Creating multiple prioritized clinical summaries using artificial intelligence
A system, method, and a computer program product for generating a clinical summary of a patient using artificial intelligence is provided. A patient data that includes unstructured data and structured data is collected from multiple computing devices. Natural language processing models determine clinical issues from the unstructured data. Active clinical issues are determined from the clinical issues. A knowledge graph generated using a relational language model determines treatments associated with the active clinical issues. Active diagnostic and treatment orders are determined from the structured data. Multiple summaries summarizing the active clinical issues, treatments, active diagnostic orders, and active treatment orders are determined using natural language generation models trained to summarize multiple tasks. The multiple summaries are aggregated into a single summary. The language of the summary is modified using a hyperparameter in a smoothing natural language model to direct the summary toward an audience having a particular type.
1 . A system comprising:
a memory configured to store instructions for an artificial intelligence engine;
a processor coupled to the memory and configured to read the instructions from the memory to cause the system to perform operations, the operations comprising:
receiving, over a network, patient data associated with a patient, the patient data including unstructured data collected from multiple computing systems, wherein the unstructured data includes at least audio data;
determining, using a plurality of natural language processing models in the artificial intelligence engine, a plurality of clinical issues of the patient based on the audio data;
summarizing, using a plurality of natural language generation models in the artificial intelligence engine that are each trained to generate a summary for a different summary task, the clinical issues into a plurality of notes that are different from each other;
aggregating, using an aggregating natural language generation model in the artificial intelligence engine, the plurality of notes into one or more aggregate notes associated with the patient, wherein the one or more aggregate notes are in a paragraph form;
smoothing, using the aggregating natural language generation model that includes a hyperparameter set to vary a language output of the aggregating natural language generation model based on a target audience, the one or more aggregate notes into smoothed one or more notes by rephrasing the one or more aggregate notes for the target audience; and
transmitting the smoothed one or more notes for display on at least one computing device.
2 . The system of claim 1 , wherein the patient data includes structured data and further comprising:
generating, using at least one predictive neural network model and the structured data, at least one risk score, wherein the at least one risk score corresponds to a patient outcome;
generating, using one of the plurality of natural language generation models and the at least one risk score, a patient note corresponding to the patient outcome; and
incorporating the patient note into the plurality of notes.
3 . The system of claim 2 , wherein the at least one risk score that corresponds to the patient outcome is a prediction of a patient discharge from a hospital.
4 . The system of claim 2 , wherein the at least one risk score that corresponds to the patient outcome is a prediction of a clinical risk of deterioration of the patient.
5 . The system of claim 1 , further comprising:
determining, using a knowledge graph, a plurality of treatments for the plurality of clinical issues; and
prioritizing the plurality of treatments for the plurality of clinical issues.
6 . The system of claim 5 , further comprising:
determining, using an auto-resolution module, that a treatment in the plurality of treatments is not required;
removing the treatment from the plurality of treatments; and
prioritizing the plurality of treatments without the removed treatment.
7 . The system of claim 6 , further comprising:
summarizing, using the plurality of natural language generation models, the prioritized plurality of treatments into a second plurality of notes; and
aggregating, using the aggregating natural language generation model, the second plurality of notes into the one or more aggregate notes associated with the patient.
8 . The system of claim 6 , further comprising:
determining, using one of the plurality of natural language processing models, that one of the plurality of treatments is a barrier to discharge; and
generating, using one of the plurality of natural language generations models, reasons for the barrier to discharge.
9 . The system of claim 1 , further comprising:
receiving an identifier in a message from the at least one computing device;
linking the identifier to a type of a user; and
setting the hyperparameter based on the type of the user,
wherein the smoothed one or more notes include different content depending on the hyperparameter.
10 . A method comprising:
receiving, over a network, patient data associated with a patient, the patient data including unstructured data collected from multiple computing systems, wherein the unstructured data includes audio data from an audio file;
determining, using a plurality of natural language processing models in an artificial intelligence engine executing on a processor, a plurality of clinical issues of the patient based on the audio data;
summarizing, using a plurality of natural language generation models in the artificial intelligence engine that are each trained to generate a summary for a different summary task, the clinical issues into a plurality of notes that are different from each other;
aggregating, using an aggregating natural language generation model in the artificial intelligence engine, the plurality of notes into one or more aggregate notes associated with the patient;
smoothing, using the aggregating natural language generation model that includes a hyperparameter set to vary a language output of the aggregating natural language generation model based on a target audience, the one or more aggregate notes into smoothed one or more notes by rephrasing the one or more aggregate notes for the target audience; and
transmitting the one or more aggregated notes for display on a computing device.
11 . The method of claim 10 , wherein the patient data includes structured data and further comprising:
generating, using at least one predictive neural network model and the structured data, at least one risk score, wherein the at least one risk score corresponds to a patient outcome;
generating, using one of the plurality of natural language generation models and the at least one risk score, a patient note corresponding to the patient outcome; and
incorporating the patient note into the plurality of notes.
12 . The method of claim 11 , wherein the at least one risk score that corresponds to the patient outcome is a prediction of a patient discharge from a hospital.
13 . The method of claim 11 , wherein the at least one risk score that corresponds to the patient outcome is a prediction of a clinical risk of deterioration of the patient.
14 . The method of claim 10 , further comprising:
determining, using a knowledge graph, a plurality of treatments for the plurality of clinical issues; and
prioritizing the plurality of treatments for the plurality of clinical issues.
15 . The method of claim 14 , further comprising:
determining, using an auto-resolution module, that a treatment in the plurality of treatments is not required;
removing the treatment from the plurality of treatments; and
prioritizing the plurality of treatments without the removed treatment.
16 . The method of claim 15 , further comprising:
summarizing, using the plurality of natural language generation models, the prioritized plurality of treatments into a second plurality of notes; and
aggregating, using the aggregating natural language generation model, the second plurality of notes into the one or more aggregate notes associated with the patient.
17 . The method of claim 15 , further comprising:
determining, using one of the plurality of natural language processing models, that one of the plurality of treatments is a barrier to discharge; and
generating, using one of the plurality of natural language generations models, reasons for the barrier to discharge.
18 . The method of claim 10 , further comprising:
setting the hyperparameter based on an identifier corresponding to a type of a user associated with the computing device.
19 . A non-transitory computer-readable medium storing instructions thereon, that when executed by a processor, cause the processor to perform operations, the operations comprising:
receiving, over a network, patient data associated with a patient, the patient data including unstructured data collected from multiple computing systems, wherein the unstructured data includes audio data;
determining, using a plurality of natural language processing models in an artificial intelligence engine, a plurality of clinical issues of the patient based on the audio data;
summarizing, using a plurality of natural language generation models in the artificial intelligence engine that are each trained to generate a summary for a different summary task, the clinical issues into a plurality of notes that are different from each other;
aggregating, using an aggregating natural language generation model in the artificial intelligence engine, the plurality of notes into one or more aggregate notes associated with the patient, wherein the one or more aggregate notes are in a paragraph form;
smoothing, using the aggregating natural language generation model that includes a hyperparameter set to vary a language output of the aggregating natural language generation model based on a target audience, the one or more aggregate notes into smoothed one or more notes by rephrasing the one or more aggregate notes for the target audience; and
transmitting the smoothed one or more notes for display on at least one computing device.
20 . The non-transitory computer-readable medium of claim 19 , wherein the audio data is extracted from an audio file recorded at the at least one computing device, wherein the smoothed one or more notes include different content depending on the hyperparameter.