Mamba-Bi-LSTM with SHAP-Guided Iterative Refinement for Multimodal ARDS Diagnosis: A Dual-System Framework

Area of Science:

• Artificial Intelligence in Medicine
• Critical Care Medicine
• Machine Learning for Healthcare

Background:

• Acute respiratory distress syndrome (ARDS) presents diagnostic challenges due to overlapping symptoms, leading to high mortality rates.
• Early diagnosis of ARDS is crucial for timely intervention and improved patient outcomes.
• Current diagnostic methods often lack the sensitivity and specificity for early detection in complex ICU environments.

Purpose of the Study:

• To develop and validate a multimodal AI framework for accurate and early diagnosis of ARDS.
• To integrate heterogeneous clinical data, including ventilator parameters, blood gas indices, chest imaging, and EEG signals.
• To enhance diagnostic performance and extend the early warning time for ARDS.

Main Methods:

• A dual-system framework combining a Mamba-Bi-LSTM primary discrimination system with a TreeSHAP-based verification system.
• Integration of multimodal data streams processed through selective state-space Mamba and bidirectional LSTM layers.
• Iterative offline refinement using TreeSHAP attribution outputs to modulate feature selection gates and a confidence-gated protocol.

Main Results:

• Achieved 92.8% accuracy and 0.889 F1 score on internal validation (MIMIC-IV) and 91.6% accuracy with 0.871 F1 on external validation (eICU).
• Extended early warning time for ARDS from 5.2 to 9.7 hours.
• EEG integration demonstrated a significant 2.7 percentage point accuracy gain and 1.9-hour warning extension (p<0.001).

Conclusions:

• The proposed dual-system AI framework offers a robust solution for early and accurate ARDS diagnosis.
• Multimodal data integration, particularly EEG, significantly enhances diagnostic performance and extends the critical warning window.
• The system's low latency (350 ms) supports real-time deployment in intensive care units for improved patient management.