Generating reasons for imaging studies
A framework for generating reasons for imaging studies. An extractor, including a reinforcement learning agent, is trained to select one or more relevant sentences from the training histories of present illness. An abstractor is further pre-trained to generate one or more reasons for study from the one or more relevant sentences. An entity linking system is pre-trained using medical text corpora to map one or more mentions in the one or more reasons for study to one or more standardized medical entities for predicting one or more diagnoses. The reinforcement learning agent may then be re-trained using one or more rewards generated by the entity linking system. One or more reasons for study may be generated from a current history of present illness using the trained extractor, abstractor and entity linking system.
1. A medical text summarization system, comprising:
a non-transitory memory device for storing computer readable program code; and
a processor device in communication with the memory device, the processor device being operative with the computer readable program code to perform steps including
receiving training histories of present illness, reference medical documents and medical text corpora,
training, based on the training histories of present illness and the reference medical documents, an extractor that selects one or more relevant sentences from the training histories of present illness, wherein the extractor comprises a reinforcement learning agent,
pre-training, based on the training histories of present illness and the reference medical documents, an abstractor that generates one or more first reasons for study from the one or more relevant sentences selected by the extractor, wherein the one or more first reasons for study comprise one or more paraphrases of the one or more relevant sentences,
pre-training an entity linking system using the medical text corpora to map one or more mentions in the one or more first reasons for study to one or more standardized entities for predicting one or more diagnoses,
re-training, based on the training histories of present illness and the reference medical documents, the reinforcement learning agent using one or more rewards generated by the entity linking system by evaluating quality of the one or more first reasons for study, and
generating one or more second reasons for study from a current history of present illness using the trained extractor, the pre-trained abstractor and the pre-trained entity linking system.
2. The medical text summarization system of claim 1 wherein the processor device is operative with the computer readable program code to generate source-target pairs from the training histories of present illness and the reference medical documents for training the extractor and pre-training the abstractor.
3. The medical text summarization system of claim 1 wherein the processor device is operative with the computer readable program code to pre-train a transformer-based contextual language model using the medical text corpora.
4. The medical text summarization system of claim 3 wherein the transformer-based contextual language model is used in the extractor, the abstractor, or a combination thereof.
5. The medical text summarization system of claim 1 wherein the processor device is operative with the computer readable program code to train the extractor using one or more heuristics.
6. The medical text summarization system of claim 1 wherein the processor device is operative with the computer readable program code to pre-train the abstractor using one or more heuristics.
7. The medical text summarization system of claim 2 wherein the processor device is operative with the computer readable program code to re-train the reinforcement learning agent by
selecting, by the extractor, one or more sentences from at least one of the source-target pairs,
generating, by the abstractor, one or more reasons for study from the selected one or more sentences,
predicting, by the entity linking system, one or more diagnoses from the generated one or more reasons for study, and
determining, by the entity linking system, the one or more rewards for re-training the reinforcement learning agent using an actor critic technique.
8. The medical text summarization system of claim 7 wherein the processor device is operative with the computer readable program code to determine the one or more rewards by measuring an overlap between the one or more diagnoses predicted by the entity linking system and reference diagnoses extracted from the reference medical documents.
9. The medical text summarization system of claim 7 wherein the processor device is operative with the computer readable program code to determine the one or more rewards by measuring an overlap between the one or more reasons for study generated by the abstractor and one or more actual reasons for study extracted from the reference medical documents.
10. A method of medical text summarization, comprising:
receiving an extractor, an abstractor, an entity linking system and a current history of present illness;
selecting, using the extractor, one or more sentences from the current history of present illness, wherein the extractor includes a reinforcement learning agent that receives a feedback of one or more rewards from the entity linking system, wherein the one or more rewards are generated by evaluating quality of one or more first reasons for study generated by the abstractor;
generating, using the abstractor, one or more second reasons for study from the one or more sentences selected by the extractor, wherein the one or more second reasons for study comprise one or more paraphrases of the one or more sentences; and
predicting, using the entity linking system, one or more diagnoses from the one or more second reasons for study generated by the abstractor.
11. The method of claim 10 further comprises extracting the current history of present illness from a current medical document associated with a specific patient.
12. The method of claim 10 wherein predicting the one or more diagnoses comprises mapping one or more mentions in the one or more generated second reasons for study to one or more standardized entities.
13. The method of claim 10 further comprises pre-training a transformer-based contextual language model using medical text corpora for use in the trained extractor, the pre-trained abstractor, or a combination thereof.
14. The method of claim 10 further comprises training the extractor using one or more heuristics.
15. The method of claim 10 further comprises pre-training the abstractor using one or more heuristics.
16. The method of claim 10 further comprises re-training the reinforcement learning agent using one or more rewards generated by the entity linking system.
17. The method of claim 16 wherein re-training the reinforcement learning agent comprises:
selecting, by the extractor, one or more sentences from at least one source-target pair;
generating, by the abstractor, one or more reasons for study from the selected one or more sentences;
predicting, by the entity linking system, one or more diagnoses from the generated one or more reasons for study; and
determining, by the entity linking system, the one or more rewards for re-training the reinforcement learning agent using an actor critic technique.
18. The method of claim 17 further comprises determining the one or more rewards by measuring an overlap between the one or more diagnoses predicted by the entity linking system and reference diagnoses.
19. The method of claim 17 further comprises determining the one or more rewards by measuring an overlap between the one or more reasons for study generated by the abstractor and one or more actual reasons for study.
20. One or more non-transitory computer-readable media embodying instructions executable by machine to perform operations, comprising:
receiving an extractor, an abstractor, an entity linking system and a current history of present illness;
selecting, using the extractor, one or more sentences from the current history of present illness, wherein the extractor includes a reinforcement learning agent that receives a feedback of one or more rewards from the entity linking system, wherein the one or more rewards are generated by evaluating quality of one or more first reasons for study generated by the abstractor;
generating, using the abstractor, one or more second reasons for study from the one or more sentences selected by the extractor, wherein the one or more second reasons for study comprise one or more paraphrases of the one or more sentences; and
predicting, using the entity linking system, one or more diagnoses from the one or more second reasons for study generated by the abstractor.