Mediastinal lymph node involvement in patients with lung cancer is an important determinant of appropriate therapy and, therefore, optimal outcomes. Randomized data show that the use of positron emission tomography (PET) decreases the incidence of futile thoracotomies by up to 50% through better detection of mediastinal and distant disease.1–3 Not surprisingly, the community at large has rapidly adopted PET for lung cancer staging.4,5 An estimated 30% to 50% of patients who undergo PET are found to have a positive mediastinum.6–8 Although there is little disagreement that a positive mediastinum by PET should be evaluated by invasive mediastinal staging,9,10 it is less clear whether a negative mediastinum by PET requires further evaluation.
Pooled data show that 15% to 26% of patients who have a negative mediastinum by PET will have pathologic N2 (pN2) disease discovered at the time of resection.6,7 Single-institution experiences from Asia, the United States, and Europe reveal that pN2 disease is detected intraoperatively in 6% to 16% of patients who had a negative mediastinum by PET.11–16 The rate of pN2 varies according to histologic findings, grade, tumor location, size, and maximized standardized uptake values (SUVMax), evidence of N1 disease by PET, and evidence of nodal disease by computed tomography (CT). These findings have led some individuals to advocate for selective invasive mediastinal staging based on the presence of these risk factors. However, there is no standard approach to quantify the probability of pN2 disease on the basis of multiple risk factors.
We sought to develop and validate a prediction model for pN2 disease. This study was based on the institutional experience at Memorial Sloan-Kettering Cancer Center, where surgeons practice selective pretreatment invasive staging.
PATIENTS AND METHODS
A retrospective cohort study was conducted using patients with radiographic early-stage lung cancer, who underwent pulmonary resection at Memorial Sloan-Kettering Cancer Center between January 2004 and May 2009. The patients excluded were: those who did not undergo PET; who had evidence of mediastinal disease or metastasis by PET; who had evidence of a T3 or T4 lesion by CT; who had suspected or confirmed synchronous, metachronous, or recurrent lung cancer; who had a questionable diagnosis of lung cancer on pathologic review; who had received induction therapy without invasive pretreatment mediastinal staging; who had undergone exploratory thoracotomy or thoracoscopy without resection; and who had no intraoperative nodal sampling performed or had missing information on pretreatment risk factors for pN2 disease. The institutional review board approved this study and waived the need for consent.
For patients with a negative mediastinum by PET, the general approach to preoperative mediastinal staging was selective invasive staging based on clinical factors suggestive of a high risk of occult pN2 and the belief that the patient would benefit more from induction rather than adjuvant chemotherapy. The decision to selectively stage the mediastinum was made on a case-by-case basis by the individual surgeon. The general approach to intraoperative mediastinal staging was sampling or dissection of lymph node stations most likely to drain the resected lung, lobe, or segment.
STATA (Special Edition 9.2; Statacorp, College Station, TX) was used to conduct all statistical analyses. Fisher’s exact test was used to compare categorical variables. A Kruskal–Wallis rank sum test was used to evaluate differences in non-normally distributed continuous variables across groups. Logistic regression analysis was used to evaluate the significance of potential two-way interactions between suspected risk factors for pN2 disease. Kaplan–Meier methods were used to estimate overall survival, starting from the date of resection. Differences in the risk of death were evaluated using Cox regression analysis adjusted for clustering among surgeons. Deaths were ascertained by linkage to the Social Security Death Index, with follow-up through December 31, 2010. p values less than 0.05 was considered to indicate significance.
