Identification and Prediction of Clinical Phenotypes in Hospitalized Patients With COVID-19: Machine Learning From Medical Records

Abstract

Background: There is significant heterogeneity in disease progression among hospitalized patients with COVID-19. The pathogenesis of SARS-CoV-2 infection is attributed to a complex interplay between virus and host immune response that in some patients unpredictably and rapidly leads to "hyperinflammation" associated with increased risk of mortality. The early identification of patients at risk of progression to hyperinflammation may help inform timely therapeutic decisions and lead to improved outcomes.

Objective: The primary objective of this study was to use machine learning to reproducibly identify specific risk-stratifying clinical phenotypes across hospitalized patients with COVID-19 and compare treatment response characteristics and outcomes. A secondary objective was to derive a predictive phenotype classification model using routinely available early encounter data that may be useful in informing optimal COVID-19 bedside clinical management.

Methods: This was a retrospective analysis of electronic health record data of adult patients (N=4379) who were admitted to a Johns Hopkins Health System hospital for COVID-19 treatment from 2020 to 2021. Phenotypes were identified by clustering 38 routine clinical observations recorded during inpatient care. To examine the reproducibility and validity of the derived phenotypes, patient data were randomly divided into 2 cohorts, and clustering analysis was performed independently for each cohort. A predictive phenotype classifier using the gradient-boosting machine method was derived using routine clinical observations recorded during the first 6 hours following admission.

Results: A total of 2 phenotypes (designated as phenotype 1 and phenotype 2) were identified in patients admitted for COVID-19 in both the training and validation cohorts with similar distributions of features, correlations with biomarkers, treatments, comorbidities, and outcomes. In both the training and validation cohorts, phenotype-2 patients were older; had elevated markers of inflammation; and were at an increased risk of requiring intensive care unit-level care, developing sepsis, and mortality compared with phenotype-1 patients. The gradient-boosting machine phenotype prediction model yielded an area under the curve of 0.89 and a positive predictive value of 0.83.

Conclusions: Using machine learning clustering, we identified and internally validated 2 clinical COVID-19 phenotypes with distinct treatment or response characteristics consistent with similar 2-phenotype models derived from other hospitalized populations with COVID-19, supporting the reliability and generalizability of these findings. COVID-19 phenotypes can be accurately identified using machine learning models based on readily available early encounter clinical data. A phenotype prediction model based on early encounter data may be clinically useful for timely bedside risk stratification and treatment personalization.

Keywords: COVID; big data; biomarkers; clinical phenotypes; critical care; early warning; electronic medical record; immune response; infection; machine learning; mortality; pathogenesis; phenotype; respiratory distress; sepsis; support tool; training; treatment; utility.

Figure 1

Missingness of clinical observations used for clustering. Clinical physiological observations associated with included patients (adults; nontransferees) with missingness of <25% in the population (N=4379) over the entire encounter. ALT: alanine transaminase; AST: aspartate aminotransferase; BUN: blood urea nitrogen; CO2: carbon dioxide; CRP: C-reactive protein; MCH: mean corpuscular hemoglobin; MCV: mean corpuscular volume; MPV: mean platelet volume; NLR: neutrophil-to-lymphocyte ratio; PLR: platelet-to-lymphocyte ratio; RBC: red blood cell count; RDW: red cell distribution width; SBP: systolic blood pressure; SFR: oxygen saturation–to–fraction of inspired oxygen ratio; SpO2: oxygen saturation; WBC: white blood cell count.

Figure 2

Heat map of correlations among clinical data used to generate clustering features showing highly uncorrelated data except for expected positive correlations between red blood cell count (RBC), hemoglobin, and hematocrit and correlations between white blood cell count (WBC) and lymphocytes or neutrophils. ALT: alanine transaminase; AST: aspartate aminotransferase; BUN: blood urea nitrogen; CO2: carbon dioxide; CRP: C-reactive protein; MCH: mean corpuscular hemoglobin; MCV: mean corpuscular volume; MPV: mean platelet volume; NLR: neutrophil-to-lymphocyte ratio; PLR: platelet-to-lymphocyte ratio; RDW: red cell distribution width; SBP: systolic blood pressure; SFR: oxygen saturation–to–fraction of inspired oxygen ratio; SpO2: oxygen saturation.

Figure 3

Weighted consensus clustering for COVID-19 phenotype identification process flow. FiO2: fraction of inspired oxygen.

