Wednesday, December 3 | 8:20 a.m.-8:30 a.m. | W1-SSNPM03-3 | S402
In centers that perform about 30 PET/CT scans daily, ChatGPT could save clinicians at least an hour if put to use assigning ICD-10 codes to exam reports, according to a group at Charité University Hospital in Berlin.
Resident Tristan Ruhwedel, MD, will present his team’s research on whether ChatGPT can potentially help alleviate administrative burdens and reduce burnout risk.
The researchers created 100 different fictitious PET/CT reports based on standardized institutional text modules for PET/CT, including contrast-enhanced diagnostic CT. Each report was assigned a single, most relevant ICD-10 code by two experienced nuclear medicine physicians. One physician used ChatGPT, while the other was limited to traditional search engines. The analysis involved identifying differences in error distribution across ICD-10 hierarchy levels using four-field tables and the McNemar test.
At the first character of the ICD-10 code, the ChatGPT-assisted coder made two errors compared with three by the manual coder. For the second and third characters, ChatGPT-based coding resulted in three errors versus five, and at the fourth and fifth characters, ChatGPT significantly outperformed manual coding, with six errors versus 17 (p = 0.035).
In addition, median coding time was 135 seconds (range, 41 seconds to 433 seconds) without ChatGPT and just 8 seconds (range, 6 seconds to 9 seconds) with it.
“Given the complexity of PET/CT reports, similar approaches might be feasible for other imaging modalities like diagnostic CT or MRI,” Ruhwedel and colleagues concluded.
This study was awarded a Trainee Research Prize in the Resident category by RSNA.