Systems and methods for weakly-supervised reportability and context prediction, and for multi-modal risk identification for patient populations
Presented herein are systems and methods for automated analysis of patient data. More particularly, in certain embodiments, the invention relates to systems and methods for predicting the context of a particular phrase (e.g. the name of a diagnosis/condition) in a clinical record of a patient using a reportability classifier. In another aspect, the invention relates to systems and methods for automatically identifying a potential care gap and/or adverse health trend for a patient from clinical data.
1 . A method for automatically assigning a context label to each of one or more named entity phrases from a clinical record of a patient, the method comprising:
receiving and/or accessing, by a processor of a computing device, the clinical record of the patient and a problem list of the patient;
automatically identifying, by the processor, the one or more named entity phrases in the clinical record of the patient;
determining, via a reportability classifier, the one or more named entity phrases correspond to an International Classification of Diseases (ICD) code at least associated with a “not reportable” code or a “family history” code, wherein the reportability classifier comprises a machine learning algorithm;
automatically assigning, by the processor, the context label to each of the one or more named entity phrases corresponding to the ICD code;
automatically triggering a human review workflow based on the context label of the one or more named entity phrases;
receiving, via a control interface, feedback from a human reviewer based at least in part on the human reviewer updating a status of a problem in the problem list; and
updating the machine learning algorithm of the reportability classifier using training data based at least in part on the feedback received from the human reviewer during the human review workflow,
wherein the training data further comprises a weakly-labeled training dataset,
wherein the reportability classifier uses distant supervision for updating the machine learning algorithm with the weakly-labeled training dataset,
wherein updating the machine learning algorithm further comprises replacing the one or more named entity phrases with a mask such that the reportability classifier is forced to key off surrounding text, wherein the mask yields one or more features from the surrounding text that are used by the reportability classifier for automatically assigning the context label to the one or more named entity phrases.
2 . The method of claim 1 , wherein the one or more named entity phrases comprises a diagnosis, a disease, a condition, a clinical entity, or the ICD code.
3 . The method of claim 1 , wherein the problem list comprises a combination of one or more of: an illness, an injury, a diagnosis, a condition, and a social problem.
4 . The method of claim 1 , wherein the problem list comprises one or more diagnoses.
5 . The method of claim 1 , wherein updating the status of the problem comprises adding the problem or removing the problem in the problem list following the feedback received from the human review workflow.
6 . The method of claim 5 , wherein the machine learning algorithm of the reportability classifier is a deep-learning algorithm.
7 . The method of claim 1 , wherein the training data further comprises data from the clinical record of the patient.
8 . The method of claim 1 , wherein the method comprises identifying, by the processor, the problem corresponding to the patient based on the one or more named entity phrases and the context label of the one or more named entity phrases.
9 . The method of claim 1 , wherein the method comprises determining, by the processor, whether each of the one or more named entity phrases is a negated concept by using the reportability classifier.
10 . The method of claim 1 , wherein the method comprises automatically identifying a care gap using the problem list.
11 . The method of claim 10 , wherein automatically identifying the care gap comprises identifying a therapy and/or a treatment the patient could be receiving but is not currently receiving.
12 . The method of claim 10 , wherein automatically identifying the care gap comprises recommending a therapy using the problem list.
13 . The method of claim 1 , wherein the clinical record is an electronic medical record (EMR).
14 . The method of claim 1 , wherein the clinical record comprises unstructured data.
15 . The method of claim 14 , wherein the unstructured data comprises clinical dictation, clinical notes, and/or clinical reports.
16 . The method of claim 14 , wherein the unstructured data comprises one or more strings of alphanumeric characters.
17 . The method of claim 1 , wherein the clinical record comprises laboratory results of the patient, patient vitals, patient demographics, patient problems, patient medications, clinical notes, clinical dictation, clinical reports, and/or past medical history.
18 . The method of claim 1 , wherein the method does not require receiving or accessing an encounter document comprising billable codes corresponding to the patient.