Figure 4

Predictive gradient-boosting machine phenotype classifier derivation process flow. AUC: area under the curve; FiO2: fraction of inspired oxygen; GBM: gradient-boosting machine; NPV: negative predictive value; PPV: positive predictive value.

Figure 5

Demonstration that k=2 is the optimal number of clusters based on instability analysis for both the training and validation data sets.

Figure 6

Rank plots showing agreement in the most significant phenotype-defining features (eg, age, blood urea nitrogen [BUN], mean corpuscular volume [MCV], creatinine, neutrophil-to-lymphocyte ratio [NLR], red blood cell count [RBC], hemoglobin, and hematocrit) across phenotypes in both the training and validation data sets. ALT: alanine transaminase; AST: aspartate aminotransferase; CO2: carbon dioxide; CRP: C-reactive protein; MCH: mean corpuscular hemoglobin; MPV: mean platelet volume; PLR: platelet-to-lymphocyte ratio; RDW: red cell distribution width; SBP: systolic blood pressure; SFR: oxygen saturation–to–fraction of inspired oxygen ratio; SpO2: oxygen saturation; WBC: white blood cell count.

Figure 7

Violin plots of clustered features showing highly similar distributions and densities of features across phenotypes in both cohorts. ALT: alanine transaminase; AST: aspartate aminotransferase; BUN: blood urea nitrogen; CO2: carbon dioxide; CRP: C-reactive protein; MCH: mean corpuscular hemoglobin; MCV: mean corpuscular volume; MPV: mean platelet volume; NLR: neutrophil-to-lymphocyte ratio; PLR: platelet-to-lymphocyte ratio; RBC: red blood cell count; RDW: red cell distribution width; SBP: systolic blood pressure; SFR: oxygen saturation–to–fraction of inspired oxygen ratio; SpO2: oxygen saturation; WBC: white blood cell count.

Figure 8

Differences in inflammatory biomarkers across phenotypes showing that phenotype 2, associated with hyperinflammatory biomarkers, was not used in clustering (D-dimer, ferritin, fibrinogen, interleukin 6 [IL6], lactate dehydrogenase [LDH], and procalcitonin [PCT]). CRP: C-reactive protein.

Figure 9

Adjusted odds ratios of comorbidities to clinical phenotypes showing similar associations between comorbidities and high severity (phenotype 2) of COVID-19 in both cohorts.

Figure 10

Principal-component analysis (PCA) biplot (training data) showing “good” cluster separation or spatial distribution and similar feature loading (correlations between key phenotype-defining features and principal components) with validation PCA. ALT: alanine transaminase; AST: aspartate aminotransferase; BUN: blood urea nitrogen; CO2: carbon dioxide; CRP: C-reactive protein; MCH: mean corpuscular hemoglobin; MCV: mean corpuscular volume; MPV: mean platelet volume; NLR: neutrophil-to-lymphocyte ratio; PLR: platelet-to-lymphocyte ratio; RBC: red blood cell count; RDW: red cell distribution width; Resp_rate: respiratory rate; SBP: systolic blood pressure; SFR: oxygen saturation–to–fraction of inspired oxygen ratio; SpO2: oxygen saturation; TEMP: temperature; WBC: white blood cell count.

Figure 11

Principal-component analysis (PCA) biplot (validation data) showing “good” cluster separation or spatial distribution and similar feature loading (direction or magnitude) of correlations between key phenotype-defining features and principal components with validation PCA. ALT: alanine transaminase; AST: aspartate aminotransferase; BUN: blood urea nitrogen; CO2: carbon dioxide; CRP: C-reactive protein; MCH: mean corpuscular hemoglobin; MCV: mean corpuscular volume; MPV: mean platelet volume; NLR: neutrophil-to-lymphocyte ratio; PLR: platelet-to-lymphocyte ratio; RBC: red blood cell count; RDW: red cell distribution width; Resp_rate: respiratory rate; SBP: systolic blood pressure; SFR: oxygen saturation–to–fraction of inspired oxygen ratio; SpO2: oxygen saturation; TEMP: temperature; WBC: white blood cell count.

Figure 12

Survival curves for patients in phenotype 2 versus phenotype 1 (days) showing significantly lower survival in phenotype 2 versus phenotype 1 in both the training and validation cohorts.