Predicting safety and efficacy of afatinib treatment of NSCLC

The primary aim of this project is to improve understanding and prediction of efficacy and toxicity associated with afatinib use for the treatment of non-small cell lung cancer.Specific focus will be on potential predictors with preliminary evidence of association with an outcome of interest. Other variables with biologically or clinically plausible association with the outcome will also be explored. The potential predictors to be explored are: •   age, •   sex •   smoking history/status, •   performance status, •   grade of adverse events, •   EGFR activating mutations, •   cancer subtype, stage •   ethnicity/race•   extent and sites of disease,•   previous treatment for lung cancer (including response to any prior EGFR TKI use) [for studies in which afatinib is not used first line]. •   family history, •   imaging •   laboratory values (excluding plasma concentration of afatinib)Data: approximately 2000 patients enrolled into the LUX-LUNG 1-6 trials were using afatinib. Assuming approximately 50% overall survival and progression free survival event rates across these trials this corresponds to approximately 1000 events. Thus, there will be at least 10 events per candidate predictor evaluated in this study. Although prediction of outcomes following afatinib therapy are of primary interest, data from comparator arms of patients not using afatinib will also be evaluated to provide insight into whether predictors identified are specific to afatinib therapy or are generally prognostic for the disease.Missing Data:Multiple imputation by chained equations (n=20) will be applied for analyses involving >5% overall missing data, otherwise a complete case analysis will be used. Quality Control Plans: Individual variables required for the analysis will be extracted/constructed based on the raw data files and data dictionaries provided. To ensure that each variable has been correctly prepared from data provided, basic analyses and descriptive statistics will be reproduced to check for consistency with pertinent results in clinical study reports. Where there are insufficient published results to confirm the proper extraction of the variable, values will be manually checked against a random sample of the original raw data file. All analyses will be coded and documented in an analysis script to facilitate transparency and reproducibility of analyses. All analysis code and results will be reviewed by a second biostatistician involved the project.Initial Evaluation of Candidate PredictorsThe Cox proportional hazards model will be primary used to evaluate the crude association between potential predictors and clinical outcomes, stratified by clinical study. Continuous variables (e.g. age, LDH) will be evaluated for non-linearity of association with outcomes by using restricted cubic splines. Multivariable analysis will be undertaken to evaluate the importance and heterogeneity of individual predictors with respect to information provided by other predictors. It is recognised that (1) patients/clinicians will need to re-evaluate decisions on whether to start/stop/continue/recommence therapy at a number of time points over the course of therapy, and (2) the predictors that are available and most useful may change over the course of therapy. A prediction model will be developed to make the best possible prediction of outcomes at baseline (before initiation of afatinib therapy), and at landmark time points early in treatment which may include early markers of outcomes as predictors. Prediction Model DevelopmentPrediction models will be developed should a sufficient number of predictors be found to have significant association with outcomes in the initial evaluation phase. Penalised Cox proportional hazards regression will be primarily utilised to develop prediction models in order to improve external prediction and promote parsimony. Specifically, the least absolute shrinkage and selection operator (Lasso) will be utilised with the penalty optimised based on cross-validation. The rationale being that a prediction model with a relatively small number of predictors is generally simpler to use in clinical practice and thereby facilitates clinical translation. The performance of each prediction model will evaluated in terms of both discrimination, using Harrell's c-index for the time-to-event, and calibration using graphic assessment and the Hosmer-Lemeshow ?2-tests for the time-to-event data. Bootstrapping (including all modelling steps) will be utilised for internal validation in order to attenuate optimistic apparent performance estimates. In the context of a pooled analysis of clinical studies an additional validation procedure will involve every study being left out once, for validation of a model based on the remaining studies. Tests for heterogeneity in predictor effects between studies will be undertaken to further evaluate the likely generalisability of model performance. Should models with reasonable prediction performance result from this project, the intention will be to externally validate the prediction performance using a separate (non clinical trial) cohort [outside the scope of this proposal].