Decision Knowledge Graphs: Construction of and Usage in Question Answering for Clinical Practice Guidelines

TL;DR

This paper introduces Decision Knowledge Graphs (DKGs) for representing and querying Clinical Practice Guidelines, significantly improving question-answering accuracy and addressing the complexity and update challenges of CPGs.

Contribution

The paper presents the first construction of DKGs tailored for CPGs, enabling effective question-answering and search functionalities in clinical guidelines.

Findings

40% increase in question-answering accuracy using DKGs

First use of DKGs for representing CPGs

Addresses complexity and update issues in clinical guidelines

Abstract

In the medical domain, several disease treatment procedures have been documented properly as a set of instructions known as Clinical Practice Guidelines (CPGs). CPGs have been developed over the years on the basis of past treatments, and are updated frequently. A doctor treating a particular patient can use these CPGs to know how past patients with similar conditions were treated successfully and can find the recommended treatment procedure. In this paper, we present a Decision Knowledge Graph (DKG) representation to store CPGs and to perform question-answering on CPGs. CPGs are very complex and no existing representation is suitable to perform question-answering and searching tasks on CPGs. As a result, doctors and practitioners have to manually wade through the guidelines, which is inefficient. Representation of CPGs is challenging mainly due to frequent updates on CPGs and decision-based structure. Our proposed DKG has a decision dimension added to a Knowledge Graph (KG) structure, purported to take care of decision based behavior of CPGs. Using this DKG has shown 40\% increase in accuracy compared to fine-tuned BioBert model in performing question-answering on CPGs. To the best of our knowledge, ours is the first attempt at creating DKGs and using them for representing CPGs.