Risk Score Scaling Using the Scorecard Method

Coefficients from the resulting logistic regression models were then combined with the WoE-transformed variables to establish scores for each feature in the Scorecard. This scorecard generation was performed using the Scorecard method implemented in the scorecardpy python package (version 0.1.9.2). As opposed to pure logistic regression models, scorecard models allow a strictly linear combination of scores that can be calculated even on a piece of paper, without the aid of any technology. Calculating the probabilities from a logistic regression model would require inverse transformations of log odds. We chose the scorecard model for the strict linear interpretation and corresponding ease of deployment anywhere.

This method requires users to select target odds and target points (a baseline number of points corresponding to a baseline score) along with the points required to double the odds. As these choices are arbitrary, we used the package defaults, which set the target odds to 1/19, the corresponding target points to 600, and the default points required to double the odds to 50. Supplemental Figure S1 shows an example of a Scorecard distribution calculated in this manner. Since the final Scorecard model is a linear function of the predictors (i.e., higher scores indicate higher COVID-19 risks), using scorecards has many benefits such as transparency, interpretability and facile implementation.