Complete Blood Count Parameters Can Outperform Regular Inflammatory Markers in Predicting COVID-19 Mortality: A Multimodal Machine Learning Model

Objectives: We identified cell counts and proportions reported in, or calculated from, a complete blood count (CBC) that independently predicted mortality in hospitalized COVID-19 patients. The primary objective was to characterize a CBC signature at presentation that might provide insights for tracking immune response and disease management.Methods/Study design: This was a retrospective longitudinal observational study. The primary outcome was in-hospital mortality. Secondary outcomes included the need for mechanical ventilation, development of sepsis, and ICU admission. Electronic medical records underwent IRB exempted extraction of clinical data. Univariate logistic regression was used to identify CBC parameters and putative inflammatory markers independently predictive of hospital mortality. Bootstrap Forest (BF) modeling was employed to aggregate predictive CBC parameters that optimized generalized coefficient of variation (R2) concurrently computing their proportion of explained variance in mortality. BF modeling was replicated with inflammatory markers and subsequently with pooled features from both models.Results: CBC parameters including segmented neutrophils, bands, ANC, and RDW-CV were significantly elevated in non-survivors compared to survivors. In addition, patients with decreased platelets, lymphocytes and monocytes were more likely to be in the non-survivor group. Incorporating these CBC parameters amplified R2 for mortality in a presentation prognostic model with only inflammatory markers including C-reactive protein, lactate dehydrogenase and ferritin.Conclusion: Routinely obtained CBC parameters improve predictive power of putative COVID-19 severity markers and enhance accuracy of mortality risk assessment at presentation. We recommend that these CBC parameters be considered for presentation risk stratification, level of care triaging decisions and longitudinal tracking of disease management.