Mapping Concepts via Locality Sensitive Hashing (LSH)

Did you know that options other than Usagi exist for medical concept mapping? At edenceHealth NV, our R&D team is actively researching new and innovating ways to efficiently create more accurate links, or mappings, between source medical codes (e.g. l'amygdalo-pharyngite) and standardized medical concepts (e.g. tonsillitis). We have developed two different products to assist data partners with this mapping process: edenceMapper, which is a medical concept suggestion engine, and edenceReviewer, which enables collaborative review and approval of suggested mappings. The latest update to the edenceMapper framework has been implemented by Kristof Van Loock during his recent internship; it is a Locality-Sensitive Hashing (LSH) approach and performs impressively in both accuracy and throughput.

LSH is a powerful technique and we use it to link a given source concept to a small set of very similar standard concepts. The main advantage of LSH is that it can avoid a full pair-wise check when searching for the most similar standard concepts. As an example, if asked to identify a particular fruit, for instance an orange, LSH would only need to check the most likely subset of possible fruits – all those that are the color orange – to generate a match, rather than needing to scan all fruit varieites before its decision. Or in other words, LSH only evaluates the most promising candidate mappings and thus enables a fast and accurate way of finding the correct match.

In comparison with Usagi, the advantages of LSH are quickly apparent. The LSH approach can map unseen data sets to standard medical concepts nearly an order of magnitude faster than Usagi while maintaining comparable (and in some cases, better) accuracy. Keep in mind that these results are just for the prototype LSH mapper on its own, and the comparison was made on a test set of ICD10 codes with ‘ground truth’ standard mappings produced largely by teams using Usagi. When integrated into edenceHealth’s broader mapping framework, we can leverage the performance of all of our existing algorithms to produce even better mappings. You can already guess what that means for your future concept mapping projects 😉

Would you like to use our mapping framework, or are you eager to help us improve it? Don’t hesitate to reach out, we’d be very happy to hear from you!