Source Data SetsThe first dataset was derived from a prospective cohort study which measured anti-SARS-COV-2 nAb half-maximal inhibitory concentration (IC50) titers against pseudotyped WA-1 (the original SARS-CoV-2 strain) at specific time periods over a year-long interval following the primary messenger RNA vaccination series and a subsequent booster vaccine dose (before dose 1, before dose 2, and after dose 2 at 1, 3, 6, and 12 months, with an additional collection 1 month after dose 3, which was approximately 10 months after dose 2). Clinical variables were collected by manual chart review as previously described. The nAb titers were found to fluctuate over time, peaking 1 month after the primary series, waning at 6 months, and rising again 1 month after the booster dose [2].
The second dataset leveraged the collected serum samples and measured IgG anti-S1/receptor-binding domain values in sera with predetermined nAb titers and clinical variables as previously described [6]. IgG and nAb IC50 titers had a robust linear correlation. Multimedia Appendix 1 illustrates this correlation for each clinical variable (COVID-19 infection status, estimated glomerular filtration rate (GFR), age, sex, history of malignancy, and race). Based on a forward stepwise linear regression, GFR (0‐30, 31‐59, and >60 mL/min/1.73 m²), and history of COVID-19 infection within the past 12 months (yes or no) were found to impact this correlation (R2=0.78) [6].
The model had undergone internal validation using repeated k-fold cross-validation. The model’s performance was evaluated using root mean squared error (RMSE), R2, and mean absolute error (MAE) across multiple versions of the model and different crossvalidation setups (ten 10-fold repeats, five 10-fold repeats, and three 5-fold repeats). For our model, which included IgG, COVID-19 infection status, and GFR, the RMSE was relatively low (around 0.27‐0.28) across validation setups, suggesting a small average error between the predicted and actual values. The R2 values were around 0.77 to 0.81 across validation setups, which indicated that the model explained about 77% to 81% of the variability in the data. The MAE values were around 0.2 for all validation setups, demonstrating a low average error magnitude (Multimedia Appendix 2).
Application DevelopmentBackend (Software)The application used the model-view-controller pattern. We initially used RStudio 2023.12.1+402 running R 4.3.2 (IBM Corp) and an integrated development environment with a set of tools to complement R. The original R script had initially used its own web application framework (Shiny) to create a local web-based application, but this framework was not compatible with the rest of the tools used in the development process. The R script was converted to Python 3.9.12 (Python Software Foundation) for ease of web development. The main challenge when developing the project was converting the original R script into Python and creating the website using an HTML and cascading style sheets (CSS). Python was chosen over R for several reasons. First, Python’s versatility and ecosystem made integration with other tools like Flask, a popular web framework that simplifies the web application development process, and Docker, which simplifies the process of containerizing applications, much easier. Second, Python’s versatility and rich ecosystem of libraries and web development tools allowed us to keep the backend and front end the same language. Third, Python could handle a high volume of web requests and was better suited for scalable applications, which was a point of consideration. Last, Python was also the better choice for long-term maintainability, as it is more readable and easier to update. The host server emphasized Docker as the primary tool for deployment, making it the easiest tool to use for the development process. The full source code is available on request.
Front End (Web Interface)The front end was converted from a Shiny package in RStudio to an HTML and CSS. The Python Flask 3.0.0 (Armin Ronacher) module was used to build the CoVimmune website. The website was launched using a YaleInformation Technology Support secured self-service sandbox environment called Spinup. The Waitress Web Server Gateway Interface was used to deploy our application and Docker 4.27.1 (Docker, Inc) was used to package, ship, and run our application.