Fig. 5: Model selection procedure.

Rounds of model selection in the multivariate probit component of the Syndromic-only and Syndromic-Rapid Antigen Test (RAT) Combined models. With 14 symptoms (5 shown for demonstration purposes) and 2 covariates there are over 131,000 possible model combinations. To make exploring these models computationally feasible and to reduce the risk of overfitting, we carried out two rounds of model selection. A subset of symptoms are identified using the strength of posterior correlation between each symptom and PCR status identified by the corresponding model, with the weakest correlated symptoms removed during each round of selection. From this subset of symptoms, a more exhaustive search of potential models is then conducted to identify the best symptom-covariate relationships, using temporal cross-validation to measure model performance. The best model for each level of complexity (i.e., number of symptoms) are then used as our candidate models. Only these final models are used for classification. This reduces the set of models tested as classifiers from >131,000 to just four per model class.