A prediction model calculates the probability of an event on the basis of the presence or absence of risk factors for that event. Logistic regression was used to predict the probability of pN2 disease—detected either by pretreatment invasive staging or at the time of operation. Selection of variables was guided by risk factors for pN2 disease, which have been previously described in the literature.11–16 Only those variables known to clinicians before the institution of first therapy were considered for inclusion in the model. For instance, several studies describe pathologic tumor size to be a risk factor for N2 disease.14,15 As pathologic tumor size cannot be known before resection, we assumed tumor size determined by CT to be a reasonable surrogate. Six variables were included in the final model: tumor location, tumor size, extent of nodal disease by CT, SUVMax of the primary tumor, N1 disease by PET, and histologic findings. Tumor location and size were determined by CT. Central lesions were defined as those located in the inner two thirds of the lung on CT.9 In a sensitivity analysis, we used an alternative definition of centrality: the inner one third of the lung.10
To evaluate the performance of the prediction model, the entire cohort was randomly split into a development set (consisting of two thirds of the patients) and a validation set (consisting of the remaining one third of the patients). Development of the model was conducted using data from the development set only and was guided by the desire to categorize risk factors in a clinically meaningful and simple fashion while avoiding a potentially overfitted model. Multiple models were considered, varying the number of categories for categorical variables, varying the approach to modeling continuous variables (categorization, splines, nonlinear parameterization guided by fractional polynomial regression, and log transformation), and including interaction terms. Validation of the model was performed using both the development and the validation sets. Calibration—how well the predicted probabilities from the multivariate model match actual probabilities—was visually assessed using calibration plots and was formally tested using the Hosmer–Lemeshow goodness-of-fit test. Discrimination was assessed using the c-statistic. This metric ranges from 0.5 (no discriminatory power equivalent to a coin toss) to 1 (perfect discrimination between 2 possible outcomes).
An exploratory analysis was conducted to compare expected outcomes associated with alternative invasive staging strategies. For a prediction model to be used in clinical practice, a cutoff for the probability of pN2 disease would have to be selected as a trigger for invasive mediastinal staging. We used an empirical approach to select a cutoff, using the Youden Index.17 This method examines all possible cutoffs in the range of a continuous variable (in this case, the probability of pN2 disease) and selects the cutoff that provides the maximum additional sensitivity for the same specificity as that of the uninformative marker. We also evaluated two clinically defined cutoffs based on a high-risk and a low-risk tolerance for failing to diagnose pN2 disease in the preoperative setting. Performance was evaluated by the use of two endpoints: (1) the proportion of patients with pN2 disease detected before treatment, and (2) the proportion of invasive procedures performed among those who ultimately did not have pN2 disease. In scenarios where patients are selected by use of the prediction model to undergo invasive staging and cases of patients undergoing routine invasive staging, only some patients will have pN2 disease detected preoperatively, as the sensitivity of invasive staging is not 100%. To estimate the proportion of patients with pN2 disease detected before treatment, we multiplied the number of patients selected for invasive staging by the sensitivity of invasive staging, divided by the total number of patients with pN2 disease. The sensitivity of invasive staging was estimated using data from this cohort.
Table 1 provides a summary of overall patient and disease characteristics, disease management, and stage-based 5-year survival for the entire cohort. Ninety-seven of 938 patients (10.3%) underwent invasive mediastinal staging, and mediastinoscopy was the most frequently used diagnostic modality. No nodal tissue was obtained from five patients (5.2%)—four of whom underwent mediastinoscopy and one who underwent transbronchial node aspiration. Among the 92 patients from whom nodal tissue was obtained, a median of three (range, 1–4) nodal stations were sampled and a median of five (range, 1–18) nodes were evaluated. The median number of nodal stations and number of nodes evaluated were greater for those who underwent mediastinoscopy. The number of nodal stations sampled during pretreatment invasive staging did not vary significantly across surgeons, but the number of nodes sampled did (Appendix Tables A1–A4). Invasive staging identified pN2 disease in nine patients before first therapy, eight of whom subsequently received induction therapy. Only one patient had multistation disease. Of the 92 patients who underwent invasive staging with successful acquisition of tissue, a total of 18 were ultimately found to have pN2 disease; nine of these cases were identified by invasive staging. Thus, the overall sensitivity of pretreatment invasive staging was 50% (95% confidence interval [CI], 26%–74%).
A total of 93 patients had pN2 disease detected either by pretreatment invasive staging or by intraoperative nodal staging. Of the nine patients who had pN2 disease identified preoperatively, six had persistent N2 disease, and two had N1 disease detected intraoperatively. Intraoperative mediastinal nodal assessment identified an additional 84 cases of previously undetected pN2 disease. A median of two (range, 1–5) mediastinal nodal stations were assessed, with a median of four (range, 1–57) nodes evaluated. Multistation disease was present in 14 patients, with one patient having three-station disease and the remaining two-station disease. The median number of nodal stations and number of nodes evaluated varied significantly (all p < 0.001) by approach to resection and by surgeon (Appendix Tables A1–A4). As expected, survival rates varied significantly by pathologic nodal status (Fig. 1). Table 2 summarizes the final stage, management, and outcomes of patients with pN2 disease.
