AI has applications across the radiology spectrum, including the use of natural language processing algorithms to help collect and process data for clinical research. The power of the algorithm may be limited by the way electronic medical records are organized however, said Marta Flory, MD, clinical assistant professor at Stanford University, in her Monday presentation.
"Machine learning and AI improves quality in radiology by advancing understanding of disease processes, enabling analysis of vast archives of data, helping us generate cohorts of the population of interest and extracting straightforward data," Dr. Flory said. "But there are challenges."
Flory
Using Natural Data Processing To Determine Which Cysts Turned Into Cancer
During the session, Dr. Flory discussed a research project she led to use natural language processing for evaluating pancreatic cysts, which can present normally, but may develop into cancer. Often, patients are observed with additional imaging for ten years so there is a lot of data in the electronic medical record about cysts that were found and observed over a number of years to see which did or did not turn into cancer.
“Our clinical question was, let's use all this data we have stored electronically over 30 years and try to mine it and shed some insight into those pancreatic cysts so we know what features may predict future malignancy,” Dr. Flory said.
The first step was to write a natural language processing algorithm to create a cohort of cases. The team had to choose words they thought radiologists would have chosen to describe the cysts or lesions and have the software to look at all reports for those words, then had a radiologist check that the algorithm was working.
This step turned out to be more difficult and time consuming than Dr. Flory expected to capture all the words radiologists might have chosen in their reports as well as how the reports were organized.
“The algorithm performed really well, but our real challenge was in how our data is actually stored,” she said.
The way things are coded and stored in the electronic medical record typically aligns with billing, not pathology, so it was challenging to know how to find the right things without knowing all the right billing codes that might lead to a report that included what the team was looking to include.
Tips For Researchers Looking at Data
She advised that research teams allow for more resources for cohort generation.
“As a clinical radiologist you are not that intimately familiar with the coding, and it is a very complicated system. It takes time and money,” Dr. Flory said. “The codes can be different across different hospital systems. In a country with a national healthcare system, all the different hospitals are using the same codes, but that is not the case in the United States. There is a lot of variability that can make it difficult to translate this work from one institution to the next.”
Once the team had their cohort, they wrote another algorithm to look at the size of the largest cyst in the report and evaluate it over time.
“We found that the current recommendation for following pancreatic cysts is a little more aggressive than it needs to be,” Dr. Flory said.
But even that question was a bit of a challenge as the algorithm was looking at the text of the report not the images so there was not a clear link labeling which lesion might have grown over time if there were multiple.
Templated reports and standardized dictation can go a long way toward help solve some of these research problems, but they need to be widely adopted to make a difference, Dr. Flory said.
“Natural language processing offers promise to be able to use electronic medical records to do research, there is a lot of data we can use, but researchers should also know about the limitations as well,” she concluded.
Access the presentation, "Creating Links Between Radiology and Health Informatics Systems to Improve Quality," (M3-CIN15) on demand at Meeting.RSNA.org.
