Table 3.

Prediction of Severe Disease

Predictora	OR (95% CI)
PMH: DM	1.51 (.51–4.45)
Smoker status	3.13 (1.08–9.38)
ED: MAP low	2.59 (.92–7.47)
ED: SpO2 low	5.36 (2.03–15.07)
ALC low	3.12 (1.00–9.8)
Lactate high	3.90 (.63–22.91)
Glucose high	2.58 (.92–7.30)
RNAemia	6.72 (2.45–19.79)

Abbreviations: ALC, Absolute Lymphocyte Count; CI, confidence interval; DM, Diabetes Mellitus; ED, emergency department; MAP, mean arterial pressure; OR, odds ratio; PMH, Past Medical History; SpO₂, oxygen saturation.

^aPotential predictors of severe disease (World Health Organization score, 5–8) included demographic features (age ≥60 or ≥80 years and sex), past medical history features (lung disease, cancer, diabetes, immunosuppression, heart disease, hypertension, angiotensin-converting enzyme inhibitor or angiotensin receptor blocker use, stroke, dementia, deep venous thrombosis or pulmonary embolus, chronic kidney disease, tobacco smoking, and obesity), binary indicators of abnormal ED vital signs (low SpO₂ and low or high MAP, heart rate, respiratory rate, and temperature), pneumonia on chest radiography or computed tomography, patient-reported symptoms (fever, chills, cough, sore throat, congestion, shortness of breath, chest pain, myalgias, nausea/vomiting/diarrhea, loss of taste, loss of smell, confusion, and headache), and binary indicators of abnormal laboratory values (high or low leukocyte or platelet count; low absolute lymphocyte count; low hemoglobin or fibrinogen levels; high levels of D-dimer, fibrinogen, C-reactive protein, lactate dehydrogenase, ferritin, troponin, lactate, serum urea nitrogen, creatinine, bilirubin, aspartate aminotransferase, alanine aminotransferase, and alkaline phosphatase; high prothrombin time or partial thromboplastin time; and high or low levels of sodium, potassium, chloride, bicarbonate, calcium, magnesium, and glucose).

To prevent overfitting, predictors were selected via elastic net regression of severe disease on these features with 10-fold cross-validation, selecting the regularization parameter λ minimizing mean cross-validated error, and yielding the features in the table above. In a logistic model regressing severe disease on these features, significant predictors of severe disease included tobacco smoking, SpO₂, and RNAemia. RNAemia was associated with 6.7 times the odds of severe disease, adjusting for other features selected by elastic net–penalized regression, an association comparable in magnitude to the association of hypoxia at initial presentation with eventual severe disease. The mean cross-validated area under the receiver operating characteristic curve of the model in predicting severe disease was 0.82. The Akaike Information Criterion (AIC) was 134.74.