A remarkable, efficient platform for understanding and acting on medical data.
Put simply, our platform ingests unstructured text and does something actionable with it.
That sounds like NLP!
Sure, we can extract entities and perform other NLP tasks. Our application though is way more than NLP. When we started the company, like others we obsessed about traditional NLP metrics. But while extracting entities via NLP are useful for some use cases, most high value use cases require much more advanced technology. At this point, NLP is sort of trivial.
The way of the graph.
Our technology uses graph theory, a knowledge graph, inference, and other types of machine learning to solve problems. What is a graph? It is a way to model the connectedness of data. This connectedness can be measured several different ways depending on the application. Using a graph to model data is also a very flexible data format which, as a side benefit, allows adjustments on the fly.
Fixed ontologies? That’s old school.
While we use a knowledge graph, that does not mean our technology is dependent on pre-configuring ontologies. Quite the contrary, our technology can ingest unknown text and still model it effectively.
Multiple-label classification, check.
One of the uses cases in healthcare that our technology can solve is multi-label classification. Our technology can ingest a set of documents with labels assigned or unassigned and correctly classify new documents. Many problems in healthcare require multi-label classification. That is each document (or a set of documents) can be assigned a single or multiple labels. A document could be an operative note, pathology report, imaging, patient progress note or any other type of unstructured or structured data. The labels could be ICD-10 codes or CPT codes. Or they could be patients’ co-morbidities.
The term big data suggests an unlimited and substantial amount of data. In practice, there is not always an unlimited amount of data for some classes. So while 80% of the data may have sufficient training data, the other 20% can be comprised of relatively low frequency events. Unless an application can candle all types of cases, it is not production ready. Being able to handle imbalanced data sets is another key component of our application because a system that cannot handle imbalanced data is of little use in healthcare.
In summary, we build technology that empowers healthcare organizations to make use of the vast volumes of unstructured data present in electronic health records, medical literature, patient surveys, and more. Our technology automates processes such as medical coding, provides input for clinical decision support, and clinical research.