An international prognostic index for patients with chronic lymphocytic leukaemia (CLL-IPI): a meta-analysis of individual patient data

Articles

An international prognostic index for patients with chronic lymphocytic leukaemia (CLL-IPI): a meta-analysis of individual patient data The International CLL-IPI working group*

Summary Background The management of patients with chronic lymphocytic leukaemia is currently undergoing improvements due to novel therapies and a plethora of biological and genetic variables that add prognostic information to the classic clinical staging systems. We established an international consortium with the aim to create an international prognostic index for chronic lymphocytic leukaemia (CLL-IPI) that integrates the major prognostic parameters. Methods We used results from a systematic search of the Cochrane Haematological Malignancies Group of MEDLINE, Embase, and Central databases for prospective, clinical phase 2 and 3 trials of chronic lymphocytic leukaemia, published between Jan 1, 1950, and Dec 31, 2010, which identified 13 trials. We contacted the principal investigators of these 13 trials, of which eight agreed to include individual patient data. We used the individual patient data from these phase 3 trials from France, Germany, Poland, the UK, and the USA to create the full analysis dataset. The full analysis dataset was randomly divided, using a random sample procedure, into training and internal-validation datasets. We did a univariate analysis and multivariate analyses using 27 baseline factors and overall survival as an endpoint. We assigned weighted risk scores to each factor included in the final multivariable model. We assessed the discriminatory value using C-statistics and also the validity and reproducibility of the CLL-IPI by subgroup analysis. We used two additional datasets from the Mayo Clinic (Rochester, MN, USA; MAYO cohort) and the SCALE Scandinavian population-based case-control study (SCAN cohort) as the external-validation datasets. Findings 3472 treatment-naive patients were included in the full analysis dataset; 2308 were randomly segregated into the training dataset and 1164 into the internal-validation dataset. 838 patients were included in the MAYO cohort and 416 in the SCAN cohort. Median age of patients in the full analysis dataset was 61 years (range 27–86). Five independent prognostic factors were identified in the training dataset: TP53 status (no abnormalities vs del[17p] or TP53 mutation or both), IGHV mutational status (mutated vs unmutated), serum β2-microglobulin concentration (≤3·5 mg/L vs >3·5 mg/L), clinical stage (Binet A or Rai 0 vs Binet B–C or Rai I–IV), and age (≤65 years vs >65 years). Using a weighted grading of the independent factors, a prognostic index was derived that identified four risk groups within the training dataset with significantly different overall survival at 5 years: low (93·2% [95% CI 90·5–96·0]), intermediate (79·3% [75·5–83·2]), high (63·3% [57·9–68·8]), and very high risk (23·3% [12·5–34·1]; log-rank test comparing survival across the four risk groups p<0·0001; C-statistic, c=0·723 [95% CI 0·684–0·752]). These risk groups were confirmed in the internal-validation and external-validation datasets.

Lancet Oncol 2016 Published Online May 13, 2016 http://dx.doi.org/10.1016/ S1470-2045(16)30029-8 See Online/Comment http://dx.doi.org/10.1016/ S1470-2045(16)30052-3 *Members of The International CLL-IPI working group are listed in the appendix Correspondence to: Dr Michael Hallek, Department I of Internal Medicine and Centre of Integrated Oncology Cologne Bonn, and Cologne Centre of Excellence for Cellular Stress Response and Ageing Related Diseases (CECAD), University of Cologne, 50937 Cologne, Germany [email protected] See Online for appendix

Interpretation The CLL-IPI combines genetic, biochemical, and clinical parameters into a prognostic model, discriminating four prognostic subgroups. The CLL-IPI will allow a more targeted management of patients with chronic lymphocytic leukaemia in clinical practice and in trials testing novel drugs. Funding José Carreras Leukaemia Foundation

Introduction The clinical staging systems for chronic lymphocytic leukaemia were developed by Rai1 and Binet2 nearly 40 years ago and represent the backbone of prognostication in clinical practice and trials. In the past two decades, the introduction of novel therapies and greater insight into the genetic and molecular biology of this cancer have led to the identification of various markers associated with survival, providing prognostic information that is complementary to the classic staging systems.3,4 In particular, molecular investigations of chronic lymphocytic leukaemia cells have shown that the deletion of the short arm of chromosome 17 (del[17p]) or mutations

of the tumour suppressor gene TP53 predict both an aggressive disease course and refractoriness to chemoimmunotherapy.5–7 The mutational status of the immunoglobulin heavy chain IGHV genes is also associated with survival, whereby patients with unmutated IGHV genes have a more aggressive disease course than do those with mutated IGHV genes.8,9 Other prognostic parameters include expression of ZAP-7010,11 and CD38,11 and biochemical parameters (eg, lactate dehydrogenase and β2-microglobulin).12 Finally, use of next-generation sequencing has identified novel gene mutations or deletions, including mutations or deletions in NOTCH1 and SF3B1,3,13,14 which might be associated

www.thelancet.com/oncology Published online May 13, 2016 http://dx.doi.org/10.1016/S1470-2045(16)30029-8

