analysis identified ventilator volume and imaging features as key predictors, with well-calibrated probability estimates.

Conclusions:

• The proposed AI framework offers precise and interpretable predictions for diaphragm dysfunction.
• It has the potential to significantly improve clinical decision-making in mechanical ventilation.
• This technology can lead to more effective, personalized interventions for high-risk ICU patients.