Machine Learning Classifier Models Can Identify Acute Respiratory Distress Syndrome Phenotypes Using Readily...

Area of Science:

• Critical Care Medicine
• Pulmonary Medicine
• Biostatistics

Background:

• Two distinct acute respiratory distress syndrome (ARDS) phenotypes with differing outcomes and treatment responses have been identified.
• Current phenotype identification relies on plasma biomarkers, which lack point-of-care assays, hindering clinical application.
• There is a need for methods to classify ARDS phenotypes using readily available clinical data.

Purpose of the Study:

• To develop and validate machine learning models for classifying ARDS phenotypes using only clinical data.
• To assess the accuracy of these models compared to latent class analysis-derived phenotypes.
• To determine if clinical data-based phenotype classification can identify patients with differential inflammatory markers and mortality.

Main Methods:

• A gradient-boosted machine algorithm was employed to build classifier models.
• Models were trained on data from three randomized controlled trial cohorts (n=2,022) and validated on a fourth cohort (n=745).
• Twenty-four enrollment variables including demographics, vital signs, and laboratory/respiratory data were utilized.

Main Results:

• The primary validation model achieved high accuracy in classifying ARDS phenotypes (AUC, 0.95).
• The model-identified hyperinflammatory phenotype showed significantly higher inflammatory biomarkers and 90-day mortality.
• Similar high accuracy was observed in secondary analyses using different validation cohorts, with significant treatment interactions noted.

Conclusions:

• Machine learning models can accurately identify ARDS phenotypes using readily available clinical data.
• This approach facilitates rapid, bedside phenotype classification.
• Clinical data-based phenotype identification holds promise for personalized treatment strategies in ARDS.