According to various studies, patient readmission rates to the intensive care unit (ICU) range from 5% to 10%.1 2 3 4 5 Consistently, readmitted patients had much poorer outcomes, higher hospital mortality, and their length of stay (LOS) in hospital was longer.1 3 5 6 7 8 9 Readmissions due to premature ICU discharge are potentially preventable, and may be attributed to deterioration of the primary or existing medical condition. Nevertheless, some readmissions are unavoidable, as there can be occurrence of new complications at any time after initial ICU discharge. Other factors possibly contributing to ICU readmissions are organisational factors, such as ICU occupancy, and availability within a step-down unit.5 10 11Although the early readmission rate has been advocated as an indicator of ICU performance, there is little evidence of a correlation between early ICU readmissions and overall quality of ICU care.2 5 12 13 Risk factors have been identified for ICU readmission.5 7 11 14 15 16 Readmitted patients tend to be older, and have higher severity scores on initial admission and on discharge.1 5 8 15 17 Recently, Gajic et al18 produced a prediction model with acceptable validity.

This present study aimed to identify factors associated with unplanned ICU readmissions by comparing severity-matched cases and controls, whilst focusing on patient variables at the time of ICU discharge. As it had been repeatedly shown that the initial disease severity of a patient was associated with readmissions, we hypothesised that by comparing severity-matched patients, we might identify modifiable risk factors for ICU readmissions, especially those that were potentially preventable.

The study was carried out in the ICU of Pamela Youde Nethersole Eastern Hospital, Hong Kong. This was a 20-bed closed system, mixed medical-surgical adult unit, which provided comprehensive intensive care service to patients in all specialties, except burns, transplant, and cardiothoracic surgery. A nested case-control design was therefore used to facilitate data collection.

Patients with unplanned ICU readmission during the same hospitalisation episode were taken as the study cases. Only the first readmission was used for analysis, whilst patients who died during their index ICU admission and those with elective readmissions were excluded. Each study case was compared with two control patients. Closest matches were selected according to the order of age (range, ± 5 years), initial disease severity according to the Acute Physiology and Chronic Health Evaluation (APACHE) IV risk of death (ROD) [range, ± 5 years], and gender. When there were more than two matched patients, the two having the closest date of ICU admission to the case were selected as controls.

Direct discharge from ICU to home or to another hospital and patients with documented “Do not resuscitate” instruction upon ICU discharge were excluded. Data from 1 January 2008 to 31 June 2010 were obtained for all cases and controls retrospectively, and included their demographic data, functional status and co-morbidities, pre-discharge physiological parameters and laboratory findings, treatments and interventions during the index admission, and time to readmission. The immediate cause of readmission was determined from detailed review of the medical record and was categorised to be of new complication (acquired after ICU discharge) or worsening of a pre-existing condition. Reasons for readmission were classified into eight major categories according to the organ system involved.

The index ICU admissions were defined as the first admission of a case, and the only admission of a control. A patient's pre-existing conditions included the chief medical problem leading to the index ICU admission and its complications. Self-care ability was according to the Karnofsky performance status score.19 Diagnosis of sepsis was based on the clinical judgement of attending physicians with or without microbiological proof. Discharges between 09:00 and 17:59 were daytime discharge. The proportion of ICU beds occupied at time 23:59 of each calendar day was regarded as the ICU occupancy for that day. Early readmissions were defined as readmissions within 72 hours of the index admission discharge, unless stated otherwise.

Values were expressed as mean ± standard deviation (SD) or the number of cases and proportions, as appropriate. Categorical variables were compared using the Pearson Chi squared test or Fisher's exact test, as appropriate. The Student t test or Mann-Whitney U test was used to compare quantitative data. Binary logistic regression with forward stepwise elimination was used for multivariate analysis. Predictor variables of readmission with P≤0.1 in the univariate analysis were included in the multivariate logistic regression. Variables with substantial missing data (>15%) were excluded.

At post-hoc analysis, the classification tree model was employed to identify risks for readmission. This is a standard data mining statistical tool, using non-parametric testing to classify cases into subgroups of the dependent variable, based on the values of the independent variables. Exhaustive Chi squared Automatic Interaction Detector (CHAID) was the splitting method. The analysis was conducted in a stepwise fashion using the Pearson Chi squared test. The predictor variable with the smallest Bonferroni adjusted P value and yielding the most significant split was chosen, and nodes were created that maximised group differences on the outcome. A terminal node was produced when the smallest adjusted P value for any predictor was not significant or the number of cases in the child node was <50. Statistical analyses were conducted using the Statistical Package for the Social Sciences (Windows version 16.0; SPSS Inc, Chicago [IL], US).

Patient characteristics are summarised in Tables 1 and 2. There were no statistical significant differences between readmissions and controls in terms of age, APACHE IV score, APACHE IV acute physiology score, and APACHE IV ROD. The mean (± SD) APACHE IV ROD was 0.3 ± 0.3 in both controls and readmitted group (P=0.84). Despite the APACHE IV score and ROD being matched, there was a statistically significant difference in the mean APACHE IV–predicted LOS between the groups (5.4 ± 2.2 days in controls vs 4.9 ± 2.2 days in the readmitted group; P=0.01).

During this 30-month period, 3202 patients were admitted to the ICU, 380 of whom died in the ICU (361 during their first ICU admission). Of the 2841 patients discharged from the ICU alive following their first ICU stay, 146 went on to have another unplanned ICU admission (ie readmission). Of the 2643 non-readmitted eligible patients who were discharged, 292 were used as matched controls (Fig 1). Thus the unplanned readmission rate was 5.1% (146/2841) among patients surviving their first ICU admission, and the early (within 72 hours) unplanned readmission rate was 2.3% (66/2841). In our case-control cohort (146 readmissions + 292 controls = 438), 191 (43.6%) patients were from general wards, 186 (42.5%) were from operating theatres, 52 (11.9%) were direct admissions from the emergency department, and the remaining admissions were from other sources including coronary care unit and other hospitals. There were 187 (42.7%) medical patients, 146 (33.3%) were surgical and 71 (16.2%) were neurosurgical patients. Of the 438 patients, 363 (82.9%) were emergency admissions.

Among the 146 readmitted patients, 36 (24.7%) had neurological diseases, 35 (24.0%) had gastro-intestinal diseases, and 28 (19.2%) had respiratory diseases as their initial/primary admission diagnosis. Readmitted patients had spent significantly more days in hospital than controls prior to their index admissions (5.2 ± 12.3 vs 2.6 ± 5.2 days; P=0.018; Table 1). Self-care ability before ICU admission and presence of co-morbidities did not differ significantly in the two groups.

