Automated summarization of a hospital stay using machine learning
A method and a device are provided for generating a summary of a hospital stay by a patient, including maintaining a database of electronic medical records (EMRs) where the EMRs include clinical notes pertaining to a patient during a time interval, identifying a set of significant physician notes for the time interval, generating a candidate set of summaries for each of the significant physician notes, for each significant physician note, analyzing the factuality of each of the candidate summaries and selecting the most factual summary, and generating a daily section for inclusion in a hospital course section of a discharge note that includes the selected factual summary for each of the significant physician notes.
1. A computer-implemented method of using a transformer machine learning model to generate a summary of a hospital stay by a patient, comprising:
maintaining a database of electronic medical records (EMRs) for patients, the EMRs include clinical notes written by clinicians that provide details about patients' hospital stays, the clinical notes including physician notes provided by physicians;
training the transformer model on a classification training dataset based on a daily course section of the clinical notes;
selecting from the database physician notes for a patient's hospital stay during a designated time interval;
identifying from the selected physician notes a set of significant physician notes for the time interval;
generating a candidate set of summaries for each of the identified significant physician notes, wherein a summary includes one or more sentences that are abstractively generated using the transformer model;
for each significant physician note, performing a factuality analysis on the candidate summaries, wherein the factuality analysis (1) eliminates any summary that includes a non-factual medical word or phrase, and (2) selects the highest probability remaining summary; and
generating a daily section for inclusion in a hospital course section of a discharge note, the generated daily section including the selected factual summary for each of the significant physician notes.
2. The method of claim 1 further comprising:
retrieving electronic medical records for a patient;
translating the received medical records into a standard format; and
storing the translated medical records in the database.
3. The method of claim 1 wherein the time interval is a day.
4. The method of claim 1 further comprising:
performing a regular expression (regex) processing step on the selected physician notes to extract the most salient and useful content for further processing.
5. The method of claim 1 wherein the physician note is selected from the group consisting of Progress Notes, Procedure Notes, Op Notes, and Consult Notes.
6. The method of claim 1 wherein identifying a set of significant physician notes is performed using a machine learning model.
7. The method of claim 6 wherein the machine learning model is a transformer model and the transformer model is fine-tuned using a dataset comprising clinical notes.
8. The method of claim 1 wherein the transformer model is fine-tuned using a dataset comprising clinical notes.
9. The method of claim 8 wherein the dataset of clinical notes comprises:
a daily narrative training set that includes sentences from hospital course sections of clinical notes, and wherein the clinical notes contain a date that falls within the time interval; and
a discharge plan training set that includes EMRs that include discharge related words.
10. The method of claim 1 wherein analyzing the factuality of each of the candidate summaries is performed by a constrained beam search algorithm.
11. The method of claim 10 wherein the constrained beam search algorithm comprises:
constructing a banned word list that specifies medical terms that cannot be included in the automatically generated summary;
eliminating any summary that includes a term included in the banned word list, and
using a transformer model to compute and select the highest probability remaining candidate summary.
12. A server computer, comprising:
a processor,
a communication interface in communication with the processor;
a data storage for storing a database of clinical notes, wherein a clinical note is written by a clinician and provides details about a patient's hospital stay, the clinical notes comprising physician notes provided by physicians; and
a memory in communication with the processor for storing instructions, which when executed by the processor, cause the server:
to train a transformer model on a classification training dataset based on a daily course section of the clinical notes;
to select from the database physician notes for a patient's hospital stay during a designated time interval;
to identify from the physician notes a set of significant physician notes for the time interval;
to generate a candidate set of summaries for each of the identified significant physician notes, wherein a summary includes one or more sentences that are abstractively generated using the transformer model;
for each significant physician note, to perform a factuality analysis on the candidate summaries, wherein the factuality analysis (1) eliminates any summary that includes a non-factual medical word or phrase, and (2) selects the highest probability remaining summary; and
to generate a daily section for inclusion in a hospital course section of a discharge note, the generated daily section including the selected factual summary for each of the significant physician notes.
13. The server computer of claim 12 wherein the instructions, when executed by the processor, further cause the server:
to retrieve electronic medical records for a patient;
to translate the received medical records into a standard format; and
to store the translated medical records in the database.
14. The server computer of claim 12 wherein the time interval is a day.
15. The server computer of claim 12 wherein the instructions, when executed by the processor, further cause the server:
to perform a regular expression (regex) processing step on the selected physician notes to extract the most salient and useful content for further processing.
16. The server computer of claim 12 wherein the physician note is selected from the group consisting of Progress Notes, Procedure Notes, Op Notes, and Consult Notes.
17. The server computer of claim 12 wherein identifying a set of significant physician notes is performed using a machine learning model.
18. The server computer of claim 17 wherein the machine learning model is a transformer model and the transformer model is fine-tuned using a dataset comprising clinical notes.
19. The server computer of claim 12 wherein the transformer model is fine-tuned using a dataset comprising clinical notes.
20. The server computer of claim 19 wherein the dataset of clinical notes comprises:
a daily narrative training set that includes sentences from hospital course sections of clinical notes, and wherein the clinical notes contain a date that falls within the time interval; and
a discharge plan training set that includes EMRs that include discharge related words.
21. The server computer of claim 12 wherein analyzing the factuality of each of the candidate summaries is performed by a constrained beam search algorithm.
22. The server computer of claim 21 wherein the constrained beam search algorithm comprises:
constructing a banned word list that specifies medical terms that cannot be included in the automatically generated summary;
eliminating any summary that includes a term included in the banned word list, and
using a transformer model to compute and selecting the highest probability remaining candidate summary.