What UC Irvine built
A team at UC Irvine's Department of Radiological Sciences, led by Gelareh Sadigh, MD, published a study in the Journal of the American College of Radiology (October 2026) describing a natural language processing pipeline built to solve a specific, unglamorous problem: finding out who in a hospital's patient population actually meets lung cancer screening eligibility criteria, as reported by Radiology Business.
USPSTF eligibility hinges on smoking history — current or former smoker, age range, pack-years, time since quitting. That information exists in nearly every chart. The problem is where it lives. The team analyzed close to 28,000 patients and roughly 41,000 EHR notes, then validated the NLP output against about 2,000 manually reviewed notes.
The result: roughly 84% agreement between the NLP classification and manual chart review. When two human reviewers independently re-checked the cases where the tool and the first reviewer disagreed, they agreed with each other about 96% of the time — a sign the discordance is mostly a documentation gap the model is closing, not model error.
The number that matters: 3x
The headline figure isn't the accuracy score — it's the yield. Using free-text notes in addition to structured EHR fields, the NLP approach identified more than three times as many lung-cancer-screening-eligible patients than structured data alone would have surfaced. The researchers concluded that NLP is "a practical and scalable approach to improving LCS eligibility ascertainment from existing clinical documentation without changing physician workflow."
Read that carefully. The patients were always in the system. A clinician had already documented smoking history in a progress note — just not in a field a rules-based query can read. No new data entry, no new clinic visit, no new checkbox was needed to find them. The information was simply written as a sentence instead of a structured value.
Unstructured data hits the screening pathway twice
It's tempting to read this as a story about patient intake and stop there. It's more useful read as a pattern: the same structural problem — clinically important information trapped in free text instead of a queryable field — shows up again later in the same pathway, after the scan has already happened. Unstructured text is costing radiology twice: once in the chart, once in the report.
| Pathway stage | Where the data goes unstructured | What closes the gap |
|---|---|---|
| Finding eligible patients (intake) | Smoking pack-years written in a progress-note sentence, not a discrete field | NLP mining of free-text notes, as in the UC Irvine study |
| Reading and reporting the CT | Nodule size, location, and Lung-RADS-relevant detail buried in narrative dictation | A structured, comprehensive report drafted with consistent fields from the start |
| Follow-up and registry tracking | Recommended next steps described in prose a scheduling system can't trigger on | Discrete, machine-readable recommendation fields carried through the report |
The intake-side fix and the reporting-side fix look different — one is a text-mining layer over existing notes, the other is how the report is produced in the first place — but they are solving the same category of problem: clinical language that downstream systems, schedulers, and registries can't act on without a human re-reading and re-typing it.
Why identification alone won't close the screening gap
Lung cancer screening uptake remains low: only about 18% of eligible US adults were up to date as of 2022, per an American Cancer Society analysis published in JAMA Internal Medicine. That shortfall has multiple causes, and this study addresses exactly one of them — making sure a health system actually knows who in its population qualifies.
It doesn't address CT scanner throughput, radiologist reading capacity, or what happens to the report once the scan is read — gaps covered in a related look at how screening-criteria changes affect CT volume and in a separate analysis of why screening programs struggle to keep patients on a completed care pathway. Finding more eligible patients without the downstream capacity to scan, read, and track them just moves the bottleneck one stage later.
Where xAID fits
The UC Irvine study is evidence for a broader point: free-text clinical documentation is a liability at every stage of a screening pathway unless something turns it into structured, queryable output. On the reporting side, that's the problem structured AI CT reporting is built to solve — a comprehensive, consistently formatted draft instead of free-text dictation, reviewed by xAID's in-house radiologist on every preliminary and delivered ready-to-sign. Solving the intake-side identification problem only pays off if the reporting side produces output a scheduling system, a registry, or a follow-up tracker can actually read.
Frequently asked questions
How is AI being used to find lung cancer screening eligibility in the EHR?
Researchers at UC Irvine's Department of Radiological Sciences built a natural language processing (NLP) pipeline that reads free-text clinical notes to extract smoking history — current/former status, pack-years, quit date — and flags patients who meet USPSTF lung cancer screening eligibility criteria. The study, led by Gelareh Sadigh, MD, and published in the Journal of the American College of Radiology (October 2026), analyzed nearly 28,000 patients and about 41,000 EHR notes.
How many more eligible patients did the NLP tool find compared to structured EHR data alone?
The NLP approach identified more than three times as many lung-cancer-screening-eligible patients as structured EHR fields alone, with about 84% agreement against roughly 2,000 manually reviewed notes. Two human reviewers independently re-checked the discordant cases and agreed with each other about 96% of the time, which points to a documentation gap rather than a measurement error.
Why do structured EHR fields miss lung cancer screening eligibility?
Smoking history is frequently recorded as a sentence in a progress note — for example, quit smoking after decades of a pack a day — rather than as a discrete pack-years field a screening algorithm can query. If a clinician documents it in prose instead of a structured intake field, rules-based eligibility screens miss it entirely, even though the information already exists in the chart.
Does finding more eligible patients close the lung cancer screening gap?
It addresses one of at least two separate bottlenecks. An American Cancer Society analysis found only about 18% of eligible US adults were up to date with lung cancer screening in 2022, and separate research shows expanding or simplifying eligibility criteria alone does not close that gap. Identifying who qualifies is necessary but not sufficient — the scans still have to be scheduled, read, and reported consistently enough to support tracking and follow-up downstream.
Source: G. Sadigh, MD, et al., UC Irvine Department of Radiological Sciences, Journal of the American College of Radiology (October 2026), as reported by Radiology Business. National screening-uptake figure (18.1% in 2022) from the American Cancer Society analysis in JAMA Internal Medicine. Figures are rounded as reported.