Such overdiagnosis was illustrated in a recent study from JAMA Internal Medicine of women in Taiwan who were screened for lung cancer using low-dose CT (LDCT).
The study, and the accompanying editorial, pointed out the potential for large databases of routine clinical data to track long-term outcomes, and potentially identify patient subgroups that could benefit from early diagnosis using digital technologies.
The Taiwanese findings echo a similar trend identified in a 2018 post hoc analysis of the Danish Lung Cancer Screening Trial, which estimated that 67.2% of cancers found during that CT screening program for current or former smokers were overdiagnosed. The authors recommended that researchers report rates of overdiagnosis in future screening studies.
The authors of the Taiwan study noted that LDCT is low cost and is frequently offered to individuals who are not considered at high risk of lung cancer, and advertisements in Taiwan often target women, who rarely smoke. The researchers examined data from the Taiwan National Cancer Registry. They looked for evidence of an increased incidence of early-stage detection and reduced incidence of late-stage diagnosis. They found that, from 2004 to 2018, there was an increase of lung cancer incidence from 2.3 to 14.4 per 100,000 (difference, 12.1; 95% confidence interval, 11.3-12.8), but no significant difference in the incidence of late-stage disease (from 18.7 to 19.3 per 100,000; difference, 0.6; 95% CI, –0.5 to 1.7).
“This combination of findings, an additional 12.1 early-stage cancers per 100,000 population and no reduction in late-stage cancers, is strongly suggestive of overdiagnosis,” the authors wrote.
It can be difficult to convince people of the potential harms of overdiagnosis, especially when patients have a nodule removed and remain healthy years later. “It’s very counterintuitive, but it’s a reality, and I think this paper paints the reality very, very clearly,” said Daniel Capurro, MD, PhD, deputy director of the Centre for the Digital Transformation of Health at University of Melbourne, and an author of the editorial.
The issue is that some lung cancers progress so slowly that they may never cause a problem clinically, and their removal can lead to unnecessary cost and risk. And it’s not just cancer. “There are a bunch of other conditions that are defined by specific criteria, but we don’t add the prognosis to that definition. At the individual patient level, we don’t know the prognosis,” Dr. Capurro said.
Dr. Capurro discussed the increasing use of digital technologies like smartphone apps. Machine learning can potentially use such data to diagnose conditions like sleep or mood disorders before they become clinically significant, allowing earlier intervention, but they could also lead to overdiagnosis. Dr. Capurro proposed using longitudinal databases to track patient outcomes, which could be applied to digital screening technologies.
“You might be able to find unknown patterns that help discriminate between these pathological definitions. You should be able to train (digital screens) with the pathological definition plus the disease trajectory as a way to improve that label,” he said.
The study was funded by the Taiwan Ministry of Health and Welfare Clinical Trial Center.