Fig. 5: SHapley Additive exPlanations (SHAP) force or explanation plots of COVID-19-positive and -negative patient encounters.

Sample observations from two patient encounters with a COVID-19-negative and b COVID-19-positive predictions. Features that are contributing to a higher and lower SHAP values are shown in red and blue, respectively, along with the size of each feature’s contribution to the model’s output. The baseline—the mean of the model output (log-odds) over the training dataset—is 0.057 (translating to a probability value of 0.5143). The first patient—who is COVID-19 negative—has a low predicted risk score of −3.27 (output probability = 0.0366). The second patient—who is COVID-19 positive—has a high predicted risk of 1.94 (output probability = 0.8744). These SHAP output values represent a “raw” log-odds value which is transformed into a probability space, to provide the final output of 0 and 1 (<0.5 and >0.5). Risk increasing effects such as Temp, SpO₂, HR, and BP were offset by decreasing effects of RR in pushing the model’s predictions towards or away from the positive class, respectively.