Extubation Decision Support in Critical Care: A Multimodal Machine Learning Framework Integrating Segmented Radiographs and Routine Clinical Data.
Authors
Affiliations (26)
Affiliations (26)
- Respiratory Therapy Room, Division of Pulmonary Medicine, Taipei Medical University-Shuang Ho Hospital, New Taipei City, 23561, Taiwan.
- Sirindhorn International Institute of Technology, Thammasat University, Pathum Thani, 12120, Thailand.
- School of Biomedical Engineering, College of Biomedical Engineering, Taipei Medical University, 250 Wuxing Street, Taipei, 11031, Taiwan.
- Research Center of Sleep Medicine, College of Medicine, Taipei Medical University, Taipei, 11031, Taiwan.
- Division of Pulmonary Medicine, Department of Internal Medicine, Taipei Medical University-Shuang Ho Hospital, New Taipei City, 23561, Taiwan.
- School of Respiratory Therapy, College of Medicine, Taipei Medical University, Taipei, 11031, Taiwan.
- Advanced Technology Lab, Wistron Corporation, Taipei, 11469, Taiwan.
- Sleep Center, Taipei Medical University-Shuang Ho Hospital, New Taipei City, 23561, Taiwan.
- TMU Research Center of Artificial Intelligence in Medicine and Health, Taipei Medical University, Taipei, 11031, Taiwan.
- Center for Artificial Intelligence and Advanced Robotics, National Taiwan University, Taipei, 106319, Taiwan.
- College of Biomedical Engineering, Taipei Medical University, Taipei, 11031, Taiwan.
- Department of Radiology, Taipei Medical University-Shuang Ho Hospital, New Taipei City, 23561, Taiwan.
- Emergency Department, Taipei Medical University-Shuang Ho Hospital, New Taipei City, 23561, Taiwan.
- Division of Pulmonary Medicine, Department of Internal Medicine, School of Medicine, College of Medicine, Taipei Medical University, Taipei, 11031, Taiwan.
- TMU Research Center for Thoracic Medicine, Taipei Medical University, Taipei, 11031, Taiwan.
- Graduate Institute of Clinical Medicine, College of Medicine, Taipei Medical University, Taipei, 11031, Taiwan.
- Department of Civil and Environmental Engineering, Imperial College London, London, SW7 2AZ, UK.
- Department of Physical Medicine and Rehabilitation, Taipei Medical University Hospital, Taipei, 11031, Taiwan.
- Graduate Institute of Nanomedicine and Medical Engineering, College of Biomedical Engineering, Taipei Medical University, Taipei, 11031, Taiwan.
- School of Biomedical Engineering, College of Biomedical Engineering, Taipei Medical University, 250 Wuxing Street, Taipei, 11031, Taiwan. [email protected].
- Research Center of Sleep Medicine, College of Medicine, Taipei Medical University, Taipei, 11031, Taiwan. [email protected].
- Division of Pulmonary Medicine, Department of Internal Medicine, Taipei Medical University-Shuang Ho Hospital, New Taipei City, 23561, Taiwan. [email protected].
- School of Respiratory Therapy, College of Medicine, Taipei Medical University, Taipei, 11031, Taiwan. [email protected].
- Sleep Center, Taipei Medical University-Shuang Ho Hospital, New Taipei City, 23561, Taiwan. [email protected].
- TMU Research Center of Artificial Intelligence in Medicine and Health, Taipei Medical University, Taipei, 11031, Taiwan. [email protected].
- TMU Research Center for Thoracic Medicine, Taipei Medical University, Taipei, 11031, Taiwan. [email protected].
Abstract
Extubation failure remains a major challenge in critically ill patients and is associated with adverse clinical outcomes. Current extubation decisions rely heavily on weaning tests and subjective interpretation of chest radiographs. This study aimed to develop a clinically feasible multimodal machine learning (ML) framework that integrates routinely available data to provide complementary support for extubation decision-making. A total of 921 individuals were included and classified into extubation-with-reintubation and extubation-without-reintubation groups. The proposed framework integrated baseline demographics, weaning measurements, radiographic assessments, and segmented post-intubation chest x-rays (i.e., tracheal, left lung, and right lung regions). Optimal base ML models for each modality were selected based on the area under the receiver operating characteristic curve and integrated using a stacking ensemble approach. Feature importance analyses were performed at both the modality and feature levels. The extubation-with-reintubation group comprised a higher proportion of elderly patients with higher Charlson comorbidity index scores than the extubation-without-reintubation group. Individuals requiring reintubation exhibited significantly higher respiratory rates, lower tidal volumes, greater rapid shallow breathing indices, and longer intervals from intubation to weaning tests and extubation (all pā<ā0.01). The multimodal ensemble outperformed rule-based and single-modality models, achieving an accuracy of 79.46%. Weaning measurements, demographics, and radiographic assessments were the most influential contributors to extubation outcome prediction. A multimodal ML framework integrating segmented post-intubation chest x-rays with routinely collected clinical data shows potential as a complementary, objective decision-support tool for extubation without requiring additional measurements. Prospective studies are needed to further validate these findings.