J. Parreco1, S. Byerly1, H. Soe-Lin1, J. Parks1, N. Namias1, R. Rattan1 1University Of Miami,Surgical Critical Care/Trauma,Miami, FL, USA
Introduction:
The reported efficacy of vitamin C and thiamine administration in sepsis has varied widely. Previous reports of classifications of patients by their phenotypic response to sepsis using unsupervised machine learning has demonstrated that treatment efficacy can be predicted. The purpose of this study was to classify patients by sepsis phenotypes and to examine the correlation with response to vitamin C and thiamine administration.
Methods:
The eICU Collaborative Research Database (eICU-CRD) 2.0 was queried for all patients aged 18 years or older with a diagnosis of sepsis within the first day of ICU admission. Four different phenotypes (α, β, γ, and δ) were identified using k-means clustering with thirty different clinical variables. The outcome of interest was mortality in the ICU. Each phenotype was evaluated for the presence of an order for vitamin C or thiamine and ICU mortality rates were compared using a chi-squared test.
Results:
There were 31,508 patients identified with sepsis during the first day of ICU admission. The overall mortality rate was 10.2% (n=3,591). The most common phenotype was α with 62.6% (n=21,956) of patients and the least common was the δ phenotype with 273 (0.8%). The figure shows a chord diagram of abnormal clinical variables grouped by organ system. The broader the ribbon, the more abnormal variables. The α phenotype had the lowest mortality rate of 3.8% (n=844) while the δ phenotype had the highest mortality rate of 46.9% (n=128, p<0.001). The overall mortality rate for patients with vitamin C ordered was 10.0% (n=85, p=0.822). There was no statistical significance in the mortality rates with vitamin C per phenotype. The overall mortality rate for patients with thiamine ordered was 10.8% (n=134, p=0.535). For patients with the β phenotype, the mortality rate with thiamine administration was 22.5% (n=38, p=0.005). The mortality rates with thiamine administration for the other phenotypes was not statistically significant.
Conclusions:
Classifying patients with sepsis by their host response patterns identified by unsupervised machine learning can reveal important aspects to the heterogeneity of treatment effects. The combination of host response patterns revealed in this study demonstrate that it is possible to identify patients who will respond poorly to thiamine administration for sepsis.
