Validating Auto-Suggested Changes for SNOMED CT in Non-Lattice Subgraphs Using Relational Machine Learning.

Abstract

An attractive feature of non-lattice-based ontology auditing methods is its ability to not only identify potential quality issues, but also automatically generate the corresponding fixes. However, exhaustive manual evaluation of the validity of suggested changes remains a challenge shared with virtually all auditing methods. To address this challenge, we explore machine learning techniques as an aid to systematically evaluate the strength of auto-suggested relational changes in the context of existing knowledge embedded in an ontology. We introduce a hybrid convolutional neural network and multilayer perception (CNN-MLP) classifier using a combination of graph, concept features and concept embeddings. We use lattice subgraphs to generate a curated, loosely-coupled training set of positive and negative instances to train the classifier. Our result shows that machine learning techniques have the potential to alleviate the manual effort required for validating and confirming changes generated by non-lattice-based auditing methods for SNOMED CT.