Systems and methods for extracting patient diagnostics from disparate

1. A method comprising,

receiving scanned documents, wherein the scanned documents comprise unstructured data;

performing optical character recognition of the scanned documents to produce text data for each page of the scanned documents, wherein the text data for each page comprises a sequence of words stored together with their location as x, y coordinates;

dividing each page of the scanned documents into subsections, wherein the dividing each page into subsections comprises applying a page blocker, wherein the page blocker identifies vectors of pixel density in the vertical and horizontal direction to identify vertical and horizontal page breaks;

using the text data to identify a structure type of each subsection of a page, wherein the structure type includes at least one of a table and text paragraph, wherein the identifying a structure type includes applying a structure classifier, wherein the structure classifier comprises a multi-stage neural network that assigns a probability of structure type to each subsection of a page;

using the text data to label each subsection of a page with a semantic type, wherein the semantic type defines a context surrounding collection of information in a subsection; and

using the text data for each subsection of a page to identify medical concepts.

2. The method of claim 1 , wherein the dividing each page into subsections includes summing the foreground color along the vertical and horizontal vectors of pixel density.

3. The method of claim 2 , wherein the dividing each page into subsections includes reviewing each vertical pixel density vector to identify an occurrence of a vertical section break, wherein the occurrence of a vertical section break comprises the summed foreground color value for a vertical pixel density vector falling below a threshold value.

4. The method of claim 3 , wherein the dividing each page into subsections includes reviewing each horizontal pixel density vector to identify an occurrence of a horizontal section break, wherein the occurrence of a horizontal section break comprises the summed foreground color value for a horizontal pixel density vector falling below a threshold value.

5. The method of claim 1 , wherein the multi-stage neural network comprises a first independent neural network comprising four (4) 2-dimension convolutional layers with 65 filters and max pooling.

6. The method of claim 5 , wherein a 100×100 matrix of vertical and horizontal pixel density values of a subsection is fed into the first independent neural network.

7. The method of claim 6 , wherein the multi-stage neural network comprises a second independent neural network comprising three (3) 1-dimension convolutional layers with a single filter and max pooling.

8. The method of claim 7 , wherein a 100 element vector of horizontal pixel density values of a subsection is fed into the second independent neural network.

9. The method of claim 8 , wherein the multi-stage neural network comprises a third independent neural network comprising three (3) 1-dimension convolutional layers with a single filter and max pooling.

10. The method of claim 9 , wherein a 100 element vector of vertical pixel density values is fed into the third independent neural network.

11. The method of claim 10 , wherein results of the first neural network, the second neural network, and the third neural network are combined and fed into a fourth fully connected 3 layer neural network for classification of a subsection.

12. The method of claim 1 , wherein the labelling of each subsection with a semantic type comprises applying a text classifier to each subsection of a page, wherein the text classifier comprises an unsupervised learning algorithm for obtaining vector representations of words in each subsection.

13. The method of claim 12 , wherein the unsupervised learning algorithm comprises a Global Vectors for Word Representation model trained on samples of text from hospital documentation.

14. The method of claim 13 , wherein the semantic type label includes at least one of medications, physical exam, lab results, radiology results, therapy notes, nursing notes, patient information, hospital boilerplate, and patient education.

15. The method of claim 14 , wherein upon an occurrence of the structure classifier assessing a subsection of a page with high probability of being a table, the page blocker tests a potential split of the subsection into subdivisions using the structure classifier and the text classifier.

16. The method of claim 15 , wherein the testing comprises assigning a semantic unity score to each subdivision, wherein the semantic unity score indicates a probability that each subdivision is correctly labelled by the text classifier, wherein the semantic unity score is computed using a Gini Impurity method.

17. The method of claim 16 , maintaining the potential subdivision upon the successful occurrence of a first event, wherein the first event includes the structure type probability of being a table for at least one of the subdivisions improving and the semantic unity score of both subdivisions improving.

18. The method of claim 1 , wherein the identifying the medical concepts comprises applying a medical concept extractor to the text data for each subsection of a page, wherein the medical concept extractor comprises a language model built using an open source ScispaCy model, wherein the model is trained on biomedical documents.

19. The method of claim 18 , translating the medical concepts and information of at least one of the structure type and semantic type into entries for a Minimum Data Set, wherein the Minimum data set comprises a set of data elements for mandatory collection and reporting assessments relating to all residents in Medicare and Medicaid certified nursing homes.