Of the 146 unplanned readmitted patients, 66 (45.2%) were early readmissions (within 72 hours of the index admission discharge), 42 (28.8%) were within 48 hours, and 31 (21.2%) within 24 hours. The overall readmission rate for daytime discharges was 5.2% (130/2500), while for nighttime discharges it was 5.1% (16/314). The early readmission rate for daytime discharges was 2.3% (57/2500), while for nighttime discharges it was 2.9% (9/314). The ICU occupancy and nighttime discharges did not have a significant impact on overall readmissions (P=0.844) and readmissions within 72 hours (P=0.096). Higher ICU occupancy was significantly associated with early readmissions (within 48 and 24 hours), compared with late readmissions beyond 48 and 24 hours (t test, P=0.029 and 0.049, respectively).

Among the unplanned readmissions (n=146), 53 (36.3%) were for respiratory causes, 82 (56.2%) for worsening of pre-existing conditions, and 64 (43.8%) for new complications. Among the 82 patients with worsening of pre-existing conditions, 22 (26.8%) had a respiratory admission diagnosis compared to 6/64 (9.4%) who were readmitted for new complications (P=0.008). Postoperative patients accounted for 32/82 (39.0%) of the patients readmitted with worsening of pre-existing conditions, as opposed to 39/64 (60.9%) who were readmitted for new complications (P=0.009).

Compared with patients readmitted for new complications, those readmitted for worsening of pre-existing conditions had significantly longer mean (± SD) index ICU LOS durations (7.2 ± 8.8 vs 4.7 ± 4.8 days; P=0.028) and shorter mean times to readmission (5.0 ± 7.6 vs 14.7 ± 23.4 days; P=0.002). Among those who were readmitted for worsening of pre-existing conditions, the highest proportion was for respiratory problems (36/82, 43.9%). The reasons for readmission for new complications were diverse, but respiratory problems were still the most common (17/64, 26.6%).

Patient outcomes in terms of hospital mortality and mean hospital LOS were significantly worse in the readmitted group, despite being matched for initial severity (Table 1). The difference in outcomes in patients readmitted for worsening of pre-existing conditions or new complications was not statistically significant (Table 2). Patients readmitted early within 72 hours (13/66, 19.7%) had significantly lower mortality than those readmitted beyond 72 hours (32/80, 40%; P=0.008).

Significant findings in the univariate analysis comparing readmissions and controls are shown in Table 1. Factors examined that were not significant included admission type (elective or emergency), admission source; self-care ability before ICU admission; presence of co-morbidities; admission diagnosis; ICU discharge time; ICU occupancy on discharge day; mean arterial blood pressure, heart rate, fractional inspired oxygen (FiO₂), Glasgow Coma Scale (GCS) score on discharge; partial pressure of carbon dioxide in arterial blood, partial pressure of oxygen in arterial blood (PaO₂), white cell count, platelet count, clotting profile, and serum levels of urea, creatinine, and total bilirubin on discharge; whether any anti-arrhythmic agents, inotropic agents, invasive mechanical ventilation, non-invasive ventilation (NIV), tracheostomy, dialysis given at any time during index admission; intubation time; and time from extubation to discharge. Characteristics of patients readmitted for worsening of pre-existing problems and for new complications are shown in Table 2. Patients readmitted for worsening of pre-existing problems had higher mean respiratory rates pre-discharge; more sepsis (especially pulmonary), and more likely to receive NIV. Similarly, patients readmitted early (within 72 hours) also had higher respiratory rates on discharge and were more likely to receive NIV than those readmitted late.

Factors identified as predisposing to ICU readmissions in the multivariate logistic regression were: positive fluid balance in the last 48 hours of the index admission, higher base excess on discharge, and longer hospital stays prior to the index admission (Table 3). Other covariates included: index admission LOS; admission type (postoperative or non-operative); physiological variables including respiratory rate, cardiac rhythm, sputum quantity, and best limb power on discharge; presence of sepsis during the index admission; haematocrit (HCT) on discharge; treatment including mechanical ventilation, re-intubation and tracheostomy during the index admission; and time to last dialysis prior to ICU discharge. Serum albumin values on discharge were excluded, because missing data exceeded 15%.

Tree model 1 shows the determinant factors associated with ICU readmission (Fig 2a). The most significant predictor was whether or not the patient suffered from sepsis during the index admission (adjusted P=0.004, χ² = 8.093). Patients with postoperative sepsis (adjusted P=0.043, χ² = 4.086), and non-operative sepsis with fluid gain on discharge (adjusted P=0.013, χ² = 13.181) increased the readmission risk further. For non-septic patients, sputum quantity on discharge had a significant impact on readmissions (adjusted P=0.006, χ² = 7.528). Tree model 2 demonstrates that septic patients without full limb power at discharge from the ICU had a higher risk of deterioration than those with any other pre-existing condition (Fig 2b). In contrast to readmissions due to new complications, postoperative patients with a HCT of ≤0.34 were at highest risk (Tree model 3, Fig 2c).

In our cohort, 5.1% of those who survived their first ICU admission were readmitted to the ICU; early readmissions amounted to 2.3%. The outcome of readmitted patients was significantly worse than that of those not readmitted, despite being matched for illness severity in terms of APACHE ROD when initially admitted to the ICU. This outcome discrepancy signifies the importance of identifying patients at high risk of deterioration after initial discharge from intensive care. The readmitted group had a significantly shorter APACHE IV–predicted LOS than the controls. Despite this, the actual ICU LOS in the controls was shorter than predicted, while in the readmitted group, it was longer than predicted. This suggested that despite being matched for initial severity, readmitted patients had poorer responses to treatment or had already endured longer initial ICU stays. Not surprisingly, delay in ICU admission increased a patient's risk of readmission; readmitted patients had significantly longer mean values for hospital LOS prior to their index ICU admission, apart from being a significant predictor of ICU readmission in the multivariate analysis. Our study also demonstrated that patients readmitted for worsening of pre-existing conditions and for new complications had different characteristics, but comparable outcomes.

The influence of pulmonary status on the risk of readmission is not debated. Previous studies found pulmonary disorder to be the leading cause of readmissions.1 3 7 15 20 21 The effect of sputum quantity on readmission was likely attributable to insufficient cough effort and retention of secretions by patients. Critically ill patients with neuromuscular complications from severe polyneuropathy and myopathy or deconditioning and weakness were at great risk of sputum retention and nosocomial pneumonia. They were also at risk of hypoventilation and type 2 respiratory failures.22 23 Similar findings were reported in patients with severe head trauma.24 25 In our cohort, patients with neurological diseases constituted the highest proportion of readmissions. Resource allocation for early rehabilitation in the ICU might be warranted.23 Good airway and pulmonary care is crucial for post-discharge patients in step-down units. On the other hand, reducing ventilator-associated pneumonia (VAP) rates by adhering to VAP prevention bundles during the ICU stays may be a way to reduce readmission rates.26 27

Another finding in this study was the effect of fluid balance in the pre-discharge period. Previous studies have illustrated the association of fluid overloading and deleterious outcomes in critically ill patients, including those with sepsis,28 acute kidney injury,29 acute lung injury,28 30 and following operations.31 A single-centre study in Japan32 found that weight gain at the time of initial ICU discharge had a negative linear relationship with the time to ICU readmission, as well as PaO₂-to-FiO₂ ratio. As vigorous fluid resuscitation is often necessary in the initial management of patients with critical illnesses, a proportion of those readmitted to the ICU with respiratory failure could have experienced lung oedema or atelectasis. The current study supports the finding that discharging patients with positive fluid balance leads to a higher readmission rate. Diuresis in critically ill patients could be recognised as a sign of recovery from their illness.

