Clinical Entity Linking

Leonor Maria Machado Barreiros

Key information

Authors:

Leonor Maria Machado Barreiros (Leonor Maria Machado Barreiros)

Supervisors:

Bruno Emanuel Da Graça Martins (Bruno Emanuel Da Graça Martins); Gonçalo Migueis de Matos Afonso Correia

Published in

November 4, 2024

Abstract

Clinical coding is a crucial task in hospitals and healthcare facilities worldwide. Traditionally, it is treated as a multi-label classification problem, which is not explainable and difficult to verify. In this work, we frame clinical coding as an entity linking task, where each document is annotated with its set of codes and respective mentions. Due to this formulation, coders can critically reason about each code and work in cooperation with the system. Entity linking solutions range from discriminative---where an entity is selected from a closed set of options---to generative methods, where an autoregressive model is asked to generate an entity identifier after the mention, according to a template. Nonetheless, since fully-annotated clinical entity linking data is scarce, this task is largely under-explored. To allow the expedition of clinical coding, we systematize the under-explored clinical entity linking problem. (1) We propose an efficiently-trained generative model that annotates entire documents and generates valid solutions through constrained decoding. (2) We propose a discriminative model trained with distant supervision, which leverages data annotated for automated clinical coding, alleviating the data scarcity problem. Because this data is noisier, we equip it with a noise detection mechanism. (3) We compare both approaches under diverse criteria (like performance and resource consumption) to determine the pros and cons of each solution. Our generative model is 5 times as accurate as the discriminative model without candidate selection and is robust in few-shot settings. Despite this, our discriminative model requires 100 times fewer resources than the generative model.