1

Articles

Research in context Evidence before this study Refining prognostication of chronic lymphocytic leukaemia has become essential, as therapeutic concepts are changing from chemoimmunotherapy to targeted drugs. Provision of a more accurate, comprehensive tool to predict the heterogeneous clinical course of this cancer would be relevant both for patient counselling and designing of future clinical trials. Between Jan 1, 2009, and Dec 31, 2010, the Cochrane Haematological Malignancies Group did a systematic database search of MEDLINE, Embase, and Central for prognostic factors in chronic lymphocytic leukaemia published between Jan 1, 1950, and Dec 31, 2010, evaluating at least one of the following new prognostic factors: del(17p), del(11q), trisomy 12, del(13q), del(6q), IGHV mutational status, ZAP-70, CD38, or TP53 mutational status. Their search identified 13 eligible phase 3 clinical trials. Added value of this study Our meta-analysis used the individual patient data of eight trials identified in the Cochrane systematic search and

with reduced survival. The abundance of prognostic markers, the insufficiency of comprehensive studies on their independent prognostic value, and the small number of proposed models showing an optimised combined use have, so far, delayed the translation of these biomarkers into general practice. To address this problem, as an international group of chronic lymphocytic leukaemia investigators, we did a comprehensive analysis using individual patient data available from prospective, controlled, randomised clinical trials with the aim to develop a prognostic index based on the most widely used clinical, biological, and genetic prognostic parameters in chronic lymphocytic leukaemia. The prognostic index will allow uniform reporting of patients across clinical trials with a classification that includes modern prognostic factors beyond clinical staging.

Methods Search strategy, selection criteria, and study population Between Jan 1, 2009, and Dec 31, 2010, the Cochrane Haematological Malignancies Group did a systematic search of MEDLINE, Embase, and Central databases for prospective, clinical phase 2 and 3 trials of chronic lymphocytic leukaemia, published between Jan 1, 1950, and Dec 31, 2010, which included at least one of the following new prognostic factors: del(17p), del(13q), del(6q), del(11q), trisomy 12, TP53 and IGHV mutational status, and ZAP-70 and CD38 expression. No language restrictions were used. From these criteria, they identified 13 eligible clinical trials and we contacted the principal investigators of these trials. Eight investigators agreed to include their individual patient data in the analysis (appendix p 3). Thus, we obtained the full analysis dataset from eight randomised, phase 3 clinical trials completed between 1997 and 2007 in France,15 Germany,6,16–18 Poland,19 the UK,20 and the USA21 2

two additional prospective cohorts to evaluate the currently known prognostic markers including novel gene mutations. The results allowed us to propose an easily applied prognostic model—the chronic lymphocytic leukaemia international prognostic index (CLL-IPI)—based on five widely available parameters defining four distinct groups of patients with very different prognoses. Implications of all the available evidence The CLL-IPI will allow uniform reporting of patients in clinical trials internationally with a classification that includes modern prognostic factors beyond clinical staging. It also prepares clinical staging of chronic lymphocytic leukaemia, for a potentially bright future of more targeted and more efficacious therapies, and will improve the precision of prognostic counselling of patients with chronic lymphocytic leukaemia regarding the implications of their disease.

(appendix p 4). All patients were treatment naive and, except in one trial,16 had an indication for treatment. The diagnosis of chronic lymphocytic leukaemia and response to therapy were assessed according to the National Cancer Institute Working Group guidelines.22 All eight trials were approved by the leading ethics committee, and written informed consent was obtained from patients according to the Declaration of Helsinki. We searched for large patient cohorts with complete cases regarding prognostic factors, and which were observed prospectively outside clinical trials to reflect real-life patients with chronic lymphocytic leukaemia. The individual patient data of two independent cohorts of newly diagnosed patients with chronic lymphocytic leukaemia were also included in the analysis as external-validation datasets. The MAYO cohort consisted of a consecutive series of newly diagnosed and prospectively observed patients with chronic lymphocytic leukaemia who were cared for at the Mayo Clinic (Rochester, MN, USA) and recruited between Jan 9, 2001, and Aug 20, 2014. Patient data of the SCAN cohort were collected from the SCALE Scandinavian populationbased case-control study,23 who were recruited between Oct 1, 1999, and Aug 30, 2002. For both cohorts, patients who had complete baseline data for all considered variables were selected for the analysis. Written informed consent was obtained for all patients in these two cohorts. We completed the CLL-IPI project in accordance with a data protection concept that was approved by the University Hospital of Cologne (Cologne, Germany).

Data analysis The data management teams of the principal investigators anonymised datasets of the individual patient data and electronically transferred these in code

www.thelancet.com/oncology Published online May 13, 2016 http://dx.doi.org/10.1016/S1470-2045(16)30029-8

