Early Immunological Biomarkers for Personalized Treatment Selection in Severe COVID-19: Post Hoc Machine Learning Analysis of a Randomized Clinical Trial.

Background: Severe COVID-19 is a global health concern despite continuous vaccination campaigns because current therapies, such as dexamethasone and remdesivir, do not considerably improve immune function, especially in high-risk individuals. SARS-CoV-2-specific T cells (CoV-2-STs) from vaccinated or convalescent donors are a promising new treatment that can enhance clinical outcomes and viral-specific immunity. CoV-2-STs improve T cell proliferation and recovery without raising safety concerns, according to randomized studies. Targeting patients for immunotherapy is made more difficult by the variability in COVID-19 progression brought on by variables like age and comorbidities. In order to further enable precision medicine and patient care, machine learning techniques are being used to analyze clinical data, predict disease severity, and optimize treatment. However, their use in guiding the treatment of novel therapies like CoV-2-STs using early cellular immunology data is limited and requires improvement. Objective: The purpose of this research was to stratify high-risk individuals using early immunological and clinical indicators, and to develop a prediction tool to enable individualized treatment decisions, either with standard of care (SoC) or with a combination of CoV-2-STs and SoC. Methods: A randomized phase 1-2 trial enrolled 87 patients with severe COVID-19 (CoV-2-STs+SoC: n=57; SoC only: n=30). We performed a post hoc machine learning analysis. Clinical and biomarker data from days 0 and 5 were analyzed longitudinally. Shrinkage linear discriminant analysis was used to create arm-specific prognostic models, and stratified cross-validation assessed performance (area under the receiver operating characteristic curve, area under the precision-recall curve, sensitivity, specificity, F1-score, and Brier score). To investigate arm-swap scenarios, Monte Carlo simulations (4 variance models) generated hypothetical early-treatment trajectories. Results: Day 5 analysis revealed a significant difference in CD3+, CD8+, CD56+, and CoV-2-STs between the 2 treatment groups (CoV-2-STs+SoC vs SoC). At Day 60, 64.9% (37/57) of patients with CoV-2-STs+SoC survived, compared with 40% (12/30) of patients with only SoC, resulting in a crude odds ratio of 2.8 (95% CI 1.1-6.9) for recovery. SoC-only models had an area under the curve between 0.72 and 0.76, while CoV-2-STs+SoC models had an area under the curve between 0.86 and 0.88. Area under the precision-recall curve values for CoV-2-STs+SOC models were 0.74-0.78, with sensitivity ranging from 0.89 to 0.91 and specificity from 0.83 to 0.87, compared with SoC-only models with a sensitivity of approximately 0.95 and a specificity of 0.58-0.62. Simulation studies indicate that CoV-2-STs+SoC may benefit up to 30% (approximately 9/30) of patients receiving SoC alone. Misclassifying candidates with CoV-2-STs+SoC as SoC-only could increase critical outcomes by up to approximately 22%. Conclusions: A robust computational tool for severe COVID-19 risk stratification and treatment selection is presented. Precision medicine and early treatment outcome prediction are supported by clinical and immunological data integration. Prospective studies are needed to confirm its clinical utility.