Figure 3.

Model architecture for the PHI detection task. First, the reports are split by a greedy chunking algorithm: chunks are cut between sentences, with no overlap. Then, chunks are fed to the transformer that leverages its attention mechanism to give a hidden representation to each token. A classification head uses these hidden representations to attribute scores, which measure the likelihood that each token belongs to each PHI class. Based on these scores, each token gets classified into its most likely PHI class. This figure only contains synthetic PHI. PHI: protected health information.