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

Area of Science:

• Critical Care Medicine
• Artificial Intelligence in Medicine
• Respiratory Physiology

Background:

• Diaphragmatic dysfunction and delirium are significant complications of mechanical ventilation, increasing mortality and intensive care unit (ICU) stay.
• Early risk stratification is challenging due to the multimodal and dynamic nature of clinical data.
• A potential lung-brain axis links respiratory muscle function to neurocognitive outcomes, but requires further investigation.

Purpose of the Study:

• To develop an interpretable multimodal deep learning model for predicting diaphragmatic dysfunction and high cognitive stress/delirium in mechanically ventilated patients.
• To identify shared predictors indicative of lung-brain crosstalk.

Main Methods:

• A multicenter retrospective study involving 25,751 mechanically ventilated ICU patients.
• A multimodal long short-term memory (LSTM) network trained on continuous clinical time-series data and diaphragm ultrasound videos.
• Post-hoc SHapley Additive exPlanations (SHAP) for quantifying feature contributions and exploring cross-modal interactions.

Main Results:

• The multimodal model achieved an AUC of 0.902 for diaphragmatic dysfunction prediction, outperforming clinical-only and video-only models.
• The model achieved an AUC of 0.792 for high cognitive stress/delirium prediction.
• SHAP analysis identified diaphragm thickening fraction (DTF) and neuromuscular blockade as key predictors for diaphragmatic dysfunction, while diaphragm excursion and heart rate variability were shared predictors for cognitive stress/delirium.

Conclusions:

• A multimodal deep learning framework effectively identifies patients at risk for diaphragmatic dysfunction and delirium.
• Integrating diaphragm ultrasound with clinical data enhances predictive accuracy compared to single-modality approaches.
• Findings support the lung-brain axis hypothesis and suggest integrated monitoring for proactive ventilator management and neuroprotection.