Application of machine learning algorithms to identify serological predictors of COVID-19 severity and outcomes

Abstract

Background: Critically ill hospitalized patients with COVID-19 have greater antibody titers than those with mild to moderate illness, but their association with recovery or death from COVID-19 has not been characterized.

Methods: In a cohort study of 178 COVID-19 patients, 73 non-hospitalized and 105 hospitalized patients, mucosal swabs and plasma samples were collected at hospital enrollment and up to 3 months post-enrollment (MPE) to measure virus RNA, cytokines/chemokines, binding antibodies, ACE2 binding inhibition, and Fc effector antibody responses against SARS-CoV-2. The association of demographic variables and more than 20 serological antibody measures with intubation or death due to COVID-19 was determined using machine learning algorithms.

Results: Predictive models reveal that IgG binding and ACE2 binding inhibition responses at 1 MPE are positively and anti-Spike antibody-mediated complement activation at enrollment is negatively associated with an increased probability of intubation or death from COVID-19 within 3 MPE.

Conclusions: At enrollment, serological antibody measures are more predictive than demographic variables of subsequent intubation or death among hospitalized COVID-19 patients.

Fig. 1. SARS-CoV-2 virus RNA and cytokine/chemokine responses among hospitalized COVID-19 patients at enrollment.

a Nasopharyngeal (NP) viral load (copies/mL, log₁₀) and (b) oropharyngeal (OP) viral load (copies/mL, log₁₀) were measured by qPCR at enrollment and compared among patients classified as moderate (WHO score 3–4), severe (WHO score 5–7), or deceased (WHO score 8). c The Spearman correlation between OP and NP viral loads at enrollment. d–i Concentrations (pg/ml) of several proinflammatory cytokines and chemokines that differed among COVID-19 hospitalized patients classified as moderate, severe, or deceased. Data are presented as means with standard deviations, indicated by error bars. p-values for statistically significant differences (p < 0.05) by Welch’s ANOVA are shown in the figures.

Fig. 2. Mucosal antibody responses among hospitalized COVID-19 patients at enrollment.

a–d Anti-nucleocapsid (N) and anti-spike (S) secretory IgA or (e–h) IgG responses were measured as median fluorescence intensity (MFI) in nasopharyngeal (NP) or oropharyngeal (OP) samples and compared among COVID-19 hospitalized patients classified as moderate (WHO score 3–4), severe (WHO score 5–7), or deceased (WHO score 8). Data are analyzed using Welch’s ANOVA and presented as means with standard deviations, indicated by error bars.

Fig. 3. Antibody responses in plasma samples of non-hospitalized and hospitalized COVID-19 patients at 1-month post-enrollment (MPE).

a–d IgG binding antibody responses against ancestral spike (S), spike receptor binding domain (S-RBD), and nucleocapsid (N) were quantified by ELISA and calculated as the binding antibody units (BAU) per ml if international standards were available or as the area under the curve (AUC) if standards were not available and titration curves only could be generated; (e) ACE2 binding inhibition antibodies were measured using MSD V-PLEX SARS-CoV-2 ACE2 kits; and (f–h) Fc effector antibody responses were quantified using complement fixation and antibody-dependent cellular cytotoxicity (ADCC) assays. All assays were run using ancestral SARS-CoV-2. Data were compared using linear regression analysis, controlling for age and biological sex, to look at differences between unvaccinated non-hospitalized and hospitalized patients at 1 MPE. Data are presented as means with standard deviations, indicated by error bars. The limit of detection (LOD) is indicated by the dashed lines. p-values for statistically significant differences (p < 0.05) are shown in the figures.

Fig. 4. Binding, ACE2 inhibition, and Fc effector antibody responses in plasma among COVID-19 hospitalized patients at enrollment and 1-month post-enrollment (MPE).

The binding (a–c) IgG and (d) IgA antibodies recognizing ancestral SARS-CoV-2 spike (S), spike receptor binding domain (S-RBD), or nucleocapsid (N) were quantified by ELISA, and measured as the binding antibody units (BAU) per mL if international standards were available or as the area under the curve (AUC) if standards were not available and titration curves could only be generated. e The percentage of ACE2 inhibition for the ancestral SARS-CoV-2 variant was calculated and arcsine transformed for analyses. f–h The Fc effector antibody responses were measured based on C1q complement fixation in response to either the spike or S-RBD or antibody dependent cellular cytotoxicity and reported as arbitrary units (AU). Antibody responses were compared among COVID-19 hospitalized patients classified as moderate (WHO score 3–4), severe (WHO score 5–7), or deceased (WHO score 8) using samples collected at enrollment vs. 1 MPE. Data are presented as means with standard deviations, indicated by error bars. p-values for statistically significant differences (p < 0.05) by linear mixed-effects regression to compare change over time or Welch’s ANOVA to compare across groups within a time point are indicated. Limit of detection (LOD) are indicated by the dashed lines.

Fig. 5. Analysis of anti-Spike (S) IgG subclasses (IgG1-4) among hospitalized COVID-19 patients at enrollment and 1-month post-enrollment (MPE).

The binding of IgG1 (a), IgG2 (b), IgG3 (c), and IgG4 (d) to ancestral SARS-CoV-2 S antigen was measured as the area under the curve (AUC). Spearman correlation of IgG1 (e), IgG2 (f), IgG3 (g), and IgG4 (h) with % ACE2 inhibition at enrollment. Hospitalized COVID-19 patients were classified as moderate (WHO score 3–4), severe (WHO score 5–7), or deceased (WHO score 8). Data are presented as means with standard deviations, indicated by error bars. The limit of detection (LOD) is indicated by the dashed lines. p-values for statistically significant differences (p < 0.05) by linear mixed-effects regression to compare change over time or Welch’s ANOVA to compare across groups within a time point are indicated.

Fig. 6. Antibody responses against ancestral SARS-CoV-2 over continuous days since enrollment until 100 days post-enrollment (DPE) or subsequent death among hospitalized COVID-19 patients.

Linear mixed-effects regression models for (a–d) anti-spike (S), anti-spike receptor binding domain (S-RBD), or anti-nucleocapsid (N) IgG or IgA, measured as the binding antibody units (BAU) per ml if international standards were available or as the area under the curve (AUC) if standards were not available and only titration curves could be generated; (e) the percentage ACE2 inhibition against ancestral SARS-CoV-2 as a surrogate of virus neutralization, and (f–h) Fc effector antibody responses as measured by complement fixation against spike or S-RBD or antibody-dependent cellular cytotoxicity (ADCC) up until 100 DPE or death among hospitalized patients classified as moderate (WHO score 3–4; n = 41), severe (WHO score 5–7; n = 40), or deceased (WHO score 8; n = 24). P-values for statistically significant differences (p < 0.05) by linear mixed-effects regression contrasts are shown within the figures.

Fig. 7. Logistic regression of COVID-19 death by indexed antibody variables among hospitalized COVID-19 patients at enrollment and 1-month post-enrollment (MPE).

Logistic regression modeling for death among hospitalized COVID-19 patients by indexed scores based on quartiles of (a, e) binding, (b, f) ACE2 inhibition, (c, g) ADCC, or (d, h) complement fixation at enrollment or 1 MPE, respectively. Predicted probabilities from logistic regression models are graphed in black with 95% confidence intervals shaded in gray. i, j Random Forest variable importance plots were used to determine the relative ranking of different demographic and serological variables in descending order of importance, expressed as mean decrease accuracy, for the classification of intubation or death among hospitalized patients at enrollment. Exclusion of serological variables from models, particularly those >10% mean decrease accuracy, would result in reduced model accuracy for classifying patients as intubated or deceased.