For example, codes are recorded for billing purposes, hence there are biases and missing codes. There are also variations between coders. Mean-pooling may be more robust against these potential noises (due to law of large number), and this may explain the results in next-disease prediction (...