Hierarchical BERT for Medical Document Understanding

Medical document understanding has gained much attention recently. One representative task is the International Classification of Disease (ICD) diagnosis code assignment. Existing work adopts either RNN or CNN as the backbone network because the vanilla BERT cannot handle well long documents (>2000 to kens). One issue shared across all these approaches is that they are over specific to the ICD code assignment task, losing generality to give the whole document-level and sentence-level embedding. As a result, it is not straight-forward to direct them to other downstream NLU tasks. Motivated by these observations, we propose Medical Document BERT (MDBERT) for long medical document understanding tasks. MDBERT is not only effective in learning representations at different levels of semantics but efficient in encoding long documents by leveraging a bottom-up hierarchical architecture. Compared to vanilla BERT solutions: 1, MDBERT boosts the performance up to relatively 20% on the MIMIC-III dataset, making it comparable to current SOTA solutions; 2, it cuts the computational complexity on self-attention modules to less than 1/100. Other than the ICD code assignment, we conduct a variety of other NLU tasks on a large commercial dataset named as TrialTrove, to showcase MDBERT's strength in delivering different levels of semantics.