Virtual Transcriptionist Dictation Assistant

Purpose	Provide a more intelligent speech recognition through AI-assisted virtual transcriptionist/dictation assistant that can address common errors that occur within radiology reports, including age, sex/gender, laterality, speech-recognition errors, and affirmative/negative correlation between findings and impression
Tag(s)	Non-Interpretative
Panel	Reading Room
Define-AI ID	19120002
Originator	Woojin Kim
Lead	Woojin Kim
Panel Chair	Ben Wandtke
Non-Interpretative Panel Chairs:	Alexander J Towbin, Adam Prater
Panel Reviewers	Reading Room Subpanel
License	Creative Commons 4.0
Status	Public Commenting

                               

Value Proposition

Radiologists have been using speech recognition (SR) to generate their radiology reports for decades, one of the early adopters of the SR technology within healthcare. While there have been advancements in the SR and reporting technology over the years, AI has the potential to provide a more intelligent virtual transcriptionist/dictation assistant function to the radiologists to reduce errors within their reports.

Narrative(s)

A 46-year-old patient calls the radiology department angrily to discuss her radiology report that contains multiple errors as well as a confusing impression. The report described her age as being 64 years old with a description of a prostate (instead of a uterus). The technique section says, “contrast menstruation.” While the findings section of the report says “no pulmonary nodule is seen,” the impression says, “1. Pulmonary nodule.”

Workflow Description

The AI algorithms can monitor the radiology report text in real-time to look for various errors, similar to how Grammarly (https://app.grammarly.com/) works to improve the grammar of the user in real-time (similar UI also recommended for consideration). The metadata, including the patient demographics, should be matched against the radiologist’s report text to detect whenever there is an error in the patient’s age, sex/gender (which includes not only male vs. female descriptions but also anatomy - e.g., prostate in a female as in the narrative example; in addition, the algorithm needs to be able to distinguish between the patient and someone described in addition to the patient - e.g., the history section may say, “mother describes cough and fever,” but the exam is that for her infant son; related, the algorithm also needs to be able to detect and keep track when more than one person has been imaged, such as the sex of the patient and the sex of the fetus/fetuses - e.g., MRI of a woman with a male fetus and a female fetus), and laterality.

In the above example, the radiologist dictated, “contrast administered,” but the SR transcribed it as “contrast menstruation.” Such SR errors are well-known to occur. AI algorithms can provide additional intelligence to know whenever there is a term that is out of place in context. This can supplement the basic spelling and grammar checks. Additional intelligence can be used to know whenever a particular body part or finding was mentioned that would not be seen with a given exam type (e.g., description of a thumb in a foot x-ray report).

Finally, the descriptions within the Findings section should be matched with those in the Impression section to ensure concordance. In the last example in the narrative, the radiologist either failed to say “no” or the SR failed to pick up “no,” resulting in what appears to be conflicting Findings vs. Impression.

• Need a robust heterogeneous training dataset as prosaic and templated styles to vary widely across institutions (not to mention countries and languages)

Inputs

 

Radiologist Report

Radiologist Voice Recording

 

 

Primary Outputs

Wrong Text Detection in Report Text

Definition	Detect the wrong word(s) within the report text.
Data Type	Text
Value Set	N/A
Units	N/A

Wrong Text Detection in Voice Recording

Definition	Detect the wrong word(s) within voice recording.
Data Type	Text
Value Set	N/A
Units	N/A

Correct Word Suggestion

Definition	Suggestion of the correct word(s).
Data Type	Text
Value Set	N/A
Units	N/A

Secondary Output 

Reason Word(s) is wrong

Definition	Provide user reason on why detected word(s) is wrong.
Data Type	Categorical
Value Set	· No issues · Misused term · Not aligned with the exam · Not aligned with the patient · Grammatical error · Conflict within the report · Conflict between the report and EHR · Missing information
Units	N/A

Not all errors come from the radiologists. Sometimes the metadata contains errors. Therefore, more advanced AI algorithms should incorporate other information. Using computer vision, for example, the laterality described in the report can be compared also with the actual images. A similar AI technique can be used to determine errors within views or technique descriptions (e.g., 3 views instead of 2, enhanced vs. unenhanced, etc.).