The association of HCT values at discharge and readmission was reported in previous studies, but a cutoff predictive value had not been specified.4 7 In the tree analysis of the subgroup readmitted for new complications, postoperative patients with HCTs of ≤0.34 were associated with an increased risk of readmission. The corresponding haemoglobin levels in patients with HCTs of 0.34 ranged between 110 and 120 g/L. Many confounders complicate the interpretation of HCT. In our cohort, control and readmitted patients were matched for age, gender, and initial disease severity. Thus, lower HCTs in the readmitted group could represent a more severe illness upon ICU discharge or more haemodilution. Yet, according to current transfusion practice in critically ill patients (based on the Transfusion Requirements in Critical Care study), outcomes in those with a restrictive transfusion threshold (7 g/L) were at least equivalent to using a liberal threshold (10 g/L).33 In critically ill patients, observational studies have shown a significant association of red cell transfusions with mortality.34 However, in a more recent multicentred study in Europe,35 an extended Cox proportional hazards analysis showed that patients who received transfusion in fact enjoyed better survival. These contradictory findings remind us that there is no single value for the haemoglobin concentration that justifies transfusion. Patients with poor cardiopulmonary reserve might benefit from a more liberal transfusion threshold.34 In our cohort, postoperative patients with lower HCT values were most vulnerable to new complications that warranted ICU readmission. The stress of major operations to the cardiopulmonary status of an anaemic patient should not be overlooked.

The influence of base excess on readmission was observed in the logistic regression model. Common causes of alkalosis in critically ill patients include contraction alkalosis and renal compensation for respiratory acidosis. It is hypothesised that the majority of our patients with alkalosis were post-hypercapnic and higher readmission rates were seen in patients with more severe hypercapnia on initial presentation. On the other hand, 45% of patients in our cohort were discharged with alkalosis (arterial pH >7.45), whilst only 3.4% (n=15) were discharged with acidosis (arterial pH <7.35). This reflects the tendency to avoid discharging patients with acidosis in our daily practice.

A few previous studies identified the GCS score upon discharge as a risk factor for ICU readmission.5 18 On the contrary, we found that whether or not a patient was discharged with full limb power predicted readmission for worsening pre-existing conditions. We hypothesise that a patient's GCS score upon ICU discharge reflects initial ICU admission severity and status, which was actually matched in our study. For example, a patient admitted with a low GCS score (and thus higher disease severity) is more likely to be discharged with a lower GCS score.

Our case-control design enabled extensive data collection on pre-discharge status. Many of the collected variables have not been reported on previously. In the current study, readmitted and non-readmitted patients were matched for initial severity of illness in terms of APACHE IV ROD. Data collection was focused on the variables that occurred after ICU admission and were modifiable. However, variables reflecting initial disease severity and associated with readmission might have been overlooked. Moreover, the data abstraction and categorisation processes were not blinded to the outcome status of the subjects, and were therefore prone to information bias. Our study did not take into account the proportion of patients who had a poor physician-predicted chance of long-term survival and were therefore not readmitted. As this was a single-centre cohort, the importance of differences in case-mix and patterns of readmission in different ICUs should be recognised.

To the best of our knowledge, this was the first study employing the classification tree for analysis of ICU readmissions. Logistic regression is valuable in providing an indication of the relative importance of each predictor. Higher-order interactions between the predictor variables could be demonstrated in the classification tree analysis. If interactions between independent variables were present, the results of the multiple logistic regression might not be valid. By contrast, factors identified using the tree models might only have an important influence in specific subgroups. For example, the association of sputum quantity with readmission could be hidden if we considered all patients, but not among non-septic patients (Tree model 1).

Our cohort was consistent with previous studies, and suggested that patients having ICU readmissions had significantly poorer outcomes in terms of hospital mortality and hospital LOS. The characteristics of patients readmitted for worsening of pre-existing conditions and for new complications appeared to differ. Incomplete resolution of respiratory conditions remained an important reason for potentially preventable ICU readmission. Attention to patients' fluid balance and sputum quantity before ICU discharge might help to prevent unplanned ICU readmissions. Further study is warranted to investigate the effect of the HCT and pH on critically ill patients.

1. Rosenberg AL, Hofer TP, Hayward RA, Strachan C, Watts CM. Who bounces back? Physiologic and other predictors of intensive care unit readmission. Crit Care Med 2001;29:511-8. CrossRef

2. Rosenberg AL, Watts C. Patients readmitted to ICUs*: a systematic review of risk factors and outcomes. Chest 2000;118:492-502. CrossRef

3. Franklin C, Jackson D. Discharge decision-making in a medical ICU: characteristics of unexpected readmissions. Crit Care Med 1983;11:61-6. CrossRef

4. Rubins HB, Moskowitz MA. Discharge decision-making in a medical intensive care unit. Identifying patients at high risk of unexpected death or unit readmission. Am J Med 1988;84:863-9. CrossRef

5. Kramer AA, Higgins TL, Zimmerman JE. Intensive care unit readmissions in U.S. hospitals: patient characteristics, risk factors, and outcomes. Crit Care Med 2012;40:3-10. CrossRef

6. Yoon KB, Koh SO, Han DW, Kang OC. Discharge decision-making by intensivists on readmission to the intensive care unit. Yonsei Med J 2004;45:193-8.

7. Durbin CG Jr, Kopel RF. A case-control study of patients readmitted to the intensive care unit. Crit Care Med 1993;21:1547-53. CrossRef

8. Alban RF, Nisim AA, Ho J, Nishi GK, Shabot MM. Readmission to surgical intensive care increases severity-adjusted patient mortality. J Trauma 2006;60:1027-31. CrossRef

9. Chan KS, Tan CK, Fang CS, et al. Readmission to the intensive care unit: an indicator that reflects the potential risks of morbidity and mortality of surgical patients in the intensive care unit. Surg Today 2009;39:295-9. CrossRef

10. Priestap FA, Martin CM. Impact of intensive care unit discharge time on patient outcome. Crit Care Med 2006;34:2946-51.

11. Baigelman W, Katz R, Geary G. Patient readmission to critical care units during the same hospitalization at a community teaching hospital. Intensive Care Med 1983;9:253-6. CrossRef

12. Cooper GS, Sirio CA, Rotondi AJ, Shepardson LB, Rosenthal GE. Are readmissions to the intensive care unit a useful measure of hospital performance? Med Care 1999;37:399-408. CrossRef

13. Berenholtz SM, Dorman T, Ngo K, Pronovost PJ. Qualitative review of intensive care unit quality indicators. J Crit Care 2002;17:1-12. CrossRef

14. Metnitz PG, Fieux F, Jordan B, Lang T, Moreno R, Le Gall JR. Critically ill patients readmitted to intensive care units—lessons to learn? Intensive Care Med 2003;29:241-8.