Univariate analyses of the relationship between known risk factors and pN2 disease were performed using the development set (Table 3). With the exception of tumor location (central versus peripheral) and histologic findings, all previously described risk factors were associated with increased rates of pN2 disease. Given these unexpected findings, a post hoc univariate analysis was conducted using the entire cohort, to evaluate whether reduced power in the small development set might have accounted for the lack of a statistically significant association. Even among the entire cohort, tumor location (p = 0.128) and pretreatment histology (p = 0.059) were not associated with pN2 disease.
A prediction model was developed using the 625 patients in the development set (Table 4). The model had a c-statistic of 0.70 (95% CI, 0.63–0.77). A c-statistic of 0.7 to 0.8 is generally considered indicative of a model with a good discriminatory ability. The goodness-of-fit test revealed a nonsignificant p value (p = 0.56) indicating no significant differences between observed and expected values across deciles of risk. A nonsignificant goodness-of-fit test is generally considered indicative of a model with a good fit. Other potential models—that categorized variables with greater granularity, that modeled continuous variables using nonlinear parameterization, or that included interaction terms—were evaluated, but resulted in unstable estimates or dropped observations. Predicted probabilities ranged from 1.6% to 58.0% in the development set and 1.6% to 55.4% in the validation set. The prediction model performed reasonably well in the validation set (c-statistic, 0.65; 95% CI, 0.56–0.74; goodness-of-fit p = 0.19). The lower c-statistic in the validation set is generally considered indicative of a model with fair discriminatory ability. Figures 2 and 3 show receiver operator curves and calibration plots, respectively, for the development and validation sets.
Because only one a priori risk factor was significantly associated with pN2 disease in the prediction model (Table 4), we conducted several post hoc analyses to further explore this surprising discovery. The use of an alternative definition for centrally located tumors did not change the results. Concerns over reduced power to detect associations in the development set led us to repeat the multivariate regression analysis using the entire cohort. In this model, evidence of N1 disease by PET, evidence of N1 disease by CT (p = 0.02), and pretreatment identification of adenocarcinoma (AC) (p = 0.041) were associated with a higher risk of pN2 disease. We examined a model that included only statistically significant risk factors: nodal status by CT, nodal status by PET, and pretreatment histology. Compared with the model in Table 4, a parsimonious model performed no better in terms of discrimination (development set c-statistic, 0.69; 95% CI, 0.63–0.75; validation set c-statistic, 0.62; 95% CI, 0.54–0.76) and was poorly calibrated when the validation set was analyzed (goodness-of-fit p < 0.001).
Another factor that may have affected the relationship between known risk factors and pN2 disease is the use of pretreatment clinical variables rather than pathologic variables. Accordingly, we evaluated discordance between pre- and postoperative tumor size and histology (Appendix Tables A5 and A6). Among patients who did not receive induction therapy, 44% (95% CI, 40%–47%) had discordance between pre- and postoperative tumor size, as defined by T classification. Among those who had a preoperative diagnosis of cancer, pre- and postoperative histology were discordant in 48% (95% CI, 43%–53%) of subjects. The use of pathologic rather than clinical size and final rather than preoperative histology improved the performance of the model in both the development set (c-statistic, 0.73; 95% CI, 0.68–0.80; goodness-of-fit p = 0.54), and the validation set (c-statistic, 0.68; 95% CI, 0.60–0.76; goodness-of-fit p = 0.68).
An exploratory analysis was conducted to examine potential outcomes associated with five alternative invasive staging strategies (Table 5). The routine invasive staging strategies, and no invasive staging, were compared with three other strategies that used the prediction model, with differing cutoffs for invasive mediastinal staging. The Youden Index was used to empirically derive the cutoff for invasive staging, which was a probability of N2 disease greater than or equal to 8.3%. For instance, for a patient with a central, 2-cm tumor with an SUVMax of 5.2, without evidence of mediastinal nodal disease by CT, without evidence of hilar nodal disease by PET, and with a pretreatment biopsy showing AC, the predicted probability of pN2 disease was 10.9%. By the use of prediction models with either an empirically or a low-risk tolerance clinically derived cutoff, this patient would be selected to undergo invasive mediastinal staging. As the empirically derived sensitivity of invasive staging is only 50%, not all patients selected for invasive staging will have pN2 disease detected preoperatively. Table 5 estimates the expected outcomes for the entire cohort of patients by use of each strategy. The use of different cutoffs to trigger invasive staging modifies the tradeoff between better pretreatment detection of pN2 disease and avoidance of invasive procedures for patients who do not ultimately have pN2 disease.