19 . The method of claim 1 , wherein the one or more named entity phrases is a disease, a condition, or a social circumstance.
20 . The method of claim 1 , wherein the method comprises automatically mapping, by the processor, each of the one or more named entity phrases to a corresponding ontological code by using a first code classifier.
21 . The method of claim 20 , wherein the corresponding ontological code is a Unified Medical Language System (UMLS) code, the ICD code, or an RxNorm Normalized Names (RXNORM) code.
22 . The method of claim 20 , wherein the corresponding ontological code is a high level code that is specific to a general disease or condition.
23 . The method of claim 20 , wherein the method comprises updating each corresponding ontological code by using a second code classifier to make each corresponding ontological code more specific.
24 . The method of claim 23 , wherein the first code classifier and/or the second code classifier is/are deep-learning algorithms.
25 . The method of claim 1 , wherein the reportability classifier comprises a neural network.
26 . The method of claim 1 , wherein the reportability classifier comprises a bidirectional long short-term memory (LSTM) network.
27 . The method of claim 1 , wherein the reportability classifier is a trained model, having been trained using a training dataset to determine the context label of the one or more named entity phrases in medical records of one or more patients.
28 . The method of claim 27 , wherein the training dataset comprises one or more clinical records comprising unstructured data from each of the one or more patients, wherein each of the one or more clinical records comprises the one or more named entity phrases.
29 . The method of claim 27 , wherein the training dataset does not include billing codes.
30 . The method of claim 27 , wherein the training dataset comprises clinical dictation and/or clinical notes for each of the one or more patients.
31 . The method of claim 27 , wherein the training dataset comprises the problem list for each of the one or more patients.
32 . The method of claim 31 , wherein the reportability classifier has been trained using the training dataset by masking the one or more named entity phrases in text of the medical records and classifying an unmasked portion of the text containing the one or more named entity phrases with the context label, wherein the context label is based on, at least, a presence or an absence of the one or more named entity phrases in the problem list of the patient.
33 . The method of claim 23 , wherein the second code classifier is a trained model, having been trained using a code training dataset to determine a specific ontological code corresponding to the one or more named entity phrases comprising a plurality of patient examples corresponding to the specific ontological code.
34 . The method of claim 33 , wherein the plurality of patient examples comprises at least 10 patient examples of the specific ontological code.
35 . The method of claim 33 , wherein the plurality of patient examples comprises between about 10 and about 25 patient examples of the specific ontological code.
36 . The method of claim 33 , wherein each of the plurality of patient examples comprises a clinical dictation comprising the one or more named entity phrases.
37 . The method of claim 1 , wherein the method comprises automatically identifying, by the processor, one or more drugs in the clinical record of the patient.
38 . The method of claim 37 , wherein the method comprises mapping, by the processor, at least one of the one or more drugs to the problem.
39 . The method of claim 38 , wherein the method comprises updating the problem list of the patient with the problem.
40 . A system for automatically assigning a context label to named entity phrases from a clinical record of a patient, the system comprising:
a processor; and
a memory having instructions stored thereon, wherein the instructions, when executed by the processor, cause the processor to perform the method of claim 1 .
41 . The method of claim 1 , wherein the reportability classifier comprises an artificial neural network (ANN) configured to process unstructured text of the clinical record.
42 . The method of claim 1 , wherein at least a portion of the reportability classifier is implemented on an application-specific integrated circuit (ASIC).
43 . The method of claim 1 , further comprising:
translating the ICD code into an appropriate “family history” ICD10 “z-code” in response to determining the one or more named entity phrases correspond to the “family history” code.
44 . The method of claim 1 , wherein the reportability classifier comprises Bidirectional Encoder Representations (BERT) trained on clinical notes.
45 . The method of claim 1 , wherein the reportability classifier comprises Bidirectional Encoder Representations (BERT).
46 . The method of claim 5 , wherein updating the status of the problem in the problem list further comprises identifying the problem as resolved, inactive, or active.
