A Knowledge Graph Enhanced Learner Model to Predict Outcomes to Questions in the Medical Field

Abstract

The training curriculum for medical doctors requires the intensive and rapid assimilation of a lot of knowledge. To help medical students optimize their learning path, the SIDES 3.0 national French project aims to extend an existing platform with intelligent learning services. This platform contains a large number of annotated learning resources, from training and evaluation questions to students' learning traces, available as an RDF knowledge graph. In order for the platform to provide personalized learning services, the knowledge and skills progressively acquired by students on each subject should be taken into account when choosing the training and evaluation questions to be presented to them, in the form of customized quizzes. To achieve such recommendation, a first step lies in the ability to predict the outcome of students when answering questions (success or failure). With this objective in mind, in this paper we propose a model of the students' learning on the SIDES platform, able to make such predictions. The model extends a state-ofthe-art approach to fit the specificity of medical data, and to take into account additional knowledge extracted from the OntoSIDES knowledge graph in the form of graph embeddings. Through an evaluation based on learning traces for pediatrics and cardiovascular specialties, we show that considering the vector representations of answers, questions and students nodes substantially improves the prediction results compared to baseline models.

Stunning Doodle: A Tool for Joint Visualization and Analysis of Knowledge Graphs and Graph Embeddings

“…The graph includes in total more than 9.2 billion triples. To be able to exploit such an amount of information in an AI-powered downstream application, a first approach would be to rely on KGEs, as done in [10]. This scenario highlights two common needs for researchers and engineers working on the project:…”

“…In the framework of the SIDES project, GEs computed from OntoSIDES have been used as input features for a ML model designed to predict students' performance on medical questions [10]. In this context, (1) interpreting the information captured by GEs, (2) identifying the best model for their computation, and (3) tuning the hyper-parameters to obtain a meaningful representation of the nodes are crucial tasks.…”

“…In our case, we discovered that GEs computed from directed graphs can capture semantic information from the KG, such as the nodes' type and the university associated with each student; while GEs obtained from undirected graphs tend to summarize the graph structure. As a consequence, if our final goal is the prediction of students' performance [10] the information about the attended university is likely relevant and GEs computed from the directed graph are more meaningful.…”

“…The fame of KGs has kept growing even more as they started to be used as information sources for many AI-powered applications in the most diverse fields, e.g. education [7,10], medicine [19] and finance [13]. One of the reasons for their increasing success is the possibility to easily employ them as input features for several Machine Learning methods by using a low-dimensional representation of such KGs, obtained through the graph embedding process.…”

A systematic approach to identify the information captured by Knowledge Graph Embeddings

“…To tackle this issue, KGE techniques create a fixed-length vector representation of the entities and relations present in the graph. These vectors can then be used for several kinds of tasks within the scope of the KG itself, such as Link prediction [20], Triple Classification [14] and Entity Resolution [21], or in separate downstream applications [9,13].…”