Finally, subgroup analyses were conducted using patients with pathologically proven AC to explore whether histologic grade or molecular information might be used to enhance the predictive ability of future models. Rates of pN2 disease varied significantly across tumor grade, with notably, no cases of pN2 disease among those with well-differentiated AC (well differentiated, 0.0%; moderately differentiated, 13.0%; poorly differentiated or undifferentiated, 16.0%; and grade not reported, 13.5%; n = 643; p < 0.001). Rates of pN2 disease were not associated with a mutation in either KRAS (KRAS mutation, 14.6%; wild-type, 11.3%; n = 265; p = 0.56) or epidermal growth factor receptor (EGFR mutation, 14.1%; wild-type, 10.8%; n = 364; p = 0.57).
We developed and validated a prediction model for pN2 disease using six previously described risk factors for pN2.11–16 The model had reasonable performance characteristics. Several unexpected findings and post hoc and exploratory analyses from this study may guide prospective, multi-institutional development and validation of a prediction model for pN2 disease.
Surprisingly, of six previously described risk factors for pN2 disease, only one (N1 disease by PET) was associated with pN2 disease in our development cohort. Post hoc analyses from our study suggested two possible contributing factors: (1) reduced power to detect associations in the smaller development set, and (2) the use of clinical variables as surrogates for pathologic variables—for instance, radiographic versus pathologic size or pretreatment biopsy versus final pathologic histology. Although the use of pathologic variables improved the performance of our prediction model, these variables cannot and should not be used for model development because this information is obviously not available to the provider before resection. Our post hoc analyses also lend support to the notion that selection of variables during model development should be guided by a priori risk factors, rather than by statistically significant relationships. Limiting our model to only variables with a statistically significant association with pN2 disease undermined the performance of the model.
Another possible reason why we did not observe a relationship between pN2 disease and previously described risk factors is the lack of a true relationship. Previous studies, like ours, have been single-institution investigations that may not be generalizable. Furthermore, variable inclusion and exclusion criteria, definitions for and the number of risk factors, and differential management of confounding factors make it difficult to interpret results across studies. For instance, among seven investigations,11,13–16,18,19 four studies described a significantly higher risk of pN2 disease among patients with central tumors, whereas three others observed no relationship. There are at least two different definitions for tumor centrality,9,10 and not all previous investigations handled confounding factors. Despite testing both definitions and using multivariate regression analysis, we found no relationship between central tumors and increased risk of pN2 disease. Given the different findings and methods across studies, there is legitimate uncertainty about a true relationship between central tumors and the risk of pN2 disease. Similarly, there may be uncertainty about the relationship between pN2 disease and other risk factors. This uncertainty argues for a prospective, multi-institutional evaluation of risk factors for pN2 disease, which uses a set of potential risk factors, on which there is a general consensus.
There are three other prediction models for pN2, two of which were published after the start of our study. The first was developed using a secondary data analysis of patients who had potentially resectable lung cancer and had enrolled in the Canadian Lung Oncology Group randomized trial of selective versus routine mediastinoscopy.20 The original trial enrolled patients between 1987 and 1990, and thus the applicability of the prediction model is limited in the current era, in which PET is usually used to stage lung cancer.4,5 Two recent prediction models evaluated patients from Asia with clinical stage IA lung cancer.19,21 Each study identified four unique predictors, and their models had reasonable performance characteristics. Our study differed in that we (1) evaluated all patients with radiographically diagnosed early-stage lung cancer with a negative mediastinum by PET, (2) used a priori–selected rather than statistically defined risk factors for pN2 disease, and (3) explored empirically and clinically derived probability cutoffs that would trigger invasive staging. These differences aside, the multinational efforts to develop prediction models for pN2 disease suggest there is international interest in using clinical decision-making tools to aid lung cancer staging.
