Extending electronic medical records vector models with knowledge graphs to improve hospitalization prediction.

BACKGROUND: Artificial intelligence methods applied to electronic medical records (EMRs) hold the potential to help physicians save time by sharpening their analysis and decisions, thereby improving the health of patients. On the one hand, machine learning algorithms have proven their effectiveness in extracting information and exploiting knowledge extracted from data. On the other hand, knowledge graphs capture human knowledge by relying on conceptual schemas and formalization and supporting reasoning. Leveraging knowledge graphs that are legion in the medical field, it is possible to pre-process and enrich data representation used by machine learning algorithms. Medical data standardization is an opportunity to jointly exploit the richness of knowledge graphs and the capabilities of machine learning algorithms. METHODS: We propose to address the problem of hospitalization prediction for patients with an approach that enriches vector representation of EMRs with information extracted from different knowledge graphs before learning and predicting. In addition, we performed an automatic selection of features resulting from knowledge graphs to distinguish noisy ones from those that can benefit the decision making. We report the results of our experiments on the PRIMEGE PACA database that contains more than 600,000 consultations carried out by 17 general practitioners (GPs). RESULTS: A statistical evaluation shows that our proposed approach improves hospitalization prediction. More precisely, injecting features extracted from cross-domain knowledge graphs in the vector representation of EMRs given as input to the prediction algorithm significantly increases the F1 score of the prediction. CONCLUSIONS: By injecting knowledge from recognized reference sources into the representation of EMRs, it is possible to significantly improve the prediction of medical events. Future work would be to evaluate the impact of a feature selection step coupled with a combination of features extracted from several knowledge graphs. A possible avenue is to study more hierarchical levels and properties related to concepts, as well as to integrate more semantic annotators to exploit unstructured data.

Biomedical word sense disambiguation with ontologies and metadata: automation meets accuracy.

BACKGROUND: Ontology term labels can be ambiguous and have multiple senses. While this is no problem for human annotators, it is a challenge to automated methods, which identify ontology terms in text. Classical approaches to word sense disambiguation use co-occurring words or terms. However, most treat ontologies as simple terminologies, without making use of the ontology structure or the semantic similarity between terms. Another useful source of information for disambiguation are metadata. Here, we systematically compare three approaches to word sense disambiguation, which use ontologies and metadata, respectively. RESULTS: The 'Closest Sense' method assumes that the ontology defines multiple senses of the term. It computes the shortest path of co-occurring terms in the document to one of these senses. The 'Term Cooc' method defines a log-odds ratio for co-occurring terms including co-occurrences inferred from the ontology structure. The 'MetaData' approach trains a classifier on metadata. It does not require any ontology, but requires training data, which the other methods do not. To evaluate these approaches we defined a manually curated training corpus of 2600 documents for seven ambiguous terms from the Gene Ontology and MeSH. All approaches over all conditions achieve 80% success rate on average. The 'MetaData' approach performed best with 96%, when trained on high-quality data. Its performance deteriorates as quality of the training data decreases. The 'Term Cooc' approach performs better on Gene Ontology (92% success) than on MeSH (73% success) as MeSH is not a strict is-a/part-of, but rather a loose is-related-to hierarchy. The 'Closest Sense' approach achieves on average 80% success rate. CONCLUSION: Metadata is valuable for disambiguation, but requires high quality training data. Closest Sense requires no training, but a large, consistently modelled ontology, which are two opposing conditions. Term Cooc achieves greater 90% success given a consistently modelled ontology. Overall, the results show that well structured ontologies can play a very important role to improve disambiguation. AVAILABILITY: The three benchmark datasets created for the purpose of disambiguation are available in Additional file 1.