TL;DR
This study introduces the HUNT Lung Cancer Model (LCM) as a superior tool for selecting individuals for lung cancer screening compared to existing criteria like NELSON and USPSTF. The HUNT LCM demonstrates significantly improved accuracy in predicting lung cancer cases, identifying more individuals at high risk while potentially reducing unnecessary screenings.
What Problem Does This Paper Address?
Lung cancer (LC) remains a leading cause of cancer mortality, largely due to late-stage diagnosis. While low-dose computed tomography (LDCT) screening has shown promise in reducing LC mortality, current selection criteria (like those from the NLST, NELSON, and USPSTF) are often too broad, leading to many individuals who develop LC being missed, or too narrow, leading to a high number of false positives and unnecessary screenings. There's a critical need for more precise and efficient methods to identify high-risk individuals for screening, balancing sensitivity (catching actual cases) with specificity (avoiding unnecessary interventions).
Methods
The researchers compared the performance of the HUNT Lung Cancer Model (HUNT LCM) against two established criteria: the NELSON trial criteria and the 2021 United States Preventive Services Task Force (USPSTF) criteria.
The HUNT LCM is a risk prediction model developed from the HUNT2 study, utilizing eight variables: sex, age, body mass index, pack-years, cigarettes per day, quit time, hours of indoor smoke exposure, and daily cough. For this study, the HUNT LCM was applied to the CONOR cohort.
To ensure a fair comparison, the study employed a "risk threshold" approach for the HUNT LCM. This meant that the model was used to rank individuals by their predicted 6-year risk of developing LC, and then a threshold was set to select the same number of individuals as the NELSON and USPSTF criteria would. This allowed for a direct comparison of efficiency and predictive power when screening the same proportion of the population.
Statistical significance was assessed using non-parametric permutation-based tests.
Datasets
The study utilized the Cohort of Norway (CONOR), a large, pan-Norwegian prospective population-based study. This dataset comprised 180,534 individuals, from which a subset of 44,831 ever-smokers with complete data was selected for the analysis. This cohort was linked with the Norwegian Cancer Registry and Death Registry for LC diagnosis and death dates.
The HUNT LCM itself was developed from the HUNT2 study, a population-based prospective study of 65,018 individuals.
Key Insights and Ideas
- Risk-based models outperform categorical criteria: The study strongly supports the idea that risk-stratification models like HUNT LCM are more effective than fixed, categorical criteria (like age and pack-year thresholds) for lung cancer screening selection.
- HUNT LCM's adaptability is a key advantage: Unlike fixed criteria, the HUNT LCM's risk threshold can be adjusted to match the screening capacity of different healthcare systems, offering practical flexibility.
- "Symptoms of daily cough" and "hours of indoor smoke" are valuable but potentially problematic: While these variables contribute to the HUNT LCM's predictive power, their subjective nature and potential unavailability in some international databases present a challenge for direct cross-model comparisons in research settings. However, they are readily answerable in clinical practice.
- Focus on "true risk" is paramount: The ultimate goal of screening selection is to identify individuals truly at high risk, and the HUNT LCM appears to achieve this more effectively.
Key Results
- Improved Prediction: When selecting the same number of individuals, the HUNT LCM identified 41.0% more LC cases than the NELSON criteria and 12.1% more LC cases than the 2021 USPSTF criteria.
- Higher Sensitivity: The HUNT LCM achieved significantly higher sensitivity (the ability to correctly identify those who will develop LC) at 52.7% (vs. 37.4% for NELSON) and 66.7% (vs. 59.5% for USPSTF).
- Reduced Number Needed to Screen (NNS): The HUNT LCM required screening fewer individuals to detect one LC case (NNS of 29 vs. 40 for NELSON, and 36 vs. 40 for USPSTF). This indicates greater efficiency.
- Top 16th Percentile Performance: When using the top 16th percentile risk score (0.98% risk in 6 years) as a cutoff, the HUNT LCM identified 70.7% of all LC cases, a 89.2% increase over NELSON and an 18.9% increase over USPSTF.
- Similar Specificity: The HUNT LCM maintained comparable specificity (the ability to correctly identify those who will not develop LC) to the other criteria, suggesting it doesn't significantly increase false positives.
- No Difference in Survival: Importantly, individuals identified by the HUNT LCM who developed LC did not show a difference in median survival compared to those identified by NELSON or USPSTF, indicating the model doesn't select for more indolent cancers.
Limitations and Future Work
- External Validation: The study acknowledges that the HUNT LCM has not yet been validated in external cohorts outside of Scandinavia.
- Incomplete Data Filtering: A potential bias exists due to filtering out incomplete data from the CONOR cohort.
- Missing Risk Factors: The HUNT LCM, developed from the HUNT2 dataset, does not include all known LC risk predictors such as history of COPD, occupational exposure to asbestos or radon, and heredity.
- Subjectivity of Some Variables: The variables "symptoms of daily cough" and "hours of indoor smoke" are subjective and may be difficult to obtain precisely in some research settings or from certain databases.
- Future Work: The authors suggest further validation in diverse populations is ongoing. They also highlight the need to address which specific risk prediction model should be used and are conducting a prospective study comparing the HUNT LCM with other models in Norway. They also mention a "reduced" HUNT model that omits the subjective variables for situations where they are unavailable.
Summary updated