One perceived limitation of this investigation is that the sensitivity of mediastinoscopy was substantially lower in our study (50%) than in reports from pooled analyses (78%).10 It is important to note that the prevalence of pN2 in pooled analyses was 39%, whereas, in our study it was 10%. Although it is a well-established epidemiologic principle that disease prevalence does not affect the sensitivity of a diagnostic test, it is usually observed that diagnostic tests perform variably in different patient populations. The underlying prevalence of pN2 is likely a good surrogate for different patient populations. Prior pooled analyses included patients with discretely enlarged mediastinal nodes, a normal mediastinum with either a central tumor, or strong suspicion of N1 disease or peripheral stage I tumors, and also some patients with extensive mediastinal infiltration.10 In contrast, our cohort included only patients with a negative mediastinum by PET. Most would agree that these two patient populations are not comparable, and thus, it is reasonable to expect mediastinoscopy to perform differently across studies.
Another perceived limitation of this study is the low overall rate of mediastinoscopy. Rates of mediastinoscopy in this investigation (~10%) are far lower than those reported in population-based databases (~20%).8 However, it is not possible to compare rates, because the patient populations are again different (i.e., operated patients with a negative mediastinum by PET versus all newly diagnosed patients with non–small-cell lung cancer). Moreover, the low mediastinoscopy rate reflects the institutional bias toward a philosophy of selective (rather than routine) invasive mediastinal staging.
Our exploratory analysis comparing various staging strategies provides an objective basis by which to assess the potential effectiveness of prediction models in cancer staging. Changing the probability cutoff for invasive staging modifies the tradeoff between better pretreatment detection of pN2 disease and avoidance of invasive procedures for patients who do not ultimately have pN2 disease. What is less obvious is the finding that a strategy of routine invasive staging does not lead to detection of all cases of pN2—instead, it is limited by the sensitivity of the diagnostic test in a given patient population. Routine invasive staging has associated risks, although small ones.22,23 If surgeons and institutions adhere to practice guidelines,9,10 most patients will undergo invasive mediastinal staging. Use of a prediction model may decrease the frequency of invasive staging procedures—thereby decreasing risks, and potentially, costs—while taking into account an institution’s and/or provider’s risk tolerance for failing to diagnose pN2 disease. An even more provocative application of prediction models in lung cancer staging is to tailor the probability cutoff to the patient’s risk tolerance, thereby promoting shared decision making between provider and patient.
Despite the potential benefits of a prediction model for lung cancer staging, there remains uncertainty about its impact on outcomes. Better staging can lead to better outcomes only if (1) the information garnered from more accurate staging leads to a change in management, which in turn leads to better outcomes; and (2) the change in management occurs frequently enough to realize these expected benefits. Meta-analyses of randomized trials show that both induction chemotherapy followed by resection24 and resection followed by adjuvant chemotherapy25 are superior to resection alone. However, in terms of long-term survival, there is no evidence that induction therapy is superior to adjuvant therapy.26–28 Evidence shows that patients tolerate chemotherapy better preoperatively than postoperatively,27 supporting the conventional wisdom. The benefits of a prediction model may be best quantified in terms of health-related quality of life.
The goal of developing future models is to improve their performance well beyond what is reported by us and other investigators.19,21 A novel approach is to identify molecular characteristics associated with nodal metastases and to incorporate molecular variables into existing prediction models. The presence of gene mutations is currently used to guide treatment decisions in lung cancer care29,30 but not diagnostic decisions. We attempted to identify an association between pN2 disease and mutations among patients who had undergone mutational analysis of EGFR or KRAS as part of their routine care. Unfortunately, we were unable to demonstrate any relationship with pN2 disease. Given the burgeoning number of mutations discovered among patients with squamous cell lung cancer and AC of the lung,31,32 opportunities remain to enhance model performance through incorporation of molecular and genetic information.
Our study has several limitations not previously discussed. A key limitation is the nonuniform approach to intraoperative nodal evaluation across surgeons. Accordingly, the true prevalence of pN2 disease is unknown and has potentially been underestimated. However, the extent to which nodal staging was misclassified is likely small because significant survival differences were observed across pathologic nodal stages. Had significant misclassification occurred, these survival differences would have been blunted or nonexistent. Nonetheless, a small degree of misclassification of pN2 status may have affected the predictive ability of the model. Another important limitation is the generalizability of our results to other centers. We attempted to define the cohort in a manner that would resemble the types of patients (straightforward, radiographically diagnosed early-stage lung cancer) that most providers in the community would likely encounter. Nevertheless, the composition of our cohort may still not be representative of patients elsewhere. Furthermore, exclusion of patients with radiographically staged T3 or T4 tumors may have also limited the generalizability of our results. Though practice guidelines recommend routine invasive mediastinal staging for these tumors,9 this practice may not be occurring in the community at large. In a national study of patients with T4 tumors, only 20% of patients undergoing an operation had previously received a mediastinoscopy.33 Even though patients with evidence of T3 or T4 tumors by CT and a negative mediastinum by PET constitute a small proportion of all patients with lung cancer, it may be worthwhile to include T3 and T4 status as variables in future prediction models. Finally, it may also be valuable to broaden the outcome measure to include patients with pN1 or pN2 disease, particularly if providers are increasingly managing preoperatively detected pN1 disease with induction therapy.
