Machine learning engine and rule engine for document auto-population using historical and contextual data
Methods, systems, and computer-readable media are disclosed herein for a machine learning engine and rule engine that leverage historical and contextual data to intelligently identify, score, and suggest one or more documents for auto-population of a graphical user interface. The machine learning and rule engine employ vectorization and clustering technique to identify, score, and suggest the most factually accurate and contextually relevant documents as selectable candidates for electronic documentation. Further, the selection, rejection, or modification of the candidate documents are ingested by the machine learning engine and/or the rule engine to update a clustering algorithm and/or to update a relevance scoring algorithm, which are then utilized for subsequent instances.
1. One or more non-transitory media having instructions that, when executed by one or more processors, cause the one or more processors to facilitate a plurality of operations, the operations comprising:
generating a combined vector, associated with healthcare information, based on a first set of historical text data and based further on a second set of historical contextual data, the first set of historical text data differing from the second set of historical contextual data;
generating a plurality of clusters based on an electronic machine-learning clustering model using the combined vector as input, wherein:
(a) the electronic machine-learning clustering model is trained based on data associated with instances of first information selected from a group comprising the first set of historical text data and the second set of historical contextual data, and
(b) the plurality of clusters includes the first set of historical text data and further include the second set of historical contextual data;
identifying a primary cluster in the plurality of clusters that is a best match to the combined vector;
performing natural language processing on second information associated with the primary cluster;
based at least on the natural language processing:
assigning a relevance score, for at least a portion of a plurality of text blocks in the primary cluster, corresponding to a respective relevance of at least a portion of the plurality of text blocks to information associated with at least a portion of the combined vector;
identifying a primary text block in the plurality of text blocks of the primary cluster having a highest relevance score; and
in response to the identifying of the primary text block in the plurality of text blocks of the primary cluster having the highest relevance score:
communicating the primary text block for display as a recommended selection for automatic population into at least a narrative field, of a healthcare related electronic document, configured to receive free-form narrative text.
2. The one or more non-transitory media of claim 1 , wherein generating the combined vector comprises:
identifying categorical data in one or more of the first set of historical text data or the second set of historical contextual data, wherein the categorical data is incompatible for input to the electronic machine-learning clustering model; and
transforming the categorical data into dummy variables, wherein the dummy variables are used to generate the combined vector.
3. The one or more non-transitory media of claim 1 , wherein the operations further comprise reducing data sparsity of the combined vector by:
applying Principle Component Analysis (PCA) to the combined vector to identify sparse columns of the combined vector, wherein the combined vector has n dimensions, one dimension for each of a plurality of features; and
projecting the combined vector having n dimensions to a k dimensional vector space to reduce a total quantity of the sparse columns, wherein k<n.
4. The one or more non-transitory media of claim 1 , wherein the electronic machine-learning clustering model includes a Density-Based Spatial Clustering of Applications with Noise (DBSCAN).
5. The one or more non-transitory media of claim 1 , wherein when generating the plurality of clusters, a total quantity of the plurality of clusters is determined by one or more characteristics of the first set of text data and the second set of historical contextual data in the combined vector, wherein a plurality of cluster tags are attached to the combined vector, and wherein the plurality of cluster tags map features of the first set of text data and the second set of historical contextual data to one or more of the plurality of clusters.
6. The one or more non-transitory media of claim 1 , wherein the plurality of clusters includes one or more portions of physician-specific patient-encounter data or patient-specific context data.
7. The one or more non-transitory media of claim 6 , wherein the primary cluster is identified, subsequent to reducing a data sparsity, by:
identifying one cluster in the plurality of clusters that corresponds to a set of text blocks in one or more of the physician-specific patient-encounter data or the patient-specific context data, wherein the set of text blocks in the one cluster have a greatest similarity to the combined vector relative to other clusters in the plurality of clusters; and
designating the one cluster as the primary cluster based on the one cluster have the greatest similarity to the combined vector relative.
8. The one or more non-transitory media of claim 1 , wherein the primary text block includes text selected from a group comprising unstructured text, text other than a medical or billing code, and text other than a medical concept.
9. The one or more non-transitory media of claim 1 , wherein the narrative field is configured to receive narrative text from a clinician describing a patient encounter, and wherein the primary text block includes unstructured text.
