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

Summary

This study developed a dual-system multimodal model combining Mamba-Bi-LSTM prediction with TreeSHAP-guided iterative feature refinement. In 3,742 held-out MIMIC-IV patients, the system achieved 92.8% accuracy and an F1 score of 0.889; external validation in 2,594 eICU patients achieved 91.6% accuracy and an F1 score of 0.871. The warning window increased from 5.2 to 9.7 hours, while EEG independently improved accuracy and warning time.

Key Findings

• The complete system achieved 92.8% accuracy and an F1 score of 0.889 on 3,742 held-out MIMIC-IV patients.
• External validation in 2,594 eICU patients achieved 91.6% accuracy and an F1 score of 0.871.
• Offline iterative refinement extended the early-warning interval from 5.2 to 9.7 hours; EEG integration added 2.7 percentage points in accuracy and 1.9 hours of warning time.

Clinical Implications

The system could support ICU surveillance and earlier recognition of ARDS, potentially enabling timely lung-protective management and escalation of care. It should not yet replace clinician assessment because the study used retrospective databases and did not evaluate prospective patient outcomes or workflow effects.

Limitations

• The retrospective database design may include selection bias, label inconsistency, and dataset-specific spectrum effects.
• Prospective clinical impact, calibration across institutions, fairness across patient subgroups, and effects on treatment outcomes were not established.

Future Directions

Prospective silent-mode and interventional studies should evaluate calibration, subgroup fairness, alarm burden, clinician interaction, and whether earlier alerts improve ventilator management, mortality, or other patient-centered outcomes.