In summary, this study provides proof of principle that prediction models for pN2 disease can be developed and validated among patients with lung cancer. A multi-institutional study is necessary to verify the external validity and generalizability of the model. As an interim step, an investigation is being designed to evaluate this model’s performance at an institution where mediastinoscopy is routinely performed for patients with central stage IA or IB or higher tumors. From a practical standpoint, use of such models may reduce unnecessary practice variation in selective invasive mediastinal staging. Furthermore, prediction models may reduce the frequency, and thus, the associated risks and costs of invasive staging procedures in lung cancer patients with a low probability of pN2 disease. The potential impact of prediction models on outcomes remains uncertain. It is to be hoped that lessons learned from this investigation will inform the development of future prediction models for pN2 disease.
1. van Tinteren H, Hoekstra OS, Smit EF, et al. Effectiveness of positron emission tomography in the preoperative assessment of patients with suspected non-small-cell lung cancer: the PLUS multicentre randomised trial. Lancet. 2002; 359:1388–1393
2. Fischer B, Lassen U, Mortensen J, et al. Preoperative staging of lung cancer with combined PET-CT. N Engl J Med. 2009; 361:32–39
3. Maziak DE, Darling GE, Inculet RI, et al. Positron emission tomography in staging early lung cancer: a randomized trial. Ann Intern Med. 2009; 151:221–228
4. Farjah F, Flum DR, Ramsey SD, Heagerty PJ, Symons RG, Wood DE. Multi-modality mediastinal staging for lung cancer among medicare beneficiaries. J Thorac Oncol. 2009; 4:355–363
5. Dinan MA, Curtis LH, Carpenter WR, et al. Stage migration, selection bias, and survival associated with the adoption of positron emission tomography among medicare beneficiaries with non-small-cell lung cancer, 1998-2003. J Clin Oncol. 2012; 30:2725–2730
6. Gould MK, Kuschner WG, Rydzak CE, et al. Test performance of positron emission tomography and computed tomography for mediastinal staging in patients with non-small-cell lung cancer: a meta-analysis. Ann Intern Med. 2003; 139:879–892
7. Silvestri GA, Gould MK, Margolis ML, et al. American College of Chest Physicians Noninvasive staging of non-small cell lung cancer: ACCP evidenced-based clinical practice guidelines (2nd edition). Chest. 2007; 132:(3 Suppl)178S–201S
8. Little AG, Gay EG, Gaspar LE, Stewart AK. National survey of non-small cell lung cancer in the United States: epidemiology, pathology and patterns of care. Lung Cancer. 2007; 57:253–260