10. The one or more non-transitory media of claim 1 , wherein the primary text block comprises a narrative note.
11. The one or more non-transitory media of claim 1 , wherein the primary text block includes a narrative note describing a patient encounter.
12. The one or more non-transitory media of claim 1 , wherein the primary text block includes a narrative note created by a clinician and describing a patient encounter.
13. The one or more non-transitory media of claim 1 , wherein the primary text block includes content other than a medical code, billing code, or medical concept.
14. The one or more non-transitory media of claim 1 , wherein the primary text block includes content other than a medical concept.
15. The one or more non-transitory media of claim 1 , wherein the primary text block includes first text of a first type and second text of a second type, wherein the first type excludes structured text, medical and billing codes, and medical concepts, and wherein the second type includes content selected from a group comprising structured text, a medical or billing code, and a medical concept.
16. The one or more non-transitory media of claim 1 , wherein the primary text block includes first text of a first type and second text of a second type, wherein the first type excludes structured text, and wherein the second type includes structured text.
17. The one or more non-transitory media of claim 1 , wherein generating the combined vector comprises generating a physician-specific vector from the first set of historical text data and a patient-specific vector from the second set of historical contextual data.
18. The one or more non-transitory media of claim 17 , wherein the combined vector is generated based on the physician-specific vector and the patient-specific vector.
19. The one or more non-transitory media of claim 1 , wherein the primary text block includes a narrative input by a clinician documenting reasons and details regarding a clinical encounter with a patient, including a medical history of the patient and a chief complaint of the patient.
20. The one or more non-transitory media of claim 1 , wherein the primary text block includes first text of a first type and second text of a second type, wherein the first type excludes structured text, medical and billing codes, and medical concepts, and wherein the second type includes a structured text element, a medical or billing code, and a medical concept.
21. The one or more non-transitory media of claim 1 , wherein the primary text block includes first text of a first type and second text of a second type, wherein the first type excludes structured text, medical and billing codes, and medical concepts, and wherein the second type includes a medical concept.
22. The one or more non-transitory media of claim 1 , wherein the primary text block includes a plurality of elements selected from a group comprising: a clinical fact, a clinical observation, and a contextual data item.
23. The one or more non-transitory media of claim 1 , wherein the primary text block includes a medical assessment describing a patient encounter.
24. The one or more non-transitory media of claim 1 , wherein the operations further comprise decreasing a dimensionality of the combined vector prior to inputting the combined vector to the electronic machine-learning clustering model to generate the plurality of clusters.
25. The one or more non-transitory media of claim 1 , wherein the operations further comprise communicating the primary text block for display as a recommended selection for automatic population of a text assessment field of an electronic document.
26. The one or more non-transitory media of claim 1 , wherein the primary text block includes a plurality of clinical assessments, and wherein the primary text block is communicated for display as a recommended selection for automatic population of a text narrative field of an electronic document.
27. The one or more non-transitory media of claim 1 , wherein the operations further comprise:
retraining the electronic machine-learning clustering model based on additional data associated with the instances; and
reinputting to the retrained electronic machine-learning clustering model information associated with an additional combined vector to generate an additional plurality of clusters.
28. A system having one or more processors configured to facilitate a plurality of operations, the operations comprising:
generating a combined vector, associated with healthcare information, based on a first set of historical text data and based further on a second set of historical contextual data, the first set of historical text data differing from the second set of historical contextual data;
generating a plurality of clusters based on an electronic machine-learning clustering model using the combined vector as input, wherein:
(a) the electronic machine-learning clustering model is trained based on data associated with instances of first information selected from a group comprising the first set of historical text data and the second set of historical contextual data, and
(b) the plurality of clusters includes the first set of historical text data and further include the second set of historical contextual data;
identifying a primary cluster in the plurality of clusters that is a best match to the combined vector;
performing natural language processing on second information associated with the primary cluster;
based at least on the natural language processing:
assigning a relevance score, for at least a portion of a plurality of text blocks in the primary cluster, corresponding to a respective relevance of at least a portion of the plurality of text blocks to information associated with at least a portion of the combined vector;
identifying a primary text block in the plurality of text blocks of the primary cluster having a highest relevance score; and
in response to the identifying of the primary text block in the plurality of text blocks of the primary cluster having the highest relevance score:
communicating the primary text block for display as a recommended selection for automatic population into at least a narrative field, of a healthcare related electronic document, configured to receive free-form narrative text.
