Ron Mazumder, Mara Aspinall, and Nick Naclerio

A recent meta-analysis published in JAMA Internal Medicine (August 2023) found no evidence that the most common types of cancer screening are providing statistically significant benefits in overall survival, with the exception of colorectal cancer screening with sigmoidoscopy.  How can this be?  We have all been taught that mammograms, colonoscopies, PSA screening, and low-dose CT save lives from breast, colon, prostate, and lung cancer, respectively. Should we stop getting screened? Not just yet.

Screening tests help identify "hidden” cases of disease in a broad population.  In the correct setting, they can identify those with occult disease at an early stage when treatments may be more effective. Screening tests are not intended to make a final determination of disease status. Further testing and clinical follow-up are almost always needed for a definitive diagnosis. Done right, screening is all about improving probabilities of early detection – by separating a smaller pool of individuals with a higher-than-average likelihood that they will be positive. If not done right, the benefits of screening tests can be offset by anxiety they create while waiting on a final diagnosis, excessive use of invasive diagnostic follow-ups, and complications from potentially unnecessary treatments.

Challenges: Biases, clinical endpoints and clinical truth. 

Proving screening’s effectiveness is complicated. Just because a screening test correctly identifies some otherwise undiagnosed cancers does not mean that it will reduce cancer deaths or that will increase overall survival at the population level.  There are multiple sources of bias that must be taken into account to appropriately evaluate the impact of screening.

• Lead time bias: Screening may find cancer earlier, but if interventions are ineffective, overall survival is not extended by the earlier identification.
• Length bias: Screening can pick up slower-growing, less aggressive cancers, which patients may die with but not from.
• Selection bias: The people being screened (and then treated) may not be typical of the overall population? Those at highest risk of disease may not be the most likely to be screened due to work and other socio-economic factors.
• Unreliable data bias: Disease (or cancer-)-specific mortality depends on consistent, accurate and unambiguous cause of death data. Most data comes from death certificates which can be highly variable making it challenging to assign an accurate cause of death, especially in the vast majority of patients with comorbidities.
• Under-powered bias: Trials to demonstrate reductions in cancer specific or all-cause mortality require very large patient cohorts monitored over very long time periods. The less common the cancer and the longer the interval typical from detection to death, the more difficult this challenge this becomes.

Combining multiple studies in a meta-analysis is difficult:  Including apples and oranges. Meta-studies need to exclude unreliable studies and adjust the data of included studies for comparability. This is easier said than done and brings in its own biases. On the one hand, relevant meaningful data gets neglected, and on the other, necessary adjustments introduce novel inaccuracies. Meta-studies are useful surveys of current knowledge, but statistics on effectiveness can be highly suspect (unless the data is so clear that a meta-study is not required).

The recent meta-analysis published in JAMA Internal Medicine of eighteen randomized clinical trials, comprising more than two million individuals, found that although six common cancer screening tests can identify hidden cases leading to individual patient treatment, with the exception of sigmoidoscopy (colorectal cancer), none extended average patient longevity in a statistically significant way.

 For all of the reasons cited above, this is NOT proof that cancer screening does not have benefits when used correctly.  It does, however, suggest that more studies are needed to prove the benefits of any given screening test on its intended population.

A Risk-based approach:

With more and more cancer screening tests being developed, one approach to mitigate the challenges of population screening is to focus on individuals with the highest risk of a particular cancer.

An example of such a trial was the National Lung Screening Trial (NLST). NLST was conducted to determine whether screening with low-dose CT could reduce mortality from lung cancer. Therapeutic intervention was not part of this trial. It enrolled 53,454 people at high risk for lung cancer based on their smoking habits. Participants were randomly assigned to undergo three annual screenings with either low-dose CT or single-view posteroanterior chest radiography. There were 247 deaths from lung cancer per 100,000 person-years in the low-dose CT group and 309 deaths per 100,000 person-years in the radiography group, representing a relative reduction in mortality from lung cancer with low-dose CT screening of 20.0% (p = 0.004). The rate of death from any cause was reduced in the low-dose CT group, as compared with the radiography group, by 6.7% (p = 0.02). This randomized, prospective study design demonstrated the benefit of screening in a high-risk population.

Artificial Intelligence could provide new ways to identify people at highest risk of specific cancers.  A recent article in Nature Medicine showed that a large language model of the type employed by ChatGPT could identify individuals at risk of pancreatic cancer by training machine learning models on the sequence of disease codes in clinical histories and predicting cancer occurrence within specific time windows.  It is possible that developers of screening tests for pancreatic cancer could use such a model to identify individuals at highest risk of the disease thereby reducing the size of trial needed to show benefits to that specific population.

Further support for a risk-based approach comes from the PATHFINDER study published in The Lancet (October 2023). The authors found that the positive predictive value of a blood-based, multicancer early detection test was 43% in people at least 50 years old with additional cancer risk but only 31% in people at least 50 years old without additional cancer risk.

We can’t afford to wait for the perfect screening tests

While we await other prospective screening trials which include all-cause mortality as a clinical endpoint, we should continue current screening programs and seek to develop better ones. In 2023, approximately 609,000 people are expected to die from cancer.  Importantly, the five-year relative survival is significantly higher for those with localized disease versus distant disease. Cancer treatments have been shown to be more effective in earlier stages of disease, where the tumor likely has less subclonal heterogeneity and the tumor microenvironment is less immunosuppressive.   Risk based screening approaches may provide the most bang-for-the-buck in certain cancers (e.g. lung) but have limitations in most cancers where the natural history including risk factors and specific biomarkers are still not well-understood.  Increasingly, targeted molecular tests have the potential to improve screening, diagnosis, and prognosis.  New targeted therapies are increasing the likelihood of effective treatment and even cures.  To avoid unnecessary cancer deaths while waiting for perfect screening protocols, we should continue to develop screening assays that are proven to improve detection of early-stage disease and that do not expose patients to unnecessary risks.