Interobserver variability of recall decisions between mammography readers in the English NHS breast screening programme: A comparison of interobserver variability measures

TL;DR

This study compares how consistently mammogram readers decide to recall women for further testing in breast cancer screening, highlighting issues with commonly used statistical measures.

Contribution

The paper evaluates and compares different measures of interobserver variability in mammography recall decisions, emphasizing the limitations of Cohen’s kappa in low-prevalence settings.

Findings

Percentage agreement, Gwet’s AC, and PABAK showed lower agreement for first screening appointments compared to subsequent ones.

Cohen’s kappa was found to be heavily distorted by outcome prevalence, making it unsuitable for low-prevalence screening settings.

Measures like Gwet’s AC and PABAK were more informative for assessing variability in challenging screening scenarios.

Abstract

To evaluate interobserver variability between mammogram readers’ recall decisions in the English NHS breast screening programme, comparing different variability measures.

Data from 401,682 women in 22 NHS centres who underwent mammographic screening interpreted independently by two mammogram readers were included. Percentage agreement, prevalence-adjusted bias-adjusted-kappa (PABAK), Gwet’s agreement coefficient (Gwet’s AC) and Cohen’s kappa were reported with 95% confidence intervals. Analyses were performed separately for women at first and subsequent screening appointments, by cancer diagnosis, reader recall rates and age group.

Of 86,287 women at first screening, 6,491 (7.5%) were recalled, compared to 9,488 (3.0%) of 315,395 at subsequent screenings. Percentage agreement, Gwet’s AC, and PABAK were lower for first screening than subsequent (93.6%, 95%CI: 93.4–93.7 vs 97.2%, 95%CI: 97.2–97.3), (92.3, 95%CI:92.1 to 92.5 vs 97.0, 95% CI: 97.0 to 97.1) and (87.2, 95%CI: 86.9–87.4 vs 94.4, 95%CI: 94.3–94.5), whereas Cohen’s kappa, which is biased downwards when prevalence of recall is lower, did not change (61.6, 95%CI: 60.7–62.5 vs 61.8, 95%CI: 61.0–62.5). Percentage agreement, Gwet’s AC, and PABAK were lower for women with cancer detected than without, but Cohen’s kappa showed the opposite pattern, driven by prevalence bias. Percentage agreement, Gwet’s AC, and PABAK were lower when one/both readers had high recall rates, but Cohen’s kappa showed no important pattern.

Percentage agreement, Gwet’s AC, and PABAK showed lower agreement for interpreting the more challenging first screen, without assistance of previous mammograms, when women had cancer and when one/both readers had a high recall rate. Cohen’s kappa was heavily distorted by outcome prevalence. Despite widespread use, Cohen’s kappa is inappropriate for low prevalence settings such as screening, or making comparisons when prevalence varies.

Figures