Abstract Rationale Acute Respiratory Distress Syndrome (ARDS) exhibits marked biological and clinical heterogeneity that complicates diagnosis and treatment. Prior approaches have classified patients into binary phenotypes based on biomarkers or ventilator parameters, providing limited insight into disease dynamics or treatment responsiveness. We developed a continuous prognostic phenotyping model that improves on binary subtyping to capture the nuanced progression of ARDS using routinely available clinical data. This enables near real-time measurement of disease severity and treatment effects at the bedside. Methods We trained a machine learning model using clinical and physiological data from the MIMIC-IV and eICU databases. Latent Profile Analysis and Support Vector Machine were employed to identify hidden prognostic subgroups of ARDS patients and transform these latent clusters into a continuous phenotype score between 0 (more severe) and 1 (less severe). Unlike previous models that depend on complex biomarkers or ventilator-specific inputs, our model uses only eight standard clinical variables—heart rate, oxygen saturation (SpO2), creatinine, glucose, hematocrit, platelet count, sodium, and white blood cell count—making it deployable across most intensive care units (ICUs). The model was validated on five randomized controlled trials and used to evaluate treatment effects within the eICU dataset through statistical analyses comparing treatment and control groups. Results The model successfully quantified prognostic heterogeneity across ARDS patients and revealed differential treatment responses. Mechanical ventilation demonstrated a significant negative slope (p 0.05) among patients with phenotype scores 0.74, indicating worse outcomes in patients with lower scores, whereas those with higher scores benefited. Methylprednisolone improved outcomes in moderate disease states (phenotype 0.25-0.50; p = 0.001), while dexamethasone showed adverse effects in milder cases (phenotype 0.75). These results demonstrate that the continuous phenotype score sensitively tracks disease transitions and treatment efficacy with temporal resolution unavailable in binary models. We also found our models to be over 2 times as effective at predicting outcomes in ARDS in contrast to generic scores such as SOFA and APACHE. Conclusions We present a novel, validated continuous phenotyping model that quantifies prognostic heterogeneity and treatment effects in ARDS using only routine clinical data. This model bridges the gap between biological complexity and clinical practicality by dynamically measuring disease trajectory. Its interpretability and real-time applicability offer a scalable pathway toward precision medicine in critical care and can inform therapeutic stratification in ARDS trials. This model will be further validated prospectively in GEn1E’s Phase 2 ARDS study (NCT05795465) partially funded by BARDA. This abstract is funded by: GEn1E Lifesciences Inc.
Lal et al. (Fri,) studied this question.