9. National comprehensive cancer network clinical practice guidelines in oncology–v.3.2012: non-small cell lung cancer.
10. Detterbeck FC, Jantz MA, Wallace M, Vansteenkiste J, Silvestri GA. American College of Chest Physicians Invasive mediastinal staging of lung cancer: ACCP evidence-based clinical practice guidelines (2nd edition). Chest. 2007; 132:(3 Suppl)202S–220S
11. Al-Sarraf N, Aziz R, Gately K, et al. Pattern and predictors of occult mediastinal lymph node involvement in non-small cell lung cancer patients with negative mediastinal uptake on positron emission tomography. Eur J Cardiothorac Surg. 2008; 33:104–109
12. Cerfolio RJ, Bryant AS. Survival of patients with unsuspected N2 (stage IIIA) nonsmall-cell lung cancer. Ann Thorac Surg. 2008; 86:362–6; discussion 366
13. Defranchi SA, Cassivi SD, Nichols FC, et al. N2 disease in T1 non-small cell lung cancer. Ann Thorac Surg. 2009; 88:924–928
14. Kanzaki R, Higashiyama M, Fujiwara A, et al. Occult mediastinal lymph node metastasis in NSCLC patients diagnosed as clinical N0-1 by preoperative integrated FDG-PET/CT and CT: risk factors, pattern, and histopathological study. Lung Cancer. 2011; 71:333–337
15. Lee PC, Port JL, Korst RJ, Liss Y, Meherally DN, Altorki NK. Risk factors for occult mediastinal metastases in clinical stage I non-small cell lung cancer. Ann Thorac Surg. 2007; 84:177–181
16. Park HK, Jeon K, Koh WJ, et al. Occult nodal metastasis in patients with non-small cell lung cancer at clinical stage IA by PET/CT. Respirol. 2010; 15:1179–1184
17. Youden WJ. Index for rating diagnostic tests. Cancer. 1950; 3:32–35
18. Gómez-Caro A, Garcia S, Reguart N, et al. Incidence of occult mediastinal node involvement in cN0 non-small-cell lung cancer patients after negative uptake of positron emission tomography/computer tomography scan. Eur J Cardiothorac Surg. 2010; 37:1168–1174
19. Zhang Y, Sun Y, Xiang J, Zhang Y, Hu H, Chen H. A prediction model for N2 disease in T1 non-small cell lung cancer. J Thorac Cardiovasc Surg. 2012; 144:1360–1364
20. Shafazand S, Gould MK. A clinical prediction rule to estimate the probability of mediastinal metastasis in patients with non-small cell lung cancer. J Thorac Oncol. 2006; 1:953–959
21. Koike T, Koike T, Yamato Y, Yoshiya K, Toyabe S. Predictive risk factors for mediastinal lymph node metastasis in clinical stage IA non-small-cell lung cancer patients. J Thorac Oncol. 2012; 7:1246–1251
22. Lemaire A, Nikolic I, Petersen T, et al. Nine-year single center experience with cervical mediastinoscopy: complications and false negative rate. Ann Thorac Surg. 2006; 82:1185–9
23. Park BJ, Flores R, Downey RJ, Bains MS, Rusch VW. Management of major hemorrhage during mediastinoscopy. J Thorac Cardiovasc Surg. 2003; 126:726–731
24. Song WA, Zhou NK, Wang W, et al. Survival benefit of neoadjuvant chemotherapy in non-small cell lung cancer: an updated meta-analysis of 13 randomized control trials. J Thorac Oncol. 2010; 5:510–516
25. Pignon JP, Tribodet H, Scagliotti GV, et al. LACE Collaborative Group Lung adjuvant cisplatin evaluation: a pooled analysis by the LACE Collaborative Group. J Clin Oncol. 2008; 26:3552–3559
26. Lim E, Harris G, Patel A, Adachi I, Edmonds L, Song F. Preoperative versus postoperative chemotherapy in patients with resectable non-small cell lung cancer: systematic review and indirect comparison meta-analysis of randomized trials. J Thorac Oncol. 2009; 4:1380–1388
27. Felip E, Rosell R, Maestre JA, et al. Preoperative chemotherapy plus surgery versus surgery plus adjuvant chemotherapy versus surgery alone in early-stage non–small-cell lung cancer. J Clin Oncol. 2010; 28:3138–3145
28. Martins RG, D’Amico TA, Loo BW Jr, et al. The management of patients with stage IIIA non-small cell lung cancer with N2 mediastinal node involvement. J Natl Compr Canc Netw. 2012; 10:599–613
29. Pao W, Girard N. New driver mutations in non-small-cell lung cancer. Lancet Oncol. 2011; 12:175–180
30. Febbo PG, Ladanyi M, Aldape KD, et al. NCCN Task Force report: evaluating the clinical utility of tumor markers in oncology. J Natl Compr Canc Netw. 2011; (9 Suppl 5)S1–32
31. Ding L, Getz G, Wheeler DA, et al. Somatic mutations affect key pathways in lung adenocarcinoma. Nature. 2008; 455:1069–1075
32. Cancer Genome Atlas Research Network Comprehensive genomic characterization of squamous cell lung cancers. Nature. 2012; 489:519–525
33. Farjah F, Wood DE, Varghese TK Jr, Symons RG, Flum DR. Trends in the operative management and outcomes of T4 lung cancer. Ann Thorac Surg. 2008; 86:368–374