Integrated gradients fades the patient in from a blank baseline and adds up how much each feature moved the prediction along the way.
A toy logistic model. "Shoe size" has no effect on the model, so it earns nothing. At the end of the path the attributions always add up to the change in prediction.