Application of Machine Learning for Clinical Subphenotype Identification in Sepsis

Sepsis is a heterogeneous clinical syndrome. Identification of sepsis subphenotypes could lead to allowing more precise therapy. However, there is a lack of models to identify the subphenotypes in such patients. Thus, we aimed to identify possible subphenotypes and compare the clinical outcomes for subphenotypes in a large sepsis cohort.

This machine learning-based, cluster analysis was performed using the Medical Information Mart in Intensive Care (MIMIC)-IV database. We enrolled all adult (> 18 years old) patients diagnosed with sepsis in the first 24 h after intensive care unit (ICU) admission. K-means cluster analysis was performed to identify the number of classes. Multivariable logistic regression models were used to estimate the association between sepsis subphenotypes and in-hospital mortality.

A total of 8817 participants with sepsis were enrolled. The median age was 66.8 (IQR, 55.9-77.1) years, and 38.1% (3361/8817) were female. Two subphenotypes resulted in optimal separation including 11 routinely available clinical variables obtained during the first 24 h after ICU admission. Participants in subphenotype B showed higher levels of lactate, glucose and creatinine, white blood cell count, sodium and heart rate and lower body temperature, platelet count, systolic blood pressure, hemoglobin and PaO₂/FiO₂ ratio. In addition, the in-hospital mortality in patients with subphenotype B was significantly higher than that in subphenotype A (29.4% vs. 8.5%, P < 0.001). The difference was still significant after adjustment for potential covariates (adjusted OR 2.214; 95% CI 1.780-2.754, P < 0.001).

Two sepsis subphenotypes with different clinical outcomes could be rapidly identified using the K-means clustering analysis based on routinely available clinical data. This finding may help clinicians to identify the subphenotype rapidly at the bedside.