Radiology department uses AI to identify lung cancer screening patients in EHR
A radiology department is detailing how it uses natural language processing to help root out screening-eligible patients from the health system’s electronic health record.
Lung cancer is the leading cause of cancer death in the U.S., with low-dose CT screening shown to reduce mortality. However, uptake remains low at about 18% nationally, experts detail Oct. 12 in the Journal of the American College of Radiology.
One key barrier is incomplete or inaccurate structured smoking history documentation in the EHR, limiting automated LCS determinations. To address this, the University of California, Irvine, has developed a natural language processing tool, which extracts smoking info and identifies screening-eligible patients from free-text clinical notes.
Experts say the intervention is showing promise, especially when compared to the previous manual method of mining this information.
“In summary, [natural language processing]-derived smoking status and [lung cancer screening] eligibility demonstrated high concordance with manual review and identified more than three times as many screening-eligible patients as structured EHR data,” corresponding author Gelareh Sadigh, MD, with the Department of Radiological Sciences at UCI, and colleagues concluded. “Our findings support NLP as a practical and scalable approach to improving LCS eligibility ascertainment from existing clinical documentation without changing physician workflow.”
The retrospective, observational study sought eligible patients who were 50 or older, had an outpatient primary care encounter, and a documented EHR note sometime between 2018 and 2023. For each patient, researchers extracted free-text notes from all well-visit encounters, together with structured smoking history data. Using 50 manually reviewed notes, they developed a new rule-based NLP pipeline to extract smoking variables from free-text EHR notes. A separate, non-overlapping sample of 2,000 notes was used to further refine the AI model.
A total of nearly 28,000 patients with almost 41,000 EHR notes were included in the study. NLP compared against 2,000 initially human-annotated notes showed nearly 84% agreement for lung cancer screening eligibility. Inter-rater agreement between the two human reviewers for 372 patients with discordant NLP-manual LCS eligibility yielded 96% agreement. NLP against the 200 annotated notes that were used for validation showed 84% agreement.
When testing AI for assessing eligibility from 2,200 notes, it demonstrated strong overall performance. The LCS ineligible group yielded the highest accuracy, followed by the LCS-unknown cohort. In adjusted multivariable analysis, overall NLP agreement with manual review for screening eligibility was lower with every one-year increase in age and when smoking status was current versus never, among other factors.
“Our NLP performance is broadly consistent with prior research, though direct comparison is limited by heterogeneity in tasks, validation methods, and reference standards,” the authors noted. “A strength of this study is the larger manually reviewed reference sample. Higher NLP–manual agreement among patients with unknown or no smoking history may reflect the need to extract fewer data elements to determine LCS eligibility in this population,” they added.
Read more, including potential study limitations, in JACR.
Marty Stempniak has covered healthcare since 2012, with his byline appearing in the American Hospital Association's member magazine, Modern Healthcare and McKnight's. Prior to that, he wrote about village government and local business for his hometown newspaper in Oak Park, Illinois. He won a Peter Lisagor and Gold EXCEL awards in 2017 for his coverage of the opioid epidemic.