29. The system of claim 28 , wherein generating the combined vector comprises:
identifying categorical data in one or more of the first set of historical text data or the second set of historical contextual data, wherein the categorical data is incompatible for input to the electronic machine-learning clustering model; and
transforming the categorical data into dummy variables, wherein the dummy variables are used to generate the combined vector.
30. The system of claim 28 , wherein the operations further comprise reducing data sparsity of the combined vector by:
applying Principle Component Analysis (PCA) to the combined vector to identify sparse columns of the combined vector, wherein the combined vector has n dimensions, one dimension for each of a plurality of features; and
projecting the combined vector having n dimensions to a k dimensional vector space to reduce a total quantity of the sparse columns, wherein k<n.
31. The system of claim 28 , wherein the electronic machine-learning clustering model includes a Density-Based Spatial Clustering of Applications with Noise (DBSCAN).
32. The system of claim 28 , wherein when generating the plurality of clusters, a total quantity of the plurality of clusters is determined by one or more characteristics of the first set of text data and the second set of historical data in the combined vector, wherein a plurality of cluster tags are attached to the combined vector, and wherein the plurality of cluster tags map features of the first set of text data and the second set of historical contextual data to one or more of the plurality of clusters.
33. The system of claim 28 , wherein the primary cluster is identified, subsequent to reducing a data sparsity, by:
identifying one cluster in the plurality of clusters that corresponds to a set of text blocks in one or more of physician-specific patient-encounter data or patient-specific context data, wherein the set of text blocks in the one cluster have a greatest similarity to the combined vector relative to other clusters in the plurality of clusters; and
designating the one cluster as the primary cluster based on the one cluster have the greatest similarity to the combined vector relative.
34. A method, comprising:
generating a combined vector, associated with healthcare information, based on a first set of historical text data and based further on a second set of historical contextual data, the first set of historical text data differing from the second set of historical contextual data;
generating a plurality of clusters based on an electronic machine-learning clustering model using the combined vector as input, wherein:
(a) the electronic machine-learning clustering model is trained based on data associated with instances of first information selected from a group comprising the first set of historical text data and the second set of historical contextual data, and
(b) the plurality of clusters includes the first set of historical text data and further include the second set of historical contextual data;
identifying a primary cluster in the plurality of clusters that is a best match to the combined vector;
performing natural language processing on second information associated with the primary cluster;
based at least on the natural language processing:
assigning a relevance score, for at least a portion of a plurality of text blocks in the primary cluster, corresponding to a respective relevance of at least a portion of the plurality of text blocks to information associated with at least a portion of the combined vector;
identifying a primary text block in the plurality of text blocks of the primary cluster having a highest relevance score; and
in response to the identifying of the primary text block in the plurality of text blocks of the primary cluster having the highest relevance score:
communicating the primary text block for display as a recommended selection for automatic population into at least a narrative field, of a healthcare related electronic document, configured to receive free-form narrative text.
35. The method of claim 34 , further comprising: reducing a data sparsity of the combined vector prior to generating the plurality of clusters using the combined vector as input to the electronic machine-learning clustering model.
36. The method of claim 35 , further comprising receiving a user input that includes a selection of the primary text block for automatic population of the healthcare related electronic document, communicating a value of one to a rule engine as feedback for subsequent cluster identifications.
37. The method of claim 34 , further comprising receiving user input that does not select the primary text block.
38. The method of claim 37 , wherein when the user input does not select the primary text block for automatic population of the electronic healthcare related document, communicating a value of zero to a rule engine, as feedback for subsequent cluster identifications.
39. The method of claim 34 , further comprising:
combining the first set of historical text data and the second set of historical contextual data to generate the combined vector, wherein:
(a) the first set of historical text data is stored at a first database of clinician notes, and
(b) the second set of historical contextual data is stored at a second database of patient demographic information.
40. The method of claim 34 , further comprising receiving user input that includes an edited version of the primary text block, wherein when the user input includes a selection of the edited version of the primary text block for automatic population of the healthcare related electronic document, communicating the edited version of the primary text block to a machine learning engine, wherein the edited version is stored as new historical data.