Figure 3

Marginal value of RT-PCR testing in patients with clinically suspected Covid-19 infection. Estimation and evaluation using testing propensity-score weights. (a) Decision paths of the tree, applied to evaluation patients. A decision tree classifier is trained on 80% of the tested patients. It predict RT-PCR positivity using 12 features: breathlessness, anorexia, tiredness, digestive signs (diarrhea/vomiting), conjunctivitis, cutaneous symptoms (rash/frostbites), shivers, myalgia, cough, fever, cardiopulmonary symptoms (breathlessness + chest pain/oppression) and chemosensory impairment (anosmia/ageusia). We evaluate the decision tree on 20% held-out patients and illustrate how it splits this population. Each node is a splitting criterion (presence of the symptoms to the top, absence to the bottom). The colour of each node corresponds to the odds ratio of being RT-PCR+ at this stage of the decision path. For each leaf, the probability and odds ratio of being RT-PCR+ are reported (see Figs. S6, S8 and Table S1 for details). (b) Permutation features importance on the evaluation set. The permutation importance is an indicator of the relevance of a feature at predicting RT-PCR positivity. It measures the decrease in the model score (here, average precision) when a single feature is randomly shuffled. We report the permutation importance on the left-out evaluation data (20% of the dataset) for each feature of the decision tree. Error bars are the standard deviations of the importance through 50 different permutations. (c) Performance of the decision tree on the test set. Precision-recall curve of the decision tree on the 20% held-out set of tested patients.