Risk-Based Selection of Individuals for Oral Cancer Screening

Abstract

Purpose: We evaluated proof of principle for resource-efficient, risk-based screening through reanalysis of the Kerala Oral Cancer Screening Trial.

Methods: The cluster-randomized trial included three triennial rounds of visual inspection (seven clusters, n = 96,516) versus standard of care (six clusters, n = 95,354) and up to 9 years of follow-up. We developed a Cox regression-based risk prediction model for oral cancer incidence. Using this risk prediction model to adjust for the oral cancer risk imbalance between arms, through intention-to-treat (ITT) analyses that accounted for cluster randomization, we calculated the relative (hazard ratios [HRs]) and absolute (rate differences [RDs]) screening efficacy on oral cancer mortality and compared screening efficiency across risk thresholds.

Results: Oral cancer mortality was reduced by 27% in the screening versus control arms (HR = 0.73; 95% CI, 0.54 to 0.98), including a 29% reduction in ever-tobacco and/or ever-alcohol users (HR = 0.71; 95% CI, 0.51 to 0.99). This relative efficacy was similar across oral cancer risk quartiles (P interaction = .59); consequently, the absolute efficacy increased with increasing model-predicted risk-overall trial: RD in the lowest risk quartile (Q1) = 0.5/100,000 versus 13.4/100,000 in the highest quartile (Q4), P trend = .059 and ever-tobacco and/or ever-alcohol users: Q1 RD = 1.0/100,000 versus Q4 = 22.5/100,000; P trend = .026. In a population akin to the Kerala trial, screening of 100% of individuals would provide 27.1% oral cancer mortality reduction at number needed to screen (NNS) = 2,043. Restriction of screening to ever-tobacco and/or ever-alcohol users with no additional risk stratification would substantially enhance efficiency (43.4% screened for 23.3% oral cancer mortality reduction at NNS = 1,029), whereas risk prediction model-based screening of 50% of ever-tobacco and/or ever-alcohol users at highest risk would further enhance efficiency with little loss in program sensitivity (21.7% screened for 19.7% oral cancer mortality reduction at NNS = 610).

Conclusion: In the Kerala trial, the efficacy of oral cancer screening was greatest in individuals at highest oral cancer risk. These results provide proof of principle that risk-based oral cancer screening could substantially enhance the efficiency of screening programs.

Trial registration: ClinicalTrials.gov NCT04494620.

FIG 1.

CONSORT diagram. The figure depicts the number of clusters, individuals, years of follow-up, screening visits (for the screening group), oral cancers diagnosed, and oral cancer deaths in the screening and control arms of the Kerala trial. The number of screen-detected cancers reported is the number with positive screening results. Of note, the number of participants (n = 3) and number of oral cancer events (n = 32 oral cancers and n = 16 oral cancer deaths) differ between our analysis and the 2005 report by Sankaranarayanan et al. These differences arise from the exclusion of duplicate participants (n = 3) and data updates available from additional follow-up.

FIG 2.

Oral cancer mortality RD (screening v control arms) across oral cancer risk in the Kerala Oral Cancer Screening Trial. Shown are adjusted oral cancer mortality RDs between the screening and control arm participants in the overall trial participants (A) and among ever-tobacco and/or ever-alcohol users (B). Estimates (circles and squares) and 95% jackknife CIs (error bars) are shown across oral cancer risk prediction model–based quartiles (defined on the basis of the control population). P values shown are for trend across quartiles. See statistical methods for additional details. RD, rate difference.

FIG 3.

Stage distribution of oral cancers detected in the screening and control arms. The figure depicts the percentage of oral cancers detected at stages I, II, III, and IV and with unknown staging in the screening and control arms in the overall trial population (A) and restricted to ever-tobacco and/or ever-alcohol users (B), the highest risk quartile in the overall trial (C), and the highest risk quartile among ever-tobacco and/or ever-alcohol users (D). Chi-square test P = .0059, P = .013, P = .019, and P = .022 for A, B, C, and D, respectively.

FIG 4.

Performance of risk-based and age-based strategies for selection of individuals for oral cancer screening in the Kerala Oral Cancer Screening Trial. The figure depicts the percentage of all oral cancer deaths in the Kerala trial population targeted at varying levels of risk-based and age-based thresholds for the counterfactual selection of individuals for oral cancer screening. Results are shown for each of the selection strategies in the overall trial population and in ever-tobacco and/or ever-alcohol users. See statistical methods for additional details.