Machine learning-based identification of efficient and restrictive physiological subphenotypes in acute respiratory distress syndrome.

INTRODUCTION

Acute respiratory distress syndrome (ARDS) is a severe condition with high morbidity and mortality, characterized by significant clinical heterogeneity. This heterogeneity complicates treatment selection and patient inclusion in clinical trials. Therefore, the objective of this study is to identify physiological subphenotypes of ARDS using machine learning, and to determine ventilatory variables that can effectively discriminate between these subphenotypes in a bedside setting with high performance, highlighting potential utility for future clinical stratification approaches.

METHODOLOGY

A retrospective cohort study was conducted using data from our ICU, covering admissions from 2017 to 2021. The study included 224 patients over 18 years of age diagnosed with ARDS according to the Berlin criteria and undergoing invasive mechanical ventilation (IMV). Data on physiological and ventilatory variables were collected during the first 24 h IMV. We applied machine learning techniques to categorize subphenotypes in ARDS patients. Initially, we employed the unsupervised Gaussian Mixture Classification Model approach to group patients into subphenotypes. Subsequently, we applied supervised models such as XGBoost to perform root cause analysis, evaluate the classification of patients into these subgroups, and measure their performance.

RESULTS

Our models identified two ARDS subphenotypes with significant clinical differences and significant outcomes. Subphenotype Efficient (n = 172) was characterized by lower mortality, lower clinical severity and presented a less restrictive pattern with better gas exchange compared to Subphenotype Restrictive (n = 52), which showed the opposite. The models demonstrated high performance with an area under the ROC curve of 0.94, sensitivity of 94.2% and specificity of 87.5%, in addition to an F1 score of 0.85. The most influential variables in the discrimination of subphenotypes were distension pressure, respiratory frequency and exhaled carbon dioxide volume.

CONCLUSION

This study presents an approach to improve subphenotype categorization in ARDS. The generation of clustering and prediction models by machine learning involving clinical, ventilatory mechanics, and gas exchange variables allowed for more accurate stratification of patients. These findings have the potential to optimize individualized treatment selection and improve clinical outcomes in patients with ARDS.