Machine learning uncovers blood test patterns subphenotypes at hospital admission discerning increased 30-day ICU mortality rates in COVID-19 elderly patients

ABSTRACT

Introduction: Since March 2020, a number of SARS-CoV-2 patients have frequently required intensive care unit (ICU) admission, associated with moderate survival outcomes and an increasing economic burden. Elderly patients are among the most numerous, due to previous comorbidities and complications they develop during hospitalization [1]. For this reason, a reliable early risk stratification tool could help estimate an early prognosis and allow for an appropriate resources allocation in favour of the most vulnerable and critically ill patients. Method(s) This retrospective study includes data from two Spanish hospitals, HU12O (Madrid) and HCUV (Valencia), from 193 patients aged > 64 with COVID-19 between February and November 2020 who were admitted to the ICU. Variables include demographics, full-blood-count (FBC) tests and clinical outcomes. Machine learning applied a non-linear dimensionality reduction by t-distributed stochastic neighbor embedding (t-SNE) [2];then hierarchical clustering on the t-SNE output was performed. The number of clinically relevant subphenotypes was chosen by combining silhouette and elbow coefficients, and validated through exploratory analysis. Result(s) We identified five subphenotypes with heterogeneous interclustering age and FBC patterns (Fig. 1). Cluster 1 was the 'healthiest' phenotype, with 2% 30-day mortality and characterized by moderate leukocytes and eosinophils. Cluster 5, the severe phenotype, showed 44% 30-day mortality and was characterized by the highest leukocyte, neutrophil and platelet count and minimal monocytes and lymphocyte count. Clusters 2-4 displayed intermediate mortality rates (20-28%). Conclusion(s) The findings of this preliminary report of Eld-ICUCOV19 patients suggest the patient's FBC and age can display discriminative patterns associated with disparate 30-day ICU mortality rates.