Articles

to the project management team (JBa, NK, and MH) of The International CLL-IPI working group. We completed all development steps of the analysis on the full analysis dataset, including the multivariate analyses and the construction of the prognostic index. As a preparatory step before the multivariate analyses but after the univariate analysis, the full analysis dataset was randomly divided into training and internal-validation datasets. The random allocation of participant data (two-thirds to the training and a third to the internal-validation datasets) was done by a random sample procedure with SPSS (version 23) software, resulting in 2308 (66%) patients in the training and 1164 (34%) in the internal-validation datasets. We consecutively did both the multivariate analyses and the construction of the prognostic index for the training dataset and then both in the internalvalidation dataset, to confirm the analysis steps. Subsequently, we analysed the external-validation datasets, using data from the MAYO and the SCAN cohorts, to confirm the findings from the full analysis dataset. The main endpoint of statistical analyses was overall survival calculated from study entry to date of death for patients of the full analysis dataset (training and internal-validation datasets) and from diagnosis to date of death for the external-validation datasets. Furthermore, we calculated time to first treatment for watch-and-wait patients from the CLL1 trial,16 which was the only trial of the full analysis dataset that included patients who did not need treatment, and for patients from the external-validation cohorts. Time to first treatment was defined as the time between diagnosis and start of first treatment for chronic lymphocytic leukaemia for the CLL1 cohort16 and the externalvalidation datasets. Patients without a documented event (for overall survival was death; for time to first treatment was start of chronic lymphocytic leukaemia treatment) were censored at the date of last observation. 27 baseline markers, assessed in samples obtained before the start of first-line treatment, were considered as covariates for construction of the prognostic index. These covariates were clinical characteristics (age, sex, time between diagnosis and study entry, Rai1 and Binet2 clinical stages, B-symptoms, and the Eastern Cooperative Oncology Group [ECOG] performance status), laboratory values (haemoglobin concentration, and platelet, leucocyte, lymphocyte, and neutrophil counts), cytogenetic abnormalities according to fluorescence in-situ hybridisation (FISH; del[17p], del[11q], trisomy 12, del[13q], and del[6q]), IGHV mutational status, gene mutations (TP53, NOTCH1, and SF3B1), expressions of ZAP-70 and CD38, serum parameters (lactate dehydrogenase, β2-microglobulin, and thymidine kinase), and variables for the study and type of first-line treatment. For the external-validation datasets, only variables included in the final model were included in the analyses.

Statistical analysis We used the Kaplan-Meier method, including the log-rank test, for estimations and comparisons of overall survival. We calculated hazard ratios (HR) using Cox proportional-hazard regression analyses. All tests were two-sided, and statistical significance was defined as a p value less than 0·05. After having analysed missing values, with the application of Little’s missing completely at random (MCAR) test,24 imputation of missing data was not applied because these were not missing completely at random, a condition for the use of such methods. Therefore, we considered whether to use a complete-case analysis at each analysis step based on the full analysis dataset. For the external-validation dataset, we analysed missing values of the SCAN cohort using Little’s MCAR test and imputed these using linear regression. There were no missing values for the MAYO cohort. We categorised laboratory variables by published thresholds and quartiles, and included these categorised variables for further analyses if the corresponding continuous variables were significantly associated with overall survival in univariate analyses. We then included all factors significantly associated with overall survival in univariate analyses for the multivariate analysis. Additionally, in the multivariate analysis we included variables for the study and type of first-line treatment. In constructing a multivariable prognostic model, we took into account the extent of completeness of covariates regarding patients and datasets. For this purpose, variables were categorised in four variable categories I–IV (appendix pp 5–6). Applying backward-stepwise and forward-stepwise proportional-hazards Cox regressions, we completed the multivariate analysis in a four-step procedure starting with variables that were categorised as being completely available (less than 15% missing values; step I). After, we added variables that were nearly complete (>15% and ≤25% missing; step II), moderately complete (>25% and ≤35% missing; step III), and insufficiently available (>35% missing; step IV; appendix pp 6–7). To identify cases with an important missing variable, missing values were labelled (as –1), by indicator variables for the variables categorised as moderately (III) and insufficiently complete (IV). Thus, the cases with missing values were not dropped off by the complete-case analysis. (appendix pp 6–7). We constructed the prognostic index on the basis of the independent prognostic factors included in the final multivariable Cox model. We assigned a weighted risk score to each factor based on the regression parameters from the Cox regression analysis. The prognostic score was then defined as the sum of single-risk parameters. We identified different risk groups on the basis of the following criteria: significant differences in overall survival between risk groups, absence of heterogeneities concerning independent factors within each risk group (assessed by undertaking several subgroup analyses,

www.thelancet.com/oncology Published online May 13, 2016 http://dx.doi.org/10.1016/S1470-2045(16)30029-8

3

Articles

Patients (n=3472)

Median overall survival (months [95% CI])

5-year overall survival (95% CI)

10-year overall survival (95% CI)

Time between diagnosis and study entry

log-rank p value

Hazard ratio (95% CI)

p value

0·53

≤1 year

2108 (62%)

97·6 (89·5–104·0)

67·6% (65·5–69·7)

41·5% (38·5–44·6)

··

1·00 (reference)

¨

>1 year

1311 (38%)

90·7 (87·1–97·8)

68·6% (66·0–71·2)

38·1% (33·7–42·5)

··

1·03 (0·9–1·2)

0·53

Age

<0·0001

≤65 years

2395 (69%)

124·0 (106·6–119·7)

74·0% (72·1–75·8)

46·9% (43·8–50·0)

··

1·00 (reference)

>65 years

1077 (31%)

87·0 (63·8–74·7)

54·9% (51·8–58·0)

26·4% (22·5–30·3)

··

1·9 (1·7–2·1)

71·9% (69·0–74·7)

50·9% (46·6–55·1)

··

1·00 (reference)

66·4% (64·4–68·3)

35·8% (32·8–38·7)

··

1·4 (1·2–1·6)

78·5% (76·4–80·5)

54·5% (50·9–58·0)

··

1·00 (reference)

··

2·0 (1·8–2·3)

<0·0001

··

2·9 (2·2–3·8)

<0·0001

Sex

¨ <0·0001

<0·0001

Female

1045 (30%)

Male

2427 (70%)

124·0 (NE) 87·0 (83·6–90·9)

ECOG performance status

¨ <0·0001

<0·0001

0

1640 (63%)

144·7 (NE)

1

858 (33%)

77·8 (73·7–84 5)

60·3% (56·8–63·7)

31·7% (26·5–36·9)

2 or 3

101 (4%)

58·6 (36·6–86 8)

49·0% (37·7–60·3)

26·2% (13·0–39·4)

B-symptoms*

¨

<0·0001

No

1615 (69%)

79·7% (77·6–81·2)

57·2% (53·7–60·6)

··

1·00 (reference)

Yes

717 (31%)

137·5 (NE) 89·7 (82·3–97·9)

69·2% (65·6–72·7)

32·4% (26·6–38·2)

··

1·7 (1·5–2·0)

¨

A

992 (32%)

NR

83·7% (81·4–86·0)

63·1% (59·4–66·9)

··

1·00 (reference)

B

1260 (41%)

86·1 (83·1–91·5)

70·1% (67·4–72·7)

30·8% (25·9–35·8)

··

2·2 (1·9–2·5)

<0·0001

C

849 (27%)

68·9 (62·3–76·4)

54·9% (51·4–58·4)

25·9% (20·3–30·4)

··

3·0 (2·6–3·5)

<0·0001

<0·0001

<0·0001

Binet stage2

¨

<0·0001

Rai stage1 0

386 (14%)

NR

91·0% (87·2–93·2)

79·1% (73·6–83·5)

··

1·00 (reference)

I/II

1428 (50%)

98·5 (95·7–107·7)

71·5% (69·2–74·1)

41·5% (37·8–45·9)

··

3·5 (2·7–4·5)

<0·0001

III/VI

1025 (36%)

69·7 (63·3–76·4)

54·8% (51·7–58·1)

29·2% (20·5–30·0)

··

5·1 (3·9–6·7)

<0·0001

Type according to the hierarchical model

¨

<0·0001

del(17p)

191 (7%)

30·9 (27·3–36·6)

23·9% (17·2–30·6)

14·9% (8·6–21·2)

··

4·2 (3·5–5·1)

<0·0001

del(11q)†

499 (17%)

72·3 (67·6–78·8)

59·9% (55·3–64·5)

16·5% (9·7–23·3)

··

1·8 (1·5–2·1)

<0·0001

Trisomy 12‡

356 (12%)

72·5% (69·9–75·7)

37·5% (29·2–45·7)

··

1·1 (0·9–1·4)

0·17

Normal§

826 (29%)

125·5 (NE)

97·4 (86·8–110·6)

75·6% (72·5–78·6)

52·7% (48·1–57·3)

··

0·8 (0·7–0·9)

0·013

del(13q)¶

989 (35%)

112·7 (100·3–121·5)

72·9% (69·9–75·7)

45·1% (40·2–50·0)

··

1·00 (reference)

¨

<0·0001

IGHV mutational status Unmutated Mutated

1432 (60%) 947 (40%)

77·3 (75·1–82·2)

64·2% (61·6–66·8)

24·0% (20·1–27·9)

··

2·8 (2·4–3·2)

NR

82·8% (80·3–85·3)

65·1% (60·7–69·5)

··

1·00 (reference)

Mutated

1557 (96%)

86·3 (82·9–90·4)

68·0% (65·2–70·4)

29·7% (25·5–33·8)

··

1·00 (reference)

69 (4%)

57·2 (45·2–66·5)

43·5% (31·5–55·4)

21·8% (9·7–34·0)

··

1·9 (1·4–2·5)

76·1% (74·1–78·1)

48·6% (45·3–51·9)

··

1·00 (reference)

162 (8%)

84·1 (74·0–94·2)

69·9% (62·8–77·1)

21·3% (10·8–31·7)

··

1·6 (1·3–2·0)

1645 (85%)

121·5 (114·0–132·3)

77·7% (75·6–79·7)

50·5% (47·0–54·0)

··

1·00 (reference)

291 (15%)

77·3 (71·9–88·9)

64·9% (59·2–70·1)

29·4% (21·0–37·9)

··

1·7 (1·5–2·1)

Negative

1126 (57%)

94·6 (88·9–101·3)

70·0% (67·2–72·8)

32·0% (26·0–38·0)

··

1·00 (reference)

Positive

850 (43%)

70·2 (64·9–73·9)

57·3% (53·6–60·6)

18·9% (13·1–24·7)

··

1·6 (1·4–1·8)

Negative

628 (49%)

88·8 (78·9–96·1)

65·0% (61·2–68·8)

32·1% (25·3–38·8)

··

1·00 (reference)

Positive

665 (51%)

71·7 (68·1–76·8)

58·8% (55·0–62·6)

18·5% (13·5–23·6)

··

1·3 (1·2–1·6)

Mutated

1806 (92%)

118·1 (109·4–124·0)

Mutated

¨ <0·0001

<0·0001

SF3B1 mutational status Unmutated

¨ <0·0001

<0·0001

NOTCH1 mutational status Unmutated

¨

<0·0001

TP53 mutational status Unmutated

<0·0001

Expression of CD38

¨ <0·0001

<0·0001

Expression of ZAP-70

¨ <0·0001

<0·0001 ¨ <0·0001

(Table 1 continues on next page)

4

www.thelancet.com/oncology Published online May 13, 2016 http://dx.doi.org/10.1016/S1470-2045(16)30029-8

Articles

Patients (n=3472)

Median overall survival (months [95% CI])

5-year overall survival (95% CI)

10-year overall survival (95% CI)

log-rank p value

Hazard ratio (95% CI)

p value

(Continued from previous page) Serum lactate dehydrogenase (U/L)

<0·0001

≤250·0

1538 (50%)

124·3 (117·0–141·7)

77·8% (75·6–80·0)

51·1% (47·5–54·7)

··

1·00 (reference)

>250·0

1556 (50%)

80·1 (75·4–82·9)

61·7% (59·1–64·2)

29·4% (59·1–64·2)

··

1·9 (1·7–2·1)

Serum thymidine kinase (U/L)

¨ <0·0001

<0·0001

≤10·0

768 (45%)

NR

91·0% (88·9–93·1)

75·2% (71·0–79·4)

··

1·00 (reference)

>10·0

921 (54%)

93·6 (83·7–98·4)

70·3% (67·2–73·3)

36·3% (31·1–41·5)

··

3·6 (3·0–4·4)

Serum β2-microglobulin (mg/L)

¨ <0·0001

<0·0001

≤3·5

1683 (66%)

>3·5

869 (34%)

141·7 (NE) 65·9 (60·5–71·5)

79·6% (77·6–81·6)

56·6% (53·1–60·1)

··

1·00 (reference)

54·0% (50·5–57·6)

22·1% (17·7–26·4)

··

2·7 (2·4–3·0)

¨ <0·0001

Data are n (%) unless otherwise stated. Except for age and sex categories, some data are missing in the covariate categories. Overall survival was calculated from study entry to date of death. NR=not reached. NE=not evaluable. ECOG=Eastern Cooperative Oncology Group. *Presence of any constitutional symptom according to the International Workshop on Chronic Lymphocytic Leukemia guidelines.25 †Excluding del(17p). ‡Excluding del(17p) and del(11q). §No abnormalities according to the Döhner hierarchical model including del(17p), del(11q), trisomy 12, and del(13q). ¶Excluding del(17p), del(11q), and trisomy 12.

Table 1: Patient characteristics of the full analysis dataset and results of the univariate analyses for overall survival

looking at the different patient subgroups in each risk group, and assessing whether patients were comparable with regard to overall survival), and equal spacing between Kaplan-Meier survival curves (visual assessment). To ensure clinical applicability, the composition of the final risk groups was intended to create homogeneity regarding treatment recommendations. To assess the discriminatory value, we calculated C-statistics (c), whereby a c value of 1·0 indicates perfect discrimination and a c value of 0·5 is equivalent to chance. We also did subgroup analyses to further explore the validity and reproducibility of the final prognostic model by applying the Kaplan-Meier method for overall survival in the subgroups provided by the independent factors of the prognostic index, as well as sex (by calculating the CLL-IPI risk groups separately for men and women). We used software SPSS (version 23) and SAS (version 9.4) for statistical analyses.

Role of funding source The funder of the study had no role in the study design, data collection, data analysis, data interpretation, or writing of the report. The project management team (JBa, NK, MH) had full access to all raw data in the study. The corresponding author had final responsibility to submit for publication.

Results Data from 3725 patients were exported from the eight trial datasets, of which 3472 unique eligible patients were included in the full analysis dataset (appendix p 4). Median age was 61 years (range 27–86) and 1542 (44%) deaths from any cause were observed after a median observation time of 79·9 months

(IQR 79·9–101·4; median overall survival was 95·3 months [95% CI 89·7–98·5]). Patient characteristics of the full analysis dataset are in table 1 and the appendix (pp 8–11). We completed the univariate analyses based on the full analysis dataset, and all 27 baseline factors were significantly associated with overall survival, except for trisomy 12 and the time between diagnosis and study entry (table 1; appendix pp 8–11). After the random sample procedure, the training dataset contained 2308 (66%) patients and the internalvalidation dataset contained 1164 (34%) patients. Both datasets had similar baseline characteristics and overall survival (appendix pp 12–14). Descriptions of the final models from multivariate analysis steps I–IV are in the appendix (pp 15–18). After the fourth step of the multivariate analysis (ie, considering variables categorised as insufficiently available; step IV; appendix pp 6–7), we identified five independent factors for overall survival: TP53 status (no abnormalities vs del[17p] or TP53 mutation or both), IGHV mutational status (mutated vs unmutated), β2-microglobulin concentration (≤3·5 mg/L vs >3·5 mg/L), clinical stage (Binet A or Rai 0 vs Binet B–C or Rai I–IV), and age (≤65 years vs >65 years). The presence of either a del(17p) or TP53 mutation was concluded to be of prognostic importance because no significant association existed between patients who exhibited only one characteristic (TP53 mutation present vs only del[17p] present; p=0·468), or between patients with both characteristics and those with just one characteristic (both del[17p] and TP53 mutation present vs only del[17p] present; p=0·590), whereas the comparison between patients without either abnormality and patients with at least one characteristic was significantly

www.thelancet.com/oncology Published online May 13, 2016 http://dx.doi.org/10.1016/S1470-2045(16)30029-8

5

Articles

Overall survival events Training dataset (n=1214)

Adverse factor

Regression Hazard coefficient ratio (95% CI)

p value

Assigned risk score

462 (38%)

TP53 status

¨

Deleted or mutated

1·434

4·2 (3·2–5·5)

<0·0001

4

IGHV mutational status

¨

Unmutated

0·950

2·6 (2·1–3·2)

<0·0001

2

β2-microglobulin concentration

¨

>3·5 mg/L

0·678

2·0 (1·6–2·4)

<0·0001

2

Clinical stage

¨

Rai I–VI or Binet B–C

0·464

1·6 (1·3–1·9)

<0·0001

1

Age

¨

>65 years

0·555

1·7 (1·4–2·1)

<0·0001

1

Internal-validation dataset (n=585)

243 (42%)

TP53 status

¨

Deleted or mutated

1·181

3·3 (2·4–4·5)

<0·0001

¨

IGHV mutational status

¨

Unmutated

0·724

2·1 (1·5–2·8)

<0·0001

¨

β2-microglobulin concentration

¨

>3·5 mg/L

0·461

1·6 (1·2–2·1)

0·00112

¨

Clinical stage

¨

Rai I–VI or Binet B–C

0·680

1·9 (1·4–2·8)

0·00012

¨

>65 years

0·527

1·6 (1·3–2·2)

0·00014

¨

Age MAYO cohort (n=838)*

144 (17%)

TP53 status†

¨

Deleted

1·384

4·0 (2·4–6·7)

<0·0001

¨

IGHV mutational status

¨

Unmutated

0·771

2·2 (1·5–3·1)

<0·0001

¨

β2-microglobulin concentration

¨

>3·5 mg/L

0·771

2·2 (1·5–3·1)

<0·0001

¨

Clinical stage

¨

Rai I–VI

0·662

1·9 (1·4–2·8)

<0·0001

¨

Age

¨

>65 years

0·796

2·2 (1·6–2·2)

<0·0001

¨

SCAN cohort (n=416)*

215 (52%)

TP53 status

¨

Deleted or mutated

1·278

3·6 (2·2–6·0)

<0·0001

¨

IGHV mutational status

¨

Unmutated

1·347

3·8 (2·8–5·2)

<0·0001

¨

β2-microglobulin concentration

>3·5 mg/L

0·181

1·6 (1·1–2·3)

Clinical stage

Binet B–C

0·681

1·9 (1·4–2·7)

<0·0001

¨

Age

>65 years

0·922

2·5 (1·9–3·3)

<0·0001

¨

0·00961

¨

Overall survival was calculated from study entry for the full analysis dataset, and from diagnosis for the external-validation datasets, to date of death. Clinical staging was done according to Rai1 or Binet2 clinical staging systems. Number of patients in the training and internal-validation datasets are only those with complete-case analysis. *MAYO and SCAN cohorts were the external-validation datasets. †Data for TP53 mutational status not available.

Table 2: Final multivariable model (training, internal-validation, and external-validation datasets)

different in the multivariate analysis (p<0·0001). These findings led to the composite factor TP53 status and the final five-factor multivariable model. The final model was based on 1214 (53%) of 2308 patients from the 6

training dataset who had complete data available for analysis and included TP53 status, IGHV mutational status, β2-microglobulin concentration, clinical stage, and age (table 2). These factors were further confirmed to be independent based on 585 (50%) of 1164 patients from the internal-validation dataset for whom complete data were available (table 2). To evaluate the potential effect of the type of first-line treatment again, the variable for study treatment was added back to the final multivariable model. This variable did not provide independent prognostic information based on the training (p=0·847) or internalvalidation datasets (p=0·335; data not shown). Other than TP53 disruption, recurrent genetic abnormalities (eg, NOTCH1 and SF3B1) did not show independent prognostic information (appendix p 18). The final multivariable model constituted the basis for construction of the CLL-IPI (table 2). On the basis of the regression parameters and subgroup analyses (concerning the independent factors with regard to overall survival), we defined the following criteria to establish individual weighted risk scores of the five independent factors. In the training dataset, the factors of age and clinical stage should receive identical scores due to similar regression coefficients, accordingly, these two factors should receive the lowest score; the weighted risk score of IGHV mutational status should be at least double that of age or clinical stage; the weighted risk score of TP53 status should be at least triple the weighted risk score of age and clinical stage, and should be double the weighted risk score of IGHV mutational status (table 2). Thus, weighted risk scores of 1 were assigned to age and clinical stage, and a score of 2 was assigned to IGHV mutational status (table 2). As per subgroup analyses, we assigned the β2-microglobulin concentrations a risk score of 2, because patients with elevated β2-microglobulin concentrations and unmutated IGHV behaved similarly with regard to overall survival in our dataset (data not shown). Accordingly, a risk score of 4 was assigned to TP53 status. The final grading of the CLL-IPI factors is shown in table 2, resulting in total risk scores between 0 and 10 (appendix pp 19–20). We derived the CLL-IPI risk categories on the basis of the training dataset (1214 patients) and segregated patients in this dataset into four risk categories: low, intermediate, high, and very high risk (table 3). Significantly different 5-year overall survival was observed for these risk groups: 93·2% (95% CI 90·5–96·0) for those at low risk, 79·3% (75·5–83·2) at intermediate, 63·3% (57·9–68·8) at high, and 23·3% (12·5–34·1) for those at very high risk (log-rank test across all four groups p<0·0001; table 3, figure 1A). With respect to discrimination, the C-statistic of this final multivariable model was 0·723 (95% CI 0·684–0·752). The proposed CLL-IPI was fully confirmed on the basis of the internal-validation dataset

www.thelancet.com/oncology Published online May 13, 2016 http://dx.doi.org/10.1016/S1470-2045(16)30029-8

Articles

CLL-IPI risk score Training dataset

Patients

Median overall survival (months [95% CI])

5-year overall survival (95% CI)

10-year overall survival (95% CI)

Comparisons

Hazard ratio (95% CI)

1214

Low

0–1

341 (28%)

NR

93·2% (90·5–96·0)

79·0% (71·8–86·3)

¨

¨

Intermediate

2–3

474 (39%)

105 (96–119)

79·3% (75·5–83·2)

39·2% (31·0–47·4)

vs 0–1

3·5 (2·5–4·8)

High

4–6

337 (28%)

75 (68–82)

63·3% (57·9–68·8)

21·9% (14·2–29·6)

Very high

7–10

62 (5%)

29 (18–40)

23·3% (12·5–34·1)

NR

90·7% (86·4–95·1)

86·5% (80·5–92·4)

¨

¨

79·8% (73·9–85·8)

40·1% (29·3–50·9)

vs 0–1

4·6 (2·8–7·4)

16·1% (6·7–25·4)

Internal-validation dataset

3·5%* (NE)

vs 2–3

1·9 (1·5–2·3)

vs 4–6

3·6 (2·6–4·8)

585

Low

0–1

186 (32%)

Intermediate

2–3

200 (34%)

104 (84–123)

High

4–6

147 (25%)

63 (51–73)

52·8% (44·5–61·1)

Very high

7–10

52 (9%)

31 (20–39)

18·6% (7·5–29·7)

MAYO cohort‡

0%† (NE)

vs 2–3

2·2 (1·6–3·0)

vs 4–6

2·6 (1·8–3·7)

838

Low

0–1

390 (47%)

NR

96·6% (94·5–98·8)

81·9% (74·4–89·5)

¨

¨

Intermediate

2–3

272 (33%)

125 (NE)

92·0% (88·3–95·7)

55·4% (41·4–69·4)

vs 0–1

2·9 (1·8–4·6)

High

4–6

149 (18%)

82 (72–101)

68·5% (60·0–77·2)

22·5% (5·9–39·1)

vs 2–3

3·2 (2·1–4·7)

Very high

7–10

27 (3%)

43 (NE)

21·2% (NE)

21·2§ (NE)

vs 4–6

3·2 (1·7–5·6)

SCAN cohort‡

416

Low

0–1

242 (58%)

92·1% (88·8–95·5)

79·0 (73·6–84·3)

¨

¨

Intermediate

2–3

104 (25%)

182 (NE) 93 (85–114)

75·0% (66·7–83·3)

36·1 (26·6–45·5)

vs 0–1

4·5 (3·3–6·2)

High

4–6

56 (14%)

78 (62–92)

64·3% (84·7–76·8)

18·6 (8·2–29·0)

Very high

7–10

14 (3%)

32 (12–66)

28·6% (5·0–52·2)

7·1%¶ (NE)

vs 2–3

1·7 (1·2–2·3)

vs 4–6

2·7 (1·5–5·1)

Data are n (%), unless otherwise stated. Overall survival was calculated from study entry for the full analysis set, and from diagnosis for the external-validation datasets, to date of death. Number of patients in the training and internal-validation datasets included are only those with complete-case analysis. CLL-IPI=chronic lymphocytic leukaemia international prognostic index. NR=not reached. NE=not evaluable. *At time of last observation (month 93). †At month 98. ‡MAYO and SCAN cohorts are the external-validation datasets. §At month 89. ¶At month 101.

Table 3: Survival data of CLL-IPI risk groups in the training, internal-validation, and external-validation datasets

for 5-year overall survival (log-rank test across all four groups p<0·0001; c=0·773 [95% CI 0·730–0·819]; table 3, figure 1B). We validated and confirmed the CLL-IPI using the external-validation datasets (two independent cohorts of newly diagnosed patients with chronic lymphocytic leukaemia). The MAYO cohort contained 838 patients with a median observation of 63·2 months (IQR 30·2–91·8), and the SCAN cohort contained 416 patients with a median observation time of 151·0 months (124·8–163·0). Baseline characteristics and results from univariate analyses for overall survival of the external-validation dataset are in the appendix (pp 21–24). As of November, 2014, 512 (61%) patients in the MAYO cohort had not received any treatment, and as of May, 2015, 192 (46%) patients in the SCAN cohort had not received any treatment; these patients were managed according to the watch-and-wait approach. In the MAYO cohort, all CLL-IPI risk factors were shown to be independent prognostic factors (table 2). Due to missing data for TP53 mutations, we used del(17p) as the sole marker of TP53 status in this cohort. The 5-year overall survival of the CLL-IPI risk groups in the MAYO cohort differed significantly (p<0·0001 [log-rank test across all four groups]; c=0·788

[95% CI 0·730–0·83]; table 3, figure 1C). β2-microglobulin concentration was not available in 17 (4%) of 416 patients in the SCAN cohort. Since these values were missing completely at random according to Little’s MCAR test (p=0·224), we applied missing data imputation using linear regression. In the SCAN cohort, the five-factor multivariable model was confirmed (table 2) and the 5-year overall survival differed significantly among all risk groups (p<0·0001 [log-rank test across all four groups]; c=0·731 [95% CI 0·674–0·810]; table 3, figure 1D). All patients from the training and internal-validation datasets with complete data for risk factors were pooled and constituted the score population (1799 [52%] of 3472 patients). Subgroup analyses based on this subpopulation provided evidence for both validity and reproducibility of the prognostic model regarding clinical stages (appendix pp 25–27), cytogenetic subgroups (appendix pp 28–29), IGHV mutational status (appendix pp 30–31), β2-microglobulin concentrations (appendix pp 30–31), as well as age and sex (appendix pp 32–33). Consistent results were obtained for the MAYO cohort (appendix pp 34–39). Subgroup analyses were not undertaken for the SCAN cohort due to its small sample size.

www.thelancet.com/oncology Published online May 13, 2016 http://dx.doi.org/10.1016/S1470-2045(16)30029-8

7

Articles

A

Training dataset Low risk Intermediate risk High risk Very high risk

100

B

Internal-validation dataset

Overall survival (%)

80

60

40

20

p<0·0001

p<0·0001

0 0

12

24

36

48

60

72

84 96 108 120 132 144 156

0

12

24

36

Time from study entry (months)

Number at risk Low risk 341 339 331 320 279 270 224 169 118 81 Intermediate risk 474 452 441 415 352 312 232 143 83 52 High risk 337 314 284 256 205 178 120 69 40 19 5 3 0 ·· Very high risk 62 46 31 25 16 13

C

48

60

72

84 96 108 120 132 144 156

Time from study entry (months) 40 27 12 ··

20 13 4 ··

8 5 1 ··

0 1 0 ··

186 181 173 168 152 146 125 84 200 191 180 168 137 125 88 55 147 130 117 98 84 69 52 24 52 43 30 20 14 8 2 1

D

MAYO cohort

56 35 17 1

39 22 13 0

26 10 6 ··

11 5 2 ··

3 1 1 ··

0 0 0 ··

SCAN cohort

100

Overall survival (%)

80

60

40

20

p<0·0001

p<0·0001

0 0

12

24

36

48

60

72

84 96 108 120 132 144 156

Time from diagnosis (months)

Number at risk Low risk 390 338 316 288 259 226 188 151 105 64 Intermediate risk 272 247 224 198 178 146 111 81 46 27 High risk 149 127 110 100 82 64 44 28 14 9 5 1 1 1 0 ·· Very high risk 27 24 18 12

0

12

24

36

48

60

72

84 96 108 120 132 144 156

Time from diagnosis (months) 34 12 3 ··

19 1 0 ··

7 1 ·· ··

3 0 ·· ··

242 238 237 233 228 223 220 210 192 168 159 144 115 80 104 103 99 95 89 78 71 61 46 42 34 25 17 12 56 55 51 47 39 36 32 25 16 11 10 6 3 2 14 10 8 7 7 4 3 1 1 0 ·· ·· ·· ··

Figure 1: Overall survival according to the CLL-IPI risk groups The full analysis dataset is comprised of (A) the training dataset of 1214 patients and (B) the internal-validation dataset of 585 patients. The external-validation datasets are (C) the MAYO cohort of 838 patients and (D) the SCAN cohort of 416 patients. The p values are the log-rank tests values across all four groups.

Additionally, we analysed time to first treatment in watch-and-wait patients from the CLL1 trial16 (403 [22%] of 1799 in the score population; 57% of 710 in the CLL1 trial). In this group, significant differences were recorded between the four risk groups (figure 2; appendix pp 40–41). Time to first treatment was also analysed for the 8

external-validation datasets. Results were consistent with that of the watch-and-wait patients in the score population (figure 2; appendix pp 40–41). We also explored the effect of Rai1 versus Binet2 staging systems on the CLL-IPI and found that 59 (5%) of 1309 patients (of the score population for whom variables

www.thelancet.com/oncology Published online May 13, 2016 http://dx.doi.org/10.1016/S1470-2045(16)30029-8

Articles

A

B

Watch-and-wait patients16

MAYO cohort

Low risk Intermediate risk High risk Very high risk

100

Time to first treatment (%)

80

60

40

20

p<0·0001

p<0·0001

0 0

12

24

36

48

60

72

84 96 108 120 132 144 156

Number at risk Low risk 264 261 253 238 222 213 182 151 105 74 Intermediate risk 110 108 93 73 54 42 31 20 12 8 7 5 3 2 2 High risk 25 23 17 13 10 4 4 2 1 1 1 1 1 1 ·· Very high risk

C

0

12

24

36

48

60

72

84 96 108 120 132 144 156

Time from diagnosis (months) 44 4 2 ··

21 4 1 ··

9 2 ·· ··

2 ·· ·· ··

390 324 291 257 219 183 155 117 79 272 174 140 110 88 66 41 23 11 149 77 59 41 32 19 10 4 2 27 7 3 2 1 0 ·· ·· ··

50 3 2 ··

25 1 1 ··

13 0 0 ··

4 ·· ·· ··

2 ·· ·· ··

SCAN cohort

100

Time to first treatment (%)

80

60

40

20

p<0·0001 0 0

12

24

36

48

60

72

84 96 108 120 132 144 156

Time from diagnosis (months)

Number at risk Low risk 242 212 202 195 190 173 168 155 141 118 111 101 81 8 6 5 Intermediate risk 104 70 45 38 28 23 21 18 12 10 6 5 4 4 4 3 2 2 2 2 High risk 56 19 10 4 3 2 1 1 1 0 ·· ·· ·· ·· ·· Very high