47 . The method of claim 46 , wherein updating the status of the problem in the problem list further comprises identifying the problem as improving or worsening.
48 . A method for automatically identifying diagnostic codes using a clinical record of a patient, the method comprising:
receiving and/or accessing, by a processor of a computing device, a problem list of the patient and the clinical record of the patient, wherein the clinical record comprises text;
automatically identifying, by the processor, one or more named entity phrases in the text of the clinical record, thereby generating one or more identified named entity phrases;
automatically classifying, by the processor, at least one of the one or more identified named entity phrases to a diagnostic code, thereby generating a classified named entity phrase, wherein the diagnostic code is an International Classification of Diseases (ICD) code at least associated with a “not reportable” code or a “family history” code by using a reportability classifier, wherein the reportability classifier comprises a machine learning algorithm;
automatically mapping, by the processor, the classified named entity phrase to the diagnostic code by using a code classifier;
automatically triggering a human review workflow based on the diagnostic code of the classified named entity phrase;
receiving, via a control interface, feedback from a human reviewer, wherein the feedback is based at least in part on the human reviewer updating a status of a problem in the problem list; and
updating the machine learning algorithm of the reportability classifier using training data based at least in part on the feedback received from the human reviewer during the human review workflow,
wherein the training data further comprises a weakly-labeled training dataset,
wherein the reportability classifier uses distant supervision for updating the machine learning algorithm with the weakly-labeled training dataset,
wherein updating the machine learning algorithm further comprises replacing the one or more identified named entity phrases with a mask such that the reportability classifier is forced to key off surrounding text, wherein the mask yields one or more features from the surrounding text that are used by the reportability classifier for automatically classifying the at least one of the one or more identified named entity phrases.
49 . A method of training a reportability classifier to assign a context label to a named entity phrase in a clinical record of a patient, the method comprising:
receiving, by a processor of a computing device, (i) the clinical record and (ii) a problem list and/or a billing history for the patient, wherein the clinical record comprises text;
identifying, by the processor, the named entity phrase in the text of the clinical record, thereby generating an identified named entity phrase;
mapping, by the processor, the identified named entity phrase to a diagnostic code, wherein the diagnostic code is an International Classification of Diseases (ICD) code at least associated with a “not reportable” code or a “family history” code, wherein the reportability classifier comprises a machine learning algorithm;
masking, by the processor, the identified named entity phrase with a mask such that the reportability classifier is forced to key off surrounding text, wherein the mask yields one or more features from the surrounding text that are used by the reportability classifier to assign the context label, thereby generating a masked portion of the text;
assigning, by the processor, the context label to the named entity phrase corresponding to the ICD code using the reportability classifier based on the masked portion of the text and the one or more features from the surrounding text, thereby generating an assigned context label;
automatically triggering, by the processor, a human review workflow based on the assigned context label of the named entity phrase;
receiving, by the processor via a control interface, feedback from a human reviewer based at least in part on the human reviewer updating a status of a problem in the problem list; and
updating, by the processor, the machine learning algorithm of the reportability classifier using training data based at least in part on the feedback received from the human reviewer during the human review workflow,
wherein the training data further comprises a weakly-labeled training dataset,
wherein the reportability classifier uses distant supervision for updating the machine learning algorithm with the weakly-labeled training dataset,
wherein the masked portion of the text is further used for updating the machine learning algorithm.
50 . The method of claim 49 , wherein the reportability classifier comprises a learning algorithm.
51 . The method of claim 49 , wherein the masked portion of the text comprises, at least, the named entity phrase.
52 . The method of claim 49 , wherein the reportability classifier is a neural network.
53 . The method of claim 49 , wherein the reportability classifier comprises a member selected from a group consisting of: Long short term memory (LSTM), Bilateral LSTM (BiLSTM), Embeddings from Language Models (Elmo), Bidirectional Encoder Representations (BERT), the BERT trained on clinical notes (Clinical BERT), Distilled BERT (DistilBERT), and A Lite BERT (ALBERT).
