Purpose:

To systematically review the evidence on effectiveness, accuracy, and harms of screening for lung cancer with low-dose computed tomography (LDCT) for populations and settings relevant to primary care in the United States.

Data Sources:

PubMed/MEDLINE, the Cochrane Library, and trial registries through May 28, 2019; reference lists of retrieved articles; outside experts; and reviewers, with surveillance of the literature through November 20, 2020.

Study Selection:

English-language controlled trials of screening for lung cancer with LDCT; studies evaluating LDCT screening accuracy; studies of risk prediction models comparing benefits and harms of screening vs. the use of trial eligibility criteria or 2013 U.S. Preventive Services Task Force recommendations; trials and prospective cohort studies of treatment for Stage I lung cancer with surgery or stereotactic body radiotherapy reporting at least 5-year survival; prospective cohort and case-control studies reporting harms.

Data Extraction:

One investigator extracted data and a second checked accuracy. Two reviewers independently rated quality for all included studies using predefined criteria.

Data Synthesis:

This review included 223 publications. Seven randomized, controlled trials (RCTs) (described in 26 articles; 86,486 participants) evaluated lung cancer screening with LDCT; the National Lung Screening Trial (NLST) and Nederlands-Leuvens Longkanker Screenings Onderzoek (NELSON) were the only RCTs that were adequately powered. The NLST found a reduction in lung cancer mortality (calculated incidence rate ratio [IRR], 0.85 [95% confidence interval {CI}, 0.75 to 0.96]) and all-cause mortality (calculated IRR, 0.93 [95% CI, 0.88 to 0.99]) with three rounds of annual LDCT screening compared with chest X-ray for high-risk current and former smokers ages 55 to 74 years. These findings indicate a number needed to screen (NNS) to prevent one lung cancer death of 323 over 6.5 years of followup. NELSON found a reduction in lung cancer mortality (calculated IRR, 0.75 [95% CI, 0.61 to 0.90]) but not all-cause mortality (calculated IRR, 1.01 [95% CI, 0.92 to 1.11]) with four rounds of LDCT screening with increasing intervals (at baseline, 1 year, 3 years, and 5.5 years) compared with no screening for high-risk current and former smokers ages 50 to 74 years. These findings indicate an NNS to prevent one lung cancer death of 130 over 10 years of followup. The sensitivity of LDCT ranged from 59 to 100 percent (13 studies; n=76,856) and was over 80 percent in most studies. The specificity ranged from 26.4 to 99.7 percent (13 studies; n=75,819) and was over 75 percent in most studies. The positive predictive value (PPV) ranged from 3.3 to 43.5 percent. The negative predictive value ranged from 97.7 to 100 percent. Evidence suggests that using the Lung-RADS™ classification system in the NLST would have increased specificity while decreasing sensitivity and increasing nodule size threshold for a positive screening result would increase PPV. Harms of screening included radiation-induced cancer (estimated 0.26 to 0.81 major cancers for every 1,000 people screened with 10 annual LDCTs), false-positive results leading to unnecessary tests and invasive procedures, overdiagnosis, incidental findings, and short-term increases in distress because of indeterminate results. For every 1,000 persons screened in the NLST, false-positive results led to 17 invasive procedures (number needed to harm, 59), resulting in less than one major complication. Using Lung-RADS could reduce false-positive results compared with the NLST criteria; estimates suggest that using Lung-RADS could have prevented about 23 percent of all invasive procedures for false positives in the NLST. Overdiagnosis estimates ranged from a 0 to 67 percent chance that a screen-detected lung cancer was overdiagnosed. The NLST data indicate approximately four cases of overdiagnosis (and 3 lung cancer deaths prevented) per 1,000 people screened over 6.5 years. Incidental findings were common and variably defined with a wide range reported across studies (4.4% to 40.7% of people screened).

Modeling studies estimated that using risk prediction models would increase the number of screen-preventable deaths, reduce the number of participants needed to screen to prevent one lung cancer death, and reduce the number of false positive selections (i.e., selecting persons to be screened who did not have or develop lung cancer or death from lung cancer) per prevented lung cancer death compared with risk factor–based screening, when NLST-like cancer detection and mortality reductions were assumed, but the strength of evidence was low because it was largely derived from post hoc application to trial data and modeling.

Limitations:

NLST and NELSON participants were younger, more highly educated, and less likely to be current smokers than the U.S. screening-eligible population, and they had limited racial and ethnic diversity. The general U.S. population eligible for lung cancer screening may be less likely to benefit from early detection compared with the NLST and NELSON participants because they face a higher risk of death from competing causes and the NLST and NELSON were mainly conducted at large academic centers, potentially limiting applicability to community-based practice. Most studies reviewed in this report (including the NLST) did not use current nodule evaluation protocols such as Lung-RADS.

Conclusions:

Screening high-risk persons with LDCT can reduce lung cancer mortality and may reduce all-cause mortality but also causes false-positive results leading to unnecessary tests and invasive procedures, overdiagnosis, incidental findings, increases in distress, and, rarely, radiation-induced cancers. The evidence for benefits comes from two RCTs that enrolled participants who were more likely to benefit than the U.S. screening-eligible population and that were mainly conducted at large academic centers, potentially limiting applicability to community-based practice. Application of lung cancer screening with current nodule management protocols (e.g., Lung-RADS) might improve the balance of benefits and harms. Use of risk prediction models might improve the balance of benefits and harms, although there remains considerable uncertainty about how such approaches would perform in actual practice because current evidence does not include prospective clinical utility studies.