15. Campbell AJ, Cook JA, Adey G, Cuthbertson BH. Predicting death and readmission after intensive care discharge. Br J Anaesth 2008;100:656-62. CrossRef

16. Ho KM, Dobb GJ, Lee KY, Finn J, Knuiman M, Webb SA. The effect of comorbidities on risk of intensive care readmission during the same hospitalization: a linked data cohort study. J Crit Care 2009;24:101-7. CrossRef

17. Frost SA, Alexandrou E, Bogdanovski T, et al. Severity of illness and risk of readmission to intensive care: a meta-analysis. Resuscitation 2009;80:505-10. CrossRef

18. Gajic O, Malinchoc M, Comfere TB, et al. The stability and workload index for transfer score predicts unplanned intensive care unit patient readmission: initial development and validation. Crit Care Med 2008;36:676-82. CrossRef

19. Yates JW, Chalmer B, McKegney FP. Evaluation of patients with advanced cancer using the Karnofsky performance status. Cancer 1980;45:2220-4. CrossRef

20. Chen LM, Martin CM, Keenan SP, Sibbald WJ. Patients readmitted to the intensive care unit during the same hospitalization: clinical features and outcomes. Crit Care Med 1998;26:1834-41. CrossRef

21. Snow N, Bergin KT, Horrigan TP. Readmission of patients to the surgical intensive care unit: patient profiles and possibilities for prevention. Crit Care Med 1985;13:961-4. CrossRef

22. Hermans G, De Jonghe B, Bruyninckx F, Van den Berghe G. Interventions for preventing critical illness polyneuropathy and critical illness myopathy. Cochrane Database Syst Rev 2009;(1):CD006832.

23. Latronico N, Bolton CF. Critical illness polyneuropathy and myopathy: a major cause of muscle weakness and paralysis. Lancet Neurol 2011;10:931-41. CrossRef

24. Sirgo G, Bodi M, Díaz E, Rello J. Pneumonia in head-injured and severe trauma patients. Semin Respir Crit Care Med 2002;23:435-41. CrossRef

25. Cavalcanti M, Ferrer M, Ferrer R, Morforte R, Garnacho A, Torres A. Risk and prognostic factors of ventilator-associated pneumonia in trauma patients. Crit Care Med 2006;34:1067-72. CrossRef

26. Morris AC, Hay AW, Swann DG, et al. Reducing ventilator-associated pneumonia in intensive care: impact of implementing a care bundle. Crit Care Med 2011;39:2218-24. CrossRef

27. O'Keefe-McCarthy S, Santiago C, Lau G. Ventilator-associated pneumonia bundled strategies: an evidence-based practice. Worldviews Evid Based Nurs 2008;5:193-204. CrossRef

28. Wiedemann HP, Wheeler AP, Bernard GR, et al. Comparison of two fluid-management strategies in acute lung injury. N Engl J Med 2006;354:2564-75. CrossRef

29. Bouchard J, Mehta RL. Fluid accumulation and acute kidney injury: consequence or cause. Curr Opin Crit Care 2009;15:509-13. CrossRef

30. Sakr Y, Vincent JL, Reinhart K, et al. High tidal volume and positive fluid balance are associated with worse outcome in acute lung injury. Chest 2005;128:3098-108. CrossRef

31. Stewart RM, Park PK, Hunt JP, et al. Less is more: improved outcomes in surgical patients with conservative fluid administration and central venous catheter monitoring. J Am Coll Surg 2009;208:725-35; discussion 735-7. CrossRef

32. Matsuoka Y, Zaitsu A, Hashizume M. Investigation of the cause of readmission to the intensive care unit for patients with lung edema or atelectasis. Yonsei Med J 2008;49:422-8. CrossRef

33. Hébert PC, Wells G, Blajchman MA, et al. A multicenter, randomized, controlled clinical trial of transfusion requirements in critical care. Transfusion Requirements in Critical Care Investigators, Canadian Critical Care Trials Group. N Engl J Med 1999;340:409-17. CrossRef

34. Marik PE, Corwin HL. Efficacy of red blood cell transfusion in the critically ill: a systematic review of the literature. Crit Care Med 2008;36:2667-74. CrossRef

35. Vincent JL, Sakr Y, Sprung C, Harboe S, Damas P. Are blood transfusions associated with greater mortality rates? Results of the sepsis occurrence in acutely ill patients study. Anesthesiology 2008;108:31-9. CrossRef

Lung cancer is the leading cause of cancer deaths worldwide (1.38 million cancer deaths, 18.2% of the total) as well as of cancer morbidity (1.61 million new cases, 12.7% of all new cancers).1 Approximately 80 to 85% of lung cancer patients have non–small-cell lung cancer (NSCLC), and around 70% of these NSCLC patients present with advanced or metastatic disease (TNM stages IIIB/IV according to the American Joint Committee on Cancer2) at the time of diagnosis.3 4 5 6 Patients with late-stage NSCLC have a very poor prognosis; only about 7% with stage IIIB and 2% of those with stage IV survive beyond 5 years.7

Evidently, NSCLC is a biological and genetic variant of lung cancer, which bears activating mutations in the tyrosine kinase domain of the epidermal growth factor receptor (EGFR). In Asian NSCLC patients, the frequency of activating EGFR mutations (EGFR MuT+) is estimated to be approximately 30 to 40%.6 8 Notably, EGFR mutations lead to structural changes, which stabilise the active form of the tyrosine kinase domain and result in a high affinity for binding EGFR tyrosine kinase inhibitors (TKIs).9

There are currently two small-molecule EGFR TKIs used in clinical practice and recommended as first-line treatment in patients with EGFR MuT+ NSCLC: erlotinib (Tarceva; F. Hoffmann-La Roche Ltd, Basel, Switzerland) and gefitinib (Iressa; AstraZeneca Ltd, London, UK).6 8

Recently published analyses concluded that these EGFR TKIs appear to be the most effective therapy in treatment-naïve cancer patients with this mutation.10 11 As a result, both therapies are competing to be the primary choice in this clinical setting.

This poses the question as to whether there are differences in efficacy and cost-effectiveness between erlotinib and gefitinib. To answer this question and to offer guidance for physicians and health care payers, we undertook comparative effectiveness and cost-effectiveness assessments (CEAs) for the health care setting of Hong Kong.

In order to base the research on the strongest available evidence, standard literature databases (PubMed, ASCO and ESMO congress databases) were screened for Asian randomised controlled phase-III trials that investigated the efficacy of erlotinib and gefitinib as first-line EGFR MuT+ NSCLC therapy. We included all Asian randomised controlled phase-III trials that investigated either gefitinib or erlotinib as first-line therapy of NSCLC, that have systematically assessed the EGFR mutation status of the included patients, and that have published sufficient information on the EGFR-mutation patient population characteristics and outcomes. By applying these criteria, four suitable Asian phase-III randomised controlled trials (RCT) were identified, one for erlotinib and three for gefitinib.

The OPTIMAL trial evaluated the efficacy and tolerability of erlotinib versus chemotherapy,12 13 and the Iressa Pan-ASia Study (IPASS),14 the North-East Japan Gefitinib Study Group trial (NEJGSG),15 and the West Japan Thoracic Oncology Group 3405 trial (WJTOG)16 evaluated the efficacy and safety of gefitinib vs chemotherapy. The following section provides the background information on these clinical trials, which is necessary as a basis for the planned comparative assessments.

As shown in Table 1, the main study characteristics, study measurements, and patient characteristics of the Asian EGFR TKI phase-III RCT for first-line treatment of EGFR MuT+ NSCLC are largely comparable but not identical.

The best fit is encountered with the OPTIMAL and IPASS trials, as the tumour assessment periodicity (6 weekly for both OPTIMAL and IPASS), median age (57 years for both trials), performance status proportion (performance status 0 or 1: OPTIMAL 92%, IPASS 90%), and the tumour stage distribution (stage IV: OPTIMAL 87%, IPASS 86%) were comparable. In contrast, on comparing OPTIMAL versus the NEJGSG and WJTOG trials, differences were evident with respect to all of the above-named factors (Table 1). Such differences are important, as at least age, performance status, and tumour stage were predictors of progression-free survival (PFS) in NSCLC.7 17 18 19

All these phase-III RCT in first-line EGFR MuT+ NSCLCs have shown significant increases in the primary endpoint, namely PFS for erlotinib (OPTIMAL trial12) and gefitinib (IPASS14, NEJGSG15, WJTOG16) in comparison to standard chemotherapy. The erlotinib OPTIMAL trial reached a median PFS of 13.7 months and a corresponding hazard ratio (HR) of 0.16 with 95% confidence intervals (CI) of 0.11-0.26 (P<0.0001).13 The gefitinib IPASS, NEJGSG, and WJTOG trials reached a respective median PFS of 9.5, 10.8, and 8.4 months with corresponding HRs of 0.48 (95% CI, 0.36-0.64; P<0.001), 0.30 (95% CI, 0.22-0.41; P<0.001), and 0.33 (95% CI, 0.20-0.54; P<0.0001).14 15 16

According to all four phase-III RCT, the EGFR TKIs showed a better tolerability profile than the chemotherapy comparators, and hence they appeared to confer less toxicity while achieving greater efficacy.1 12 14 15 16

The most common serious adverse event (SAE) reported for erlotinib is elevation of alanine aminotransferase level (3.6%), which nevertheless compares favourably with gefitinib (27.6%).12 16 Other SAEs with the highest frequency for erlotinib also compare favourably with gefitinib, namely rash (2.4% vs 5.3%) and diarrhoea (1.2% vs 3.8%).12 14 15 Infection is the only SAE that has been reported for erlotinib (1.2%) but not in gefitinib trials.12 All other SAEs reported for gefitinib (aspartate aminotransferase elevation, neutropenia, fatigue, anaemia, anorexia, leukopenia, nausea, paronychia, and sensory disturbance) have not been reported for erlotinib. Irrespective of the small deviations observed when comparing the frequency of single adverse effects between erlotinib and gefitinib, the toxicity of these two TKIs can be regarded as comparable.12

As both EGFR TKIs have shown favourable outcomes compared to chemotherapy, both are currently competing to be the primary choice in treatment-naïve EGFR MuT+ NSCLC patients. Thus, in the absence of a direct head-to-head comparison, there is a need for an indirect comparative effectiveness assessment.

This assessment used the accepted and most widely applied indirect comparison methods introduced by Bucher et al in 1997.20 The Canadian Agency for Drugs and Technologies in Health21 and others22 23 have identified this method as the most suitable approach for performing indirect comparisons of RCT outcomes.

According to the Bucher method,20 the chemotherapy comparator arm (C) of each trial has been used as a ‘bridge’ to connect and compare the efficacy of the investigational treatment arms, namely erlotinib (A) and gefitinib (B). The PFS HRs were selected as the basis for this indirect treatment comparison (ITC), as this efficacy measurement accounts for censoring and incorporates time-to-event information.24 As an outcome of the comparative effectiveness assessment, the ITC HRs of erlotinib versus gefitinib are provided with 95% CIs. The applied ITC approach uses an effect size (PFS HR) that is expressed relative to the comparator (A vs C and B vs C; hence the comparator is used as a ‘bridge’) to perform a so-called ‘adjusted ITC’ of the investigational treatment arms (A vs B). The related formula for the ITC HR is HR_AB = HR_AC /HR_BC and the formula for the ITC 95% CI is HR_AB ± 1.96 x SQRT(VAR[HR_AB]).

In order to test for statistical significance, P values were calculated by means of a two-sided Z test, using the methodology of Snedecor and Cochran 1989.25 The null hypothesis that the PFS of the compared therapy options is equal was to be rejected if P<0.05. All calculations were performed using Excel 2003. The ITC calculations could be re-performed using the ITC tool26 provided by the Canadian Agency for Drugs and Technologies in Health, thus ensuring maximum transparency.

Due to the good fit in prognostic patient characteristics, the key ITC was based on the OPTIMAL versus IPASS phase-III PFS HR outcomes. Furthermore, OPTIMAL was compared with the pooled Asian gefitinib evidence. This pooled evidence was obtained by applying a random effect pooling (PFS HR of gefitinib vs chemotherapy = 0.37; 95% CI, 0.27-0.51; P<0.0001) and a fixed effect pooling (PFS HR of gefitinib vs chemotherapy = 0.38; 95% CI, 0.31-0.46; P<0.0001) to the PFS HR outcomes of the IPASS, NEJGSG, and WJTOG trials.

Phase-III RCT evidence was used as the basis for the CEA. Evidence from OPTIMAL was used for erlotinib and evidence from IPASS for gefitinib, as these studies were the most comparable with respect to prognostic characteristics of the patients (Table 1).

The CEA model uses a Markov approach that simulates the transition between the health states: PFS, progression, and death, in monthly cycles and over a life-time horizon. Patients with stage IIIB/IV EGFR MuT+ NSCLC enter the model in PFS. Transition from PFS to progression is simulated by the published phase-III Kaplan-Meier estimates (erlotinib: OPTIMAL13; gefitinib: IPASS14). For the transition from progression to death, the same transition probability was applied for both EGFR TKIs using the final overall survival results from IPASS.27 This procedure was necessary, as OPTIMAL survival data are currently immature and follow-up is ongoing.12 13

To estimate the Hong Kong–specific drug costs, the licensed dosage (same as in the phase-III RCT) was applied; hence a daily dose of 150 mg for erlotinib and a daily dose of 250 mg for gefitinib were simulated. The drug costs per daily dose of US$74.94 for erlotinib and US$59.98 for gefitinib were based on gross ex-factory prices from October 2011. In order to transfer the local currency (HK$) to US$, the average exchange rates (October 2010 to October 2011) from the Reserve Bank of Australia were used (1 US$ = 7.78 HK$). These drug costs have been simulated until disease progression or death (therapy until progression).

In order to simulate quality-adjusted life years (QALYs), published health utility values according to Nafees et al28 were applied to the health states PFS (0.653) and progression (0.473). A health utility of zero (0) was applied to the health state death. The CEA outcomes were expressed as cost per life year gained, cost per progression-free life year (PF-LY) gained, and as cost per QALY gained for erlotinib and gefitinib. The simulation results were based on a Monte-Carlo simulation using 1000 iterations; all simulations were performed in Excel 2003. Costs and effects have been discounted by 3% per annum according to regional pharmacoeconomic recommendations.29 Sensitivity analyses of the treatment effect on the cost-effectiveness results were performed by applying the extreme bounds (lower and upper 95% CIs) of the PFS Kaplan-Meier estimates for erlotinib and gefitinib.

The World Health Organization (WHO) criterion (incremental cost-effectiveness ratio [ICER] <3 times of the Hong Kong GDP/capita,30 which gave a figure of <US$102 582 or approximately <HK$798 078) was used for this purpose.31

Comparing the PFS HRs of erlotinib versus gefitinib in first-line EGFR MuT+ NSCLC based on OPTIMAL and IPASS resulted in a statistically significant PFS difference in favour of erlotinib (ITC HR=0.33; 95% CI, 0.19-0.58; P=0.0001). As shown in Figure 1, comparing erlotinib versus the pooled gefitinib phase-III evidence confirmed these findings.

According to the CEA model outcomes, erlotinib was more effective in terms of life years gained, in terms of PF-LY gained, and in terms of QALYs gained when compared with gefitinib in first-line EGFR MuT+ NSCLC therapy (Table 2).

The therapy costs of erlotinib were higher than those of gefitinib, as shown in Table 2. Besides higher daily therapy costs, the superior efficacy of erlotinib was the reason for this cost difference. The longer time in PFS compared with gefitinib increased its total therapy duration (therapy until the disease progressed or death), which translated into higher total costs.

To determine whether the additional total therapy costs of erlotinib therapy were reasonable in relation to the efficacy benefit obtained, an incremental cost-effectiveness analysis was performed. The cost per life year gained by erlotinib was US$41 494 (incremental US$ costs 14 061/ incremental life years 0.34), the cost per PF-LY gained by erlotinib was US$39 431 (incremental costs US$14 061/incremental PF-LY 0.36), and the cost per QALY gained by erlotinib was US$62 419 (incremental costs US$14 061/incremental QALY was 0.23) [Fig 2].

These ICERs were well within a range usually regarded as cost-effective using WHO cost-effectiveness criteria. According to these, a therapy is ‘highly cost-effective’ if the ICERs are less than the gross domestic product (GDP) per capita (<US$34 194), ‘cost-effective’ if the ICERs are between 1 (US$34 194) and 3 times (US$102 582) the GDP per capita, and ‘not cost-effective’ if more than 3 times the GDP per capita (>US$102 582).

As shown in Table 3, sensitivity analyses on the treatment effect confirmed the robustness of the cost-effectiveness outcomes as almost all ICERs remained below the WHO cost-effectiveness threshold (<US$102 582).

To offer guidance for physicians and health care payers, comparative effectiveness and CEAs were performed to compare erlotinib versus gefitinib in the treatment of treatment-naïve patients with EGFR MuT+ NSCLC in the health care setting of Hong Kong. Both the comparative assessments used state-of-the-art methods; however, specific limitations had to be taken into account.

The main limitation related to these comparative effectiveness assessments was that our findings were based on indirect evidence. Such ITCs have to be regarded as complementary to clinical trials, because they cannot substitute direct evidence. However, in the absence of any head-to-head comparison, the ITC approach can be regarded as the most valuable way of estimating comparative treatment effects in a statistically accurate manner.

Another limitation was the difference in prognostic patient characteristics between the phase-III trials used. Whereas OPTIMAL and IPASS showed a relatively good fit, the OPTIMAL versus NEJGSG or WJTOG comparisons showed a mismatch of prognostic factors. For this reason, the comparative effectiveness assessment used only the OPTIMAL and the IPASS trials as a basis for the key comparison. Focusing on this key comparison was a necessary precondition for the validity of the ITC, as it ensured that the results were primarily influenced by the treatment effect and not the ‘base risk profiles’. To avoid this confounding factor, the NEJGSG and the WJTOG trials were only considered in a pooled gefitinib PFS HR analysis, which confirmed the findings of the key comparison (OPTIMAL vs IPASS).

Another Asian study, namely the First-Signal study,32 was excluded from our assessment because relevant details on the EGFR MuT+ subpopulation were not published. Besides, the EGFR MuT+ status in this study was assessed only in a limited number of patients and the trial failed to meet its primary endpoint. However, an inclusion of First-Signal study would have worsened the pooled gefitinib results presented in our assessment, as the PFS HR obtained for the EGFR MuT+ population in First-Signal study was the highest obtained within each gefitinib study (PFS HR=0.54; 95% CI, 0.27-1.10) and did not reach statistical significance compared to chemotherapy.

For erlotinib, there was another phase-III RCT available named EURTAC that was performed in a European patient population, and resulted in a PFS HR of 0.37 (95% CI, 0.25-0.54),33 which was not included in our assessment. Compared to patients in the Asian phase-III gefitinib trials,14 15 16 patients in the EURTAC trial33 had the highest median age, the highest proportion with a worse performance status, and the highest proportion with stage IV disease, apart from using a Caucasian patient population which in itself was an important prognostic factor.34 35 36 Thus, according to these prognostic patient characteristics, the only phase-III trial performed in Caucasian patients (EURTAC) was not comparable to any other phase-III trial and hence warranted assessment separately from the Asian evidence. Notably, this was our rationale for basing our assessment for the Hong Kong health care setting on the available Asian evidence only.

The median PFS values of the chemotherapy comparator arms of the selected Asian phase-III trials were 4.6 months in OPTIMAL,12 13 5.4 months in NEJGSG,15 and 6.3 months in the IPASS14 and WJTOG.16 These differences in the median PFS times of chemotherapy have raised doubts about the PFS HRs of the OPTIMAL trial, since it seemed that erlotinib treatment was compared to a comparator arm with the worst performance. This is a frequently applied misinterpretation of the data, as the median PFS values reflect only one point in time in the PFS Kaplan-Meier curve. In order to determine whether one chemotherapy arm shows a better performance than the other (eg comparison between the OPTIMAL chemotherapy arm and the IPASS chemotherapy arm), a comparison of both chemotherapy PFS curves over time on the basis of patient level data from both clinical studies is required. Our comparison approach was based on the HRs (the standard measure for determining the efficacy of oncology drugs), which reflects the area between the PFS Kaplan-Meier curves of the EGFR TKIs versus the chemotherapy comparators, taking into account the whole study period, hence it is not influenced by the different median PFS values.

A possible reason for the PFS difference observed between the two EGFR TKIs might be related to the differences in the chemical structure of erlotinib and gefitinib. These structural differences influence the metabolism of the two drugs by the human liver enzymes. Erlotinib is less susceptible to the metabolizing enzymes than gefitinib and therefore, at an approved dose of 150 mg once daily, it achieves approximately a 3.5-fold higher steady-state plasma concentration than gefitinib administered at the recommended dose of 250 mg once daily.37 This higher circulating level of erlotinib might provide a clinical advantage over gefitinib38 and explain the better efficacy of erlotinib39 compared with gefitinib in the treatment-naïve Asian EGFR MuT+ NSCLC patients.

One limitation of the CEA performed was that total therapy costs were only estimated on the basis of drug costs. In order to perform an adequate total cost assessment, further cost components such as prescription costs, adverse effect costs, and EGFR mutation testing costs usually have to be taken into account. However, as the cost-effectiveness analysis was based on an incremental assessment of erlotinib versus gefitinib, the correctness of results depended on assessing all relevant differences in costs. These differences were considered adequately reflected by differences in drug costs and differences in the therapy duration (difference in PFS) simulated. The rationale for this was that both therapies have comparable prescription and EGFR testing costs, which make no difference when calculating the incremental costs between the two therapies. Only the costs of adverse effects might influence the incremental costs. However, these costs are hard to assess. Although erlotinib shows less SAEs than gefitinib, the difference in the related costs in favour of erlotinib was estimated to be minor.

Another limitation of the cost-effectiveness analysis was the assumption that both TKI therapies present a similar survival probability after disease progression. The survival probability after disease progression was simulated on the basis of the IPASS overall survival outcomes. This assumption was necessary, as the overall survival results from OPTIMAL are still immature. As a result of this assumption, the PFS benefit of erlotinib was transferred to the overall survival outcome. How strongly this assumption impacts the results is currently difficult to determine. Future CEAs using the final OPTIMAL overall survival data (currently immature) are necessary to eliminate this uncertainty

Furthermore, the cost-effectiveness results are not transferable to other health care settings, as they are dependent on country-specific drug prices. Hence, the results presented have to be regarded as specific to the health care setting of Hong Kong and any possible similar findings in other countries and health care settings need to be confirmed in separate analyses.

To the authors’ knowledge, this is the first ITC and CEA performed for treatment-naïve EGFR MuT+ NSCLC in Asian patients. Hence, currently there are no other publications confirming or conflicting with these findings.

The CEA for Hong Kong showed that the cost per life year gained, the cost per PF-LY gained, and the cost per QALY gained by erlotinib were well within an acceptable range in relation to the survival benefit obtained. In conclusion, erlotinib was cost-effective compared to gefitinib as first-line EGFR MuT+ NSCLC in Hong Kong.

This work was funded by Roche Hong Kong Limited (Roche). Roche was involved in gathering the country-specific input data, and in reviewing and commenting the manuscript. All the authors made the decision to submit the manuscript for publication and guarantee the accuracy and completeness of the data. Prof Vivian WY Lee has received educational grant, research contracts, and donations from pharmaceutical companies including AstraZeneca, Boehringer Ingelheim, Eisai, Janssen, Pfizer, Novartis, and Roche.

1. Ferlay J, Shin HR, Bray F, Forman D, Mathers C, Parkin DM. Estimates of worldwide burden of cancer in 2008: GLOBOCAN 2008. Int J Cancer 2010;127:2893-917. CrossRef

2. American Joint Committee on Cancer. AJCC Cancer Staging Manual. 7th ed. New York, NY: Springer; 2010.

3. D’Addario G, Früh M, Reck M, Baumann P, Klepetko W, Felip E. Metastatic non-small-cell lung cancer: ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up. Ann Oncol 2010;21 Suppl 5:v116-9. CrossRef

4. Yang P, Allen MS, Aubry MC, et al. Clinical features of 5,628 primary lung cancer patients: experience at Mayo Clinic from 1997 to 2003. Chest 2005;128:452-62. CrossRef

5. Howlader N, Noone AM, Krapcho M, et al. SEER Cancer Statistics Review, 1975-2008. Available from: http://seer. cancer.gov/csr/1975_2008/. Accessed 15 Sep 2011.

6. National Comprehensive Cancer Network. NCCN Practice Guidelines in Oncology—Asian Consensus Statement—non-small cell lung cancer (Version 1.2009). Available from: http://www.nccn.org/professionals/physician_gls/PDF/nscl-asia. pdf. Accessed 21 Oct 2011.

7. Goldstraw P, Crowley J, Chansky K, et al. The IASLC Lung Cancer Staging Project: proposals for the revision of the TNM stage groupings in the forthcoming (seventh) edition of the TNM classification of malignant tumours. J Thorac Oncol 2007;2:706-14. CrossRef

8. National Comprehensive Cancer Network. NCCN Practice Guidelines in Oncology—Non-Small Cell Lung Cancer (Version 2.2012). Available from: http://www.nccn.org/ professionals/physician_gls/f_guidelines.asp. Accessed 10 Oct 2011.

9. Carey KD, Garton AJ, Romero MS, et al. Kinetic analysis of epidermal growth factor receptor somatic mutant proteins shows increased sensitivity to the epidermal growth factor receptor tyrosine kinase inhibitor, erlotinib. Cancer Res 2006;66:8163-71. CrossRef

10. Paz-Ares L, Soulières D, Klughammer B, Melezínek I, Moecks J, Mok T. Pooled analysis of clinical outcomes in studies of patients with EGFR mutations treated with either an EGFR TKI or chemotherapy. 2009 Jul 16; 2009. Available from: http://www.mojemedicina.cz/files/leciva/prezentace/tarceva/wclcsf/egfr-mutations-analysis-wclc-2009-paz-ares. ppt. Accessed Oct 2013.

11. Bria E, Milella M, Cuppone F, et al. Outcome of advanced NSCLC patients harboring sensitizing EGFR mutations randomized to EGFR tyrosine kinase inhibitors or chemotherapy as first-line treatment: a meta-analysis. Ann Oncol 2011;22:2277-85. CrossRef

12. Zhou C, Wu YL, Chen G, et al. Erlotinib versus chemotherapy as first-line treatment for patients with advanced EGFR mutation-positive non-small-cell lung cancer (OPTIMAL, CTONG-0802): a multicentre, open-label, randomised, phase 3 study. Lancet Oncol 2011;12:735-42. CrossRef

13. Zhou C, Wu J, Chen G, et al. Updated efficacy and quality of life (QoL) analyses in OPTIMAL, a phase III, randomized, open-label study of first-line erlotinib vs gemcitabine/ carboplatin in patients with EGFR activating-mutation positive (EGFR Act Mut+) advanced non-small cell lung cancer (NSCLC). 2011 ASCO Annual Meeting. J Clin Oncol 2011;29 Suppl: abstract 7520.

14. Mok TS, Wu YL, Thongprasert S, et al. Gefitinib or carboplatin-paclitaxel in pulmonary adenocarcinoma. N Engl J Med 2009;361:947-57. CrossRef

15. Maemondo M, Inoue A, Kobayashi K, et al. Gefitinib or chemotherapy for non-small-cell lung cancer with mutated EGFR. N Engl J Med 2010;362:2380-8. CrossRef

16. Mitsudomi T, Morita S, Yatabe Y, et al. Gefitinib versus cisplatin plus docetaxel in patients with non-small-cell lung cancer harbouring mutations of the epidermal growth factor receptor (WJTOG3405): an open label, randomised phase 3 trial. Lancet Oncol 2010;11:121-8. CrossRef

17. Kim ST, Lee J, Sun JM, et al. Prognostic model to predict outcomes in non-small cell lung cancer patients with erlotinib as salvage treatment. Oncology 2010;79:78-84. CrossRef

18. Sardari Nia P, Van Marck E, Weyler J, Van Schil P. Prognostic value of a biologic classification of non-small-cell lung cancer into the growth patterns along with other clinical, pathological and immunohistochemical factors. Eur J Cardiothorac Surg 2010;38:628-36. CrossRef

19. Kefeli U, Kaya S, Ustaalioglu BO, et al. Prognostic factors in elderly patients with non-small cell lung cancer: a two-center experience. Med Oncol 2011;28:661-6. CrossRef

20. Bucher HC, Guyatt GH, Griffith LE, Walter SD. The results of direct and indirect treatment comparisons in meta-analysis of randomized controlled trials. J Clin Epidemiol 1997;50:683-91. CrossRef

21. Wells GA, Sultan SA, Chen L, Khan M, Coyle D. Indirect evidence: indirect treatment comparisons in meta-analysis. Available from: http://www.cadth.ca/index.php/en/hta/ reports-publications/search/publication/884. Accessed 19 Oct 2011.

22. Song F, Altman DG, Glenny AM, Deeks JJ. Validity of indirect comparison for estimating efficacy of competing interventions: empirical evidence from published meta-analyses. BMJ 2003;326:472. CrossRef

23. Tudur C, Williamson PR, Khan S, Best LY. The value of the aggregate data approach in meta-analysis with time-to-event outcomes. J R Stat Soc Ser A Stat Soc 2001;164:357- 70. CrossRef

24. Woods BS, Hawkins N, Scott DA. Network meta-analysis on the log-hazard scale, combining count and hazard ratio statistics accounting for multi-arm trials: a tutorial. BMC Med Res Methodol 2010;10:54. CrossRef

25. Snedecor GW, Cochran WG. Statistical methods. 8th ed. Iowa, US: Iowa State University Press; 1989: 64-82.

26. Wells GA, Sultan SA, Chen L, Khan M, Coyle D. Indirect treatment comparison software application (Version 1.0). Available from: http://www.cadth.ca/en/resources/about-this- guide/download-software. Accessed 21 Oct 2011.

27. Yang CH, Fukuoka M, Mok T, et al. Final overall survival results from a phase III, randomised, open-label, first-line study of gefitinib versus carboplatin/paclitaxel in clinically selected patients with advanced non–small-cell lung cancer in Asia. Presented at 35th ESMO Congress; 2010 Oct 8; Milan, Italy: Abstract LBA2.

28. Nafees B, Stafford M, Gavriel S, Bhalla S, Watkins J. Health state utilities for non small cell lung cancer. Health Qual Life Outcomes 2008;6:84. CrossRef

29. Taiwan Society for Pharmacoeconomic and Outcomes Research. Guidelines of Methodological Standards for Pharmacoeconomic Evaluations in Taiwan 2006. Available from: http://www.ispor. org/PEguidelines/countrydet.asp?c=31&t=1. Accessed 19 Oct 2011.

30. Census and Statistics Department, HKSAR. Hong Kong in figures—2012 Edition. Available from: http://www.statistics. gov.hk/pub/B10100032012AN12B0100.pdf. Accessed 21 Oct 2013.

31. Tan-Torres Edejer R, Baltussen R, Adam T, et al. Making choices in health: WHO guide to cost-effectiveness analysis. Geneva: WHO; 2003. Available from: http://www.who.int/ choice/publications/p_2003_generalised_cea.pdf. Accessed 21 Oct 2013.

32. Han JY, Park K, Kim SW, et al. First-SIGNAL: first-line single-agent iressa versus gemcitabine and cisplatin trial in never-smokers with adenocarcinoma of the lung. J Clin Oncol 2012;30:1122-8. CrossRef

33. Rosell R, Carcereny E, Gervais R, et al. Erlotinib versus standard chemotherapy as first-line treatment for European patients with advanced EGFR mutation-positive non-small-cell lung cancer (EURTAC): a multicentre, open-label, randomised phase 3 trial. Lancet Oncol 2012;13:239- 46. CrossRef

34. Ahn MJ, Lee J, Park YH, et al. Korean ethnicity as compared with white ethnicity is an independent favorable prognostic factor for overall survival in non-small cell lung cancer before and after the oral epidermal growth factor receptor tyrosine kinase inhibitor era. J Thorac Oncol 2010;5:1185- 96. CrossRef

35. Soo RA, Loh M, Mok TS, et al. Ethnic differences in survival outcome in patients with advanced stage non-small cell lung cancer: results of a meta-analysis of randomized controlled trials. J Thorac Oncol 2011;6:1030-8. CrossRef

36. Soo RA, Kawaguchi T, Loh M, et al. Differences in outcome and toxicity between Asian and caucasian patients with lung cancer treated with systemic therapy. Future Oncol 2012;8:451-62. CrossRef

37. Li J, Zhao M, He P, Hidalgo M, Baker SD. Differential metabolism of gefitinib and erlotinib by human cytochrome P450 enzymes. Clin Cancer Res 2007;13:3731-7. CrossRef

38. Ranson M, Hammond LA, Ferry D, et al. ZD1839, a selective oral epidermal growth factor receptor-tyrosine kinase inhibitor, is well tolerated and active in patients with solid, malignant tumors: results of a phase I trial. J Clin Oncol 2002;20:2240-50. CrossRef

39. Hidalgo M, Siu LL, Nemunaitis J, et al. Phase I and pharmacologic study of OSI-774, an epidermal growth factor receptor tyrosine kinase inhibitor, in patients with advanced solid malignancies. J Clin Oncol 2001;19:3267-79.