Figure 3
SHAP values for each attendance in the hold-out test set. (a) Plot summarizing the SHAP values for the ten most important variables (by mean absolute SHAP value) for each attendance in the test set. They are ordered by the global impact the feature has on the explanation (equal to the mean absolute SHAP value of the feature across all attendances). For the binary variables (i.e., the condition indicators) this favours variables with a high number of occurrences (i.e., more common conditions), not necessarily those associated with the high reattendance risk. (b) SHAP value against recorded hour of day of attendance registration (dots). (c) SHAP value against number of emergency department visits in the 30 days prior to the given attendance (dots). In panels (b) and (c) vertical dispersion is the result of interaction with other variables. All panels are coloured by the magnitude of the respective variable for the given data point, with lighter colours indicating higher values (e.g., inspect panels (b) and (c)). Grey data points correspond to non-binary, nominal categorical variables and therefore have no natural ordering which could be used to colour the data points.