54 . The method of claim 49 , wherein the method does not require receiving an encounter document of the patient.
55 . The method of claim 49 , wherein the method does not require receiving one or more billable codes corresponding to a disease or a condition of the patient.
56 . The method of claim 49 , wherein the clinical record comprises clinical notes and/or clinical dictation.
57 . The method of claim 49 , wherein the clinical record comprises unstructured text.
58 . The method of claim 49 , wherein the context label is indicative of whether or not the named entity phrase is reportable.
59 . The method of claim 49 , further comprising:
translating the ICD code into an appropriate “family history” ICD10 “z-code” in response to determining the named entity phrase corresponds to the “family history” code.
60 . The method of claim 49 , wherein the reportability classifier comprises Bidirectional Encoder Representations (BERT) trained on clinical notes.
61 . The method of claim 49 , wherein the reportability classifier comprises a Long short term memory (LSTM).
62 . A method for automatically identifying a potential care gap and/or a potential adverse health trend for a patient, the method comprising:
receiving and/or accessing, by a processor of a computing device, clinical data related to the patient and a problem list of the patient;
automatically identifying, by the processor, the potential care gap and/or the potential adverse health trend for the patient from the clinical data using a rule-based algorithm and a machine learning algorithm;
determining, via a reportability classifier, one or more named entity phrases correspond to an International Classification of Diseases (ICD) code at least associated with a “not reportable” code or a “family history” code, wherein the reportability classifier comprises the rule-based algorithm and the machine learning algorithm;
automatically assigning, by the processor, a context label to each of the one or more named entity phrases corresponding to the ICD code, thereby generating an assigned context label;
automatically triggering a human review workflow based on the assigned context label of the one or more named entity phrases;
receiving, via a control interface, feedback from a human reviewer based at least in part on the human reviewer updating a status of a problem in the problem list; and
updating the machine learning algorithm of the reportability classifier using training data based at least in part on the feedback received from the human reviewer during the human review workflow
wherein the training data further comprises a weakly-labeled training dataset,
wherein the reportability classifier uses distant supervision for updating the machine learning algorithm with the weakly-labeled training dataset,
wherein updating the machine learning algorithm further comprises replacing the one or more named entity phrases with a mask such that the reportability classifier is forced to key off surrounding text, wherein the mask yields one or more features from the surrounding text that are used by the reportability classifier for automatically assigning the context label to the one or more named entity phrases.
63 . The method of claim 62 , wherein the machine learning algorithm comprises natural language processing (NLP) software.
64 . The method of claim 62 , wherein the method further comprises notifying the patient of a finding of the potential care gap and/or the potential adverse health trend for the patient in response to automatically triggering the human review workflow.
65 . The method of claim 62 , wherein the training data further comprises data from a clinical record of the patient.
66 . The method of claim 62 , wherein the clinical data related to the patient comprises a healthcare provider narrative that references a medication indicating a condition that is missing from an electronic health record of the patient, thereby triggering the human review workflow.
67 . The method of claim 62 , wherein automatically identifying the potential care gap and/or the potential adverse health trend comprises a member selected from a group consisting of: (i) determining whether a HEDIS-defined (Healthcare Effectiveness Data and Information Set) intervention applies for the patient, (ii) tracking changes in one or more vital signs for the patient over time and alerting a care navigator for the patient of the potential adverse health trend based on the changes, (iii) identifying a change in medication prescribed and/or taken by the patient and alerting the care navigator for the patient of the change, (iv) identifying that the patient may not be following a prescribed protocol or initiative and alerting the care navigator for the patient accordingly, and (v) monitoring data from a remote monitoring program and alerting the care navigator for the patient of an anomalous reading.
68 . A system for automatically identifying a potential care gap and/or adverse health trend for a patient, the system comprising:
a processor; and
a memory having instructions stored thereon, wherein the instructions, when executed by the processor, cause the processor to perform the method of claim 61 .