No abstract

Abstract. Biomedical ontologies have become a mainstream topic in medical research. They represent important sources of evolved knowledge that may be automatically integrated in decision support methods. Grounding clinical and radiographic findings in concepts defined by a biomedical ontology, e.g., the Human Phenotype Ontology, enables us to compute semantic similarity between them. In this paper, we focus on using such similarity measures to predict disorders on undiagnosed patient cases in the bone dysplasia domain. Different methods for computing the semantic similarity have been implemented. All methods have been evaluated based on their support in achieving a higher prediction accuracy. The outcome of this research enables us to understand the feasibility of developing decision support methods based on ontology-driven semantic similarity in the skeletal dysplasia domain.

BackgroundLately, ontologies have become a fundamental building block in the process of formalising and storing complex biomedical information. With the currently existing wealth of formalised knowledge, the ability to discover implicit relationships between different ontological concepts becomes particularly important. One of the most widely used methods to achieve this is association rule mining. However, while previous research exists on applying traditional association rule mining on ontologies, no approach has, to date, exploited the advantages brought by using the structure of these ontologies in computing rule interestingness measures.ResultsWe introduce a method that combines concept similarity metrics, formulated using the intrinsic structure of a given ontology, with traditional interestingness measures to compute semantic interestingness measures in the process of association rule mining. We apply the method in our domain of interest – bone dysplasias – using the core ontologies characterising it and an annotated dataset of patient clinical summaries, with the goal of discovering implicit relationships between clinical features and disorders. Experimental results show that, using the above mentioned dataset and a voting strategy classification evaluation, the best scoring traditional interestingness measure achieves an accuracy of 57.33%, while the best scoring semantic interestingness measure achieves an accuracy of 64.38%, both at the recall cut-off point 5.ConclusionsSemantic interestingness measures outperform the traditional ones, and hence show that they are able to exploit the semantic similarities inherently present between ontological concepts. Nevertheless, this is dependent on the domain, and implicitly, on the semantic similarity metric chosen to model it.

Finding, capturing and describing characteristic features represents a key aspect in disorder definition, diagnosis and management. This process is particularly challenging in the case of rare disorders, due to the sparse nature of data and expertise. From a computational perspective, finding characteristic features is associated with some additional major challenges, such as formulating a computationally tractable definition, devising appropriate inference algorithms or defining sound validation mechanisms. In this paper we aim to deal with each of these problems in the context provided by the skeletal dysplasia domain. We propose a clear definition for characteristic phenotypes, we experiment with a novel, class association rule mining algorithm and we discuss our lessons learned from both an automatic and human-based validation of our approach.

A lack of mature domain knowledge and well established guidelines makes the medical diagnosis of skeletal dysplasias (a group of rare genetic disorders) a very complex process. Machine learning techniques can facilitate objective interpretation of medical observations for the purposes of decision support. However, building decision support models using such techniques is highly problematic in the context of rare genetic disorders, because it depends on access to mature domain knowledge. This paper describes an approach for developing a decision support model in medical domains that are underpinned by relatively sparse knowledge bases. We propose a solution that combines association rule mining with the Dempster-Shafer theory (DST) to compute probabilistic associations between sets of clinical features and disorders, which can then serve as support for medical decision making (e.g., diagnosis). We show, via experimental results, that our approach is able to provide meaningful outcomes even on small datasets with sparse distributions, in addition to outperforming other Machine Learning techniques and behaving slightly better than an initial diagnosis by a clinician.