Calibration and validation of the LLP lung cancer risk stratification model


Session type:

Daniel Vulkan, John Field, Michael Davies, Stephen Duffy, Rhian Gabe



Evidence from two large randomised trials estimates a significant 20-24% reduction in lung cancer mortality in the low dose CT screening arm. Future lung cancer screening programmes require identification of high-risk populations to optimise detection in those who would benefit from screening and provide a cost-effective programme. The Liverpool Lung Project risk model (LLPv1) was developed in 2008, validated in three international datasets, and amended for use in the UKLS trial (LLPv2). However, it was based on data from the north west of England 1997-2005.


The model was calibrated using more recent incidence data and standardised incidence ratios from England as a whole, to create LLPv3. In order to validate the calibration of the predictive model, we looked at the 75,958 respondents to the first approach UKLS trial questionnaire, for each of whom we obtained at least 5 years of follow-up data. A ROC curve was plotted to assess the ability of the LLP model to distinguish between those individuals who went on to develop lung cancer. We also calculated the proportionate difference (the percentage excess or deficit in the number of observed cancers compared to those predicted by the model).


The ROC curves for LLPv2 and LLPv3 were almost identical, as would be expected, since the coefficients pertaining to the (non age- and sex-specific) risk factors are the same; the ranking of most individuals is unchanged from LLPv2 to LLPv3. The AUC in each case was 0.81 (95% CI 0.79 - 0.82). 

Under LLPv2, the proportionate difference was -58.3% (95% CI -61.6% to -54.8%) ie the actual number of cancers was only 42% of the number predicted. Using LLPv3, this improved to -22.0% (95% CI -28.1% to -15.5%).


The discrimination of LLPv2 and LLPv3 was excellent. LLPv3, which was calibrated to contemporary, English incidence, achieved more accurate prediction of absolute incidence than LLPv2 (considerably less overestimation), and would be more effective in selecting a high-risk group for surveillance in England today.

Impact statement

LLPv3 has been shown to be the more appropriate risk model for identifying individuals who would benefit from a national lung cancer screening programme.