Extraction of patient-level clinical events from unstructured clinical documentation
View Patent ↗Some embodiments of the present disclosure provide a framework for using unsupervised artificial intelligence to automatically abstract and align clinical facets. The approach of the present application may be shown to reduce human involvement and, accordingly, enhance privacy compliance. Aspects of the present application relate to a process of self-learning from the data available. Accordingly, aspects of the present application may be shown to be resilient to the appearance of new concepts and facets in future data. Additionally, aspects of the present application may be shown to adapt well when presented with different languages, different styles of documentation and different clinical domains. Aspects of the present application relate to processing unstructured, non-fielded data, such as clinical notes, admission and discharge summaries, surgical notes, lab reports and imaging reports. These notes may be considered to contain hidden insights in the clinical domain. Additionally, these notes may be considered to contain data that may not be captured elsewhere in a readily usable way. Aspects of the present application may be shown to support analysis of large size populations at a relatively low incremental cost.
1. A method comprising:
accessing, as text input, unstructured clinical documentation;
converting the unstructured clinical documentation to vector representations of semantics for the text input in context of a given subdomain;
using an artificial intelligence (AI) model to encode the vector representations to obtain embeddings representative of the text input; and
processing the embeddings to, thereby, generate, as output, a journal specific to a patient, the journal containing a plurality of entries arranged by facet through time, wherein, for a given facet, each entry, among the plurality of entries, includes a sentence from the text input, wherein the sentence has been determined to be related to the given facet.
2. The method of claim 1 , further comprising training the AI model.
3. The method of claim 2 , wherein the training the AI model comprises pre-training the AI model on a language to obtain a hypothesis function and embedding boundaries.
4. The method of claim 3 , wherein the pre-training the AI model comprises using a Query-Key-Value (QKV) softmax loss function.
5. The method of claim 3 , wherein the pre-training the AI model comprises using a mask infill task.
6. The method of claim 3 , wherein the pre-training the AI model comprises using a next sentence prediction task.
7. The method of claim 3 , wherein the training the AI model further comprises adjusting the hypothesis function.
8. The method of claim 7 , wherein the adjusting the hypothesis function comprises using a modified softmax function with dot product.
9. The method of claim 3 , wherein the training the AI model further comprises soft adjusting the embedding boundaries.
10. The method of claim 9 , wherein the soft adjusting the embedding boundaries comprises using a modified triplet loss function.
11. The method of claim 1 , wherein the facet comprises an attribute of the patient.
12. The method of claim 11 , wherein the attribute comprises a medication.
13. The method of claim 11 , wherein the attribute comprises a comorbidity.
14. The method of claim 1 , further comprising determining that the given entry is relevant to the facet.
15. The method of claim 1 , wherein the given subdomain comprises an English medical subdomain.