Development and External Validation of a Detection Model to Retrospectively Identify Patients With Acute Respiratory...

Area of Science:

• Computational Medicine
• Health Informatics
• Critical Care Medicine

Background:

• Acute Respiratory Distress Syndrome (ARDS) diagnosis can be challenging, impacting timely treatment.
• Electronic Health Records (EHR) contain vast data potentially useful for identifying ARDS.
• Developing automated methods for ARDS identification is crucial for clinical research and patient care.

Purpose of the Study:

• To develop and validate a machine-learning (ML) model for retrospective identification of ARDS patients.
• To utilize EHR data, including structured data and radiology reports, for ARDS classification.
• To externally validate the model's performance across different institutions.

Main Methods:

• Retrospective cohort study involving physician-adjudicated ARDS cases.
• Training ML models using vital signs, respiratory support, labs, medications, radiology reports, and clinical notes.
• Internal and external validation using metrics like AUROC, ICI, sensitivity, specificity, and PPV.

Main Results:

• The best-performing model integrated structured EHR data and radiology reports.
• External validation showed an AUROC of 0.88 and an ICI of 0.13.
• The model achieved 80% sensitivity and 64% PPV, identifying ARDS a median of 2.2 hours after meeting criteria.

Conclusions:

• Machine-learning models analyzing EHR data are effective for retrospective ARDS identification.
• The developed model demonstrates generalizability across different healthcare systems.
• This approach facilitates efficient and accurate ARDS case ascertainment in clinical practice.