US 11,790,171 B2
Computer-implemented natural language understanding of medical reports
Ron Vianu, New York, NY (US); W. Nathaniel Brown, New York, NY (US); Gregory Allen Dubbin, New York, NY (US); Daniel Robert Elgort, New York, NY (US); Benjamin L. Odry, West New York, NJ (US); Benjamin Sellman Suutari, New York, NY (US); and Jefferson Chen, New York, NY (US)
Assigned to Covera Health, New York, NY (US)
Filed by Covera Health, New York, NY (US)
Filed on Apr. 15, 2020, as Appl. No. 16/849,506.
Application 16/849,506 is a continuation in part of application No. 16/386,006, filed on Apr. 16, 2019, granted, now 11,521,716.
Prior Publication US 2020/0334416 A1, Oct. 22, 2020
Int. Cl. G06F 40/295 (2020.01); G16H 15/00 (2018.01); G06V 30/40 (2022.01); G06F 40/10 (2020.01); G06V 10/82 (2022.01); G06V 30/148 (2022.01); G06V 30/262 (2022.01); G06V 10/764 (2022.01); G06N 3/08 (2023.01); G06V 30/10 (2022.01)
CPC G06F 40/295 (2020.01) [G06F 40/10 (2020.01); G06V 10/764 (2022.01); G06V 10/82 (2022.01); G06V 30/153 (2022.01); G06V 30/262 (2022.01); G06V 30/40 (2022.01); G16H 15/00 (2018.01); G06N 3/08 (2013.01); G06V 30/10 (2022.01)] 20 Claims
OG exemplary drawing
 
1. A natural language understanding method, the method comprising:
obtaining a radiological report text, the radiological report text containing one or more clinical findings;
in response to detecting one or more errors in the radiological report text, substituting a best replacement candidate for each of the one or more errors, the best replacement candidate calculated by analyzing a plurality of character-level optical transformation costs weighted by a frequency analysis over a corpus corresponding to the radiological report text;
for each given word within the radiological report text:
obtaining a word embedding;
determining a plurality of character-level embeddings; and
concatenating the word embedding and the plurality of character-level embeddings to a neural network;
generating, using the neural network, a plurality of NER tagged spans for the radiological report text, the generating based on the concatenated word and character-level embedding s;
calculating a set of linked relationships for the plurality of NER tagged spans by:
generating one or more masked text sequences based on the radiological report text and determined pairs of potentially linked NER spans;
calculating a dense adjacency matrix based on attention weights obtained from providing the one or more masked text sequences to a Transformer deep learning network; and
performing graph convolutions over the calculated dense adjacency matrix via a graph convolutional network.