Nomogram to predict the presence of EGFR activating mutation in lung adenocarcinoma

Patients

We retrospectively collected clinical and pathological data on 3,006 patients with NSCLC, including 2,856 with adenocarcinomas, who had their tumours genotyped for EGFR activating mutations at five institutions worldwide: 1,990 adenocarcinoma patients were from Memorial Sloan-Kettering Cancer Center (MSKCC) (New York, NY, USA); 297 were from the Dana-Farber Cancer Institute (DFCI) (Boston, MA, USA); 269 were from Massachusetts General Hospital Cancer Center (MGH) (Boston, MA); 213 were from the Cancer Hospital of Fudan University (CHFU) (Shanghai, China); and 87 were from the National Taiwan University Hospital (NTUH) (Taipei, Taiwan) [14]. The Catalan Institute of Oncology (CIO) (Hospital Germans Trias I Pujol, Badalona, Spain) also contributed data on 217 patients (198 adenocarcinomas) who were ultimately not included in the analyses because of missing data on key variables, such as smoking history.

Starting in 2004 at NTUH, DFCI, MGH and MSKCC, 2005 at CIO, and 2007 at CHFU, EGFR genotyping was systematically and prospectively performed for all consecutive, cytologically or histologically proven adenocarcinoma tumours with available material for testing, without any selection based on additional clinical or pathological criteria. Therefore, patients included in our analysis are representative of all patients treated at the five institutions included.

Data were collected in July 2009 under local institutional review board (IRB)-approved protocols. The present study was approved by each institution IRB.

Mutational analyses

For this analysis, we considered only drug-sensitive EGFR activating mutations, i.e. kinase domain mutations associated with sensitivity to EGFR TKIs: exon 19 deletions, exon 21 mutation (i.e. L858R) and other point mutations (G719A/C/S and L861Q). Tumours harbouring other EGFR mutations were excluded. Direct sequencing was the standard method used for EGFR genotyping (exons 18–21) at DFCI, MGH, CIO, CHFU and NTUH. At MSKCC, EGFR mutational analysis was performed both using direct sequencing and PCR-based, mutation-specific assays to detect the two most common EGFR mutations, i.e. exon 19 deletions and L858R point mutations [15]. We also retrieved the KRAS mutational status of NSCLC tumours, if available. KRAS mutations were assessed using direct sequencing [16].

Clinical and pathological data

Clinical and pathological variables of interest included: age, sex, race/ethnicity, histological type (adenocarcinoma, squamous cell carcinoma or large cell carcinoma), and for adenocarcinoma tumours, histologically predominant subtype [17], tumour stage according to the 6th American Joint Committee on Cancer TNM (tumour, metastasis and node) system that was standard at time the data were generated [18], and smoking variables (smoking status, smoking quantity (measured in pack-yrs) for smokers and time interval since quitting for former smokers). Patients were categorised as never-smokers (<100 lifetime cigarettes), former smokers (quit >1 yr prior to diagnosis) or current smokers (continued smoking within 1 yr of diagnosis). Pathological review was performed at each institution. We also retrieved results from routine immunohistochemical studies performed with an antibody against thyroid transcription factor (TTF)-1, a marker of terminal respiratory unit adenocarcinoma, which is more likely to harbour EGFR mutation [19]. However, because data were lacking, we could not include this covariate in the final model.

Statistical analysis

Analyses were performed on patients diagnosed with NSCLC adenocarcinoma, separately for non-Asian and Asian patients, resulting in one model for each of the two populations. The Chi-squared and two-sample paired t-tests were used to compare the univariate association between clinical parameters and the presence of EGFR or KRAS mutations, for categorical and continuous variables, respectively. The methodology used to develop the prediction nomogram was previously described [13, 20]. Briefly, for each model developed, we randomly divided the available data from patients diagnosed with lung adenocarcinoma into a training set (including two-thirds of the patients), used to develop a prediction model, and a validation set (including the remaining one-third). A model-building exercise was performed to develop multivariate logistic regression models that used available clinical and pathological factors to predict the likelihood of an EGFR mutation. All variables that were univariately associated with EGFR status at a level of α<0.2 were candidates for inclusion in the multivariate model, and were retained in the model if they remained significantly associated with the outcome. Those variables with a markedly skewed distribution were log-transformed; the linearity assumption was further checked for continuous and ordinal variables and, if the functional form of the relationship with the outcome suggested, they were modelled using quadratic polynomials. Levels of the categorical variables were combined if they had similar effects. Interactions between all combinations of two variables were tested but were not included in the final model because their effect did not reach a level of significance of α<0.05 and their addition to the model did not improve the model’s predictive accuracy by >1%.

The discriminatory ability of the model was quantified using the concordance index, which is numerically equivalent to the area under the receiver operating characteristic curve. Concordance index indicates the probability that, in a randomly selected pair in which one patient has and the other patient does not have EGFR mutation, the model correctly predicts a higher probability of mutation for the first patient. In practice, concordance index ranges from 0.5 (no discriminatory ability, equivalent to a coin toss) to 1.0 (perfect discrimination) [21].

Calibration was assessed by plotting the probabilities predicted by the logistic model versus the actual probability. Since the individual actual probability is either 1 or 0 (for patients with and without EGFR mutation, respectively), calibration plots were built using local regression, a method commonly used for data smoothing that relies on approximating each data point with a low-degree polynomial using the neighbouring values. A plot along the 45° line will correspond to a model in which the predicted probabilities are identical with the actual probabilities, therefore indicating perfect calibration. Both discrimination and calibration of the prediction models were calculated on the independent validation dataset.

To protect against the influence of the initial random split, the analysis was repeated in 10 randomly split training and validation datasets, using 10-fold cross-validation, refitting the model to omit one-tenth of the observations to obtain prediction for the omitted tenth. Statistical calculations were performed using SAS version 9.2 (SAS Institute Inc., Cary, NC, USA), and the Design and locfit packages in R (The R Foundation for Statistical Computing, Vienna, Austria).