The Scan Found a Lung Nodule. Then the Hospital Had to Find the Patient.

An emergency-department CT scan is ordered to answer an urgent question. Is this chest pain a pulmonary embolism? Did the fall cause internal injury? Is the shortness of breath pneumonia?

Presented By Our Partners

The scan may also reveal something nobody was looking for: a small nodule in the lung.

Most pulmonary nodules are not cancer. Some need no surveillance at all. Others, depending on their size, appearance, and the patient’s risk, call for another scan, a specialist evaluation, or more testing. The finding can be medically important without being the reason the patient came to the emergency department.

That is precisely why it is so easy to lose.

The radiologist records it. The emergency clinician is managing the immediate problem. The patient goes home with several pages of instructions and perhaps no established primary-care doctor. The nodule exists in the record, but its next step belongs to no one in particular.

A study published July 25 in Respiratory Medicine tested a modest use of artificial intelligence against this very human failure. The technology did not diagnose lung cancer. It did not decide who needed surgery. It read radiology reports for possible incidental pulmonary nodules and helped trigger an outreach workflow so a care team could notify patients and facilitate follow-up.

The result was not a miracle. It was something more useful: a measurable improvement in whether the patient actually heard about the finding and returned to have it addressed.

The distance between a finding and care

The researchers studied an academic-affiliated community hospital with a multidisciplinary lung-cancer program. They compared two groups of emergency-department patients who had chest CT scans.

In the first period, from January through March 2023, the hospital used its usual process. In the later period, from June through August 2025, a natural-language-processing system screened radiology reports for possible nodules, after which patients were contacted to help arrange follow-up.

The comparison involved 228 patients before the new workflow and 252 afterward.

Patient notification rose from 75 percent to 88.4 percent. Nodule-specific follow-up rose from 53.5 percent to 67.9 percent.

Those numbers deserve to be read in both directions.

The workflow moved roughly 14 additional patients per hundred into nodule-specific follow-up. For a relatively light-touch intervention at a single hospital, that is meaningful. The finding did not merely become more visible to a machine; it became more likely to reach a person and produce another care event.

But 67.9 percent is not a closed loop. Nearly one patient in three still did not complete the documented follow-up. The researchers identified an especially ordinary barrier: after the emergency visit, some patients could not be reached by phone.

No model can schedule a patient whose number no longer works. No alert can supply transportation, paid time off, insurance approval, trust, or a primary-care relationship. AI may locate the care gap, but closing it still depends on people, contact information, capacity, and a system willing to own the next step.

Why emergency departments lose incidental findings

Emergency medicine is built around the problem that might harm the patient today. Incidental findings operate on a different clock.

Featured Partner

Invest in the Infrastructure Behind Modern Medicine

As healthcare expands beyond hospital walls, the buildings and campuses supporting that shift are generating compelling returns for investors who move early. The Healthcare Real Estate Fund offers qualified investors direct access to a curated portfolio of medical office, outpatient, and specialty care facilities.

Learn More →

A lung nodule may require a repeat CT months later. The recommendation may depend on details such as whether the nodule is solid or subsolid, its diameter, whether it has grown, and the patient’s smoking history or other cancer risks. Fleischner Society recommendations are deliberately risk-stratified because indiscriminate follow-up can expose patients to unnecessary radiation, anxiety, procedures, and cost.

The right response is therefore not “every nodule is cancer.” It is “every finding that warrants follow-up needs a reliable owner.”

Prior research has shown how fragile that ownership can be. Published reports have put follow-up rates for incidental nodules across settings in a wide range, roughly 30 to 77 percent. The emergency department is an especially difficult handoff because its clinicians may never see the patient again and may not have a durable relationship with the physician expected to receive the recommendation.

This is a recurring pattern in healthcare. The technical act is completed; the clinical journey is not.

A radiology report can be accurate while the patient remains uninformed. A recommendation can be appropriate while no appointment is scheduled. An order can exist while the scan never occurs. Each step can appear successful inside its own software system even as the overall pathway fails.

The new study matters because it measured beyond the first step. It asked not only whether a report contained a finding, but whether the patient was notified and whether nodule-specific follow-up occurred.

That is a better standard for healthcare AI.

The AI was the tripwire, not the care

It is tempting to describe this as an AI success story. That is only partly right.

The software acted as a tripwire. It searched language in radiology reports for cases that might otherwise depend on a busy clinician, a manually maintained list, or a patient deciphering discharge paperwork. The benefit came from attaching that detection to outreach.

This distinction matters. An algorithm that flags more cases but sends them into an unattended queue has improved detection without improving care. An algorithm that generates more false alarms can create work and anxiety without benefit. A high-performing model deployed without a follow-up team may simply produce a more organized inventory of unfinished business.

The intervention in this study was a system, not just a model: report review, case identification, human contact, and facilitation of the next step.

That makes the remaining misses informative. The workflow improved notification to 88.4 percent, but it did not reach everyone. It improved follow-up to 67.9 percent, but a sizable gap remained between being told and being cared for.

The drop-off points point to the next design questions. Should outreach use text messages, portal messages, mail, and primary-care notification rather than phone alone? Who tries again when the first call fails? Can navigators see whether a recommended CT was scheduled, completed, and interpreted? What happens when the patient is uninsured, lives far from imaging, or cannot afford the next test?

These are not edge cases around the technology. They are the pathway the technology is supposed to improve.

What the study did not prove

The study was retrospective and compared patients from two different time periods. It was conducted at one hospital that already had a multidisciplinary lung-cancer program. Changes in staffing, practice patterns, patient mix, or other parts of care between 2023 and 2025 could have influenced the results.

The researchers also found no significant difference in lung-cancer stage at diagnosis between the two cohorts.

That does not make the workflow unimportant. It defines the claim honestly.

The study showed better notification and better follow-up. It did not show that AI reduced lung-cancer deaths, improved survival, or shifted cancers into earlier stages. A larger, longer, preferably prospective evaluation would be needed to answer those questions.

Stage still explains why follow-up matters. CDC data indicate that nearly half of U.S. lung cancers diagnosed from 2019 through 2023 were already distant at diagnosis, while just over one-quarter were localized. But those population statistics cannot be used to infer that this particular workflow changed anyone’s prognosis.

The strongest healthcare evidence keeps these layers separate:

• Did the tool find the right case?
• Did the patient learn about it?
• Was the recommended next step completed?
• Was a clinically important diagnosis made?
• Did treatment begin sooner?
• Did health outcomes improve?

This study reached the third question. That is farther than many healthcare AI claims go, and not as far as the headline “AI improves cancer outcomes” would imply.

The real product is accountability

Healthcare is filling with systems that can recognize language, images, risk patterns, and missing data. The scarce capability is increasingly not recognition. It is accountability after recognition.

Incidental pulmonary nodules are a clean example because the finding begins as a fragment of information inside a report. To become care, it must cross several boundaries: radiology to emergency medicine, hospital to home, finding to explanation, recommendation to appointment, and appointment to completed imaging or specialist review.

AI can help watch those boundaries. It can search reports consistently, prioritize cases, and reduce dependence on memory. But the value appears only when a health system treats the alert as the beginning of work rather than the end.

For patients, that may be the most important difference between healthcare AI that performs and healthcare AI that matters.

The machine noticed the nodule.

The care team still had to find the patient.

Sources

• Li W. et al. “Improving Follow-Up of Incidental Pulmonary Nodules in the Emergency Department Using an Artificial Intelligence-Supported Workflow.” Respiratory Medicine. Published online July 25, 2026. PubMed PMID: 42501874.
• MacMahon H. et al. “Guidelines for Management of Incidental Pulmonary Nodules Detected on CT Images: From the Fleischner Society 2017.” Radiology. 2017;284(1):228–243. doi:10.1148/radiol.2017161659.
• American College of Radiology. “Incidentally Detected Indeterminate Pulmonary Nodule.” ACR Appropriateness Criteria.
• American College of Radiology Learning Network. Recommendations Follow-Up Improvement Collaborative reporting on incidental pulmonary nodule recommendation adherence and imaging completion.
• Centers for Disease Control and Prevention. “U.S. Cancer Statistics: Lung Cancer Stat Bite.” Stage distribution for diagnoses from 2019–2023.

Free Daily Briefing

The Latest Longevity Science.
Delivered Every Morning.

Join researchers, physicians, and health professionals getting daily breakthroughs in AI-driven medicine, epigenetics, and longevity research.

Support the research that powers this editorial

Subscribe Free

No spam. Unsubscribe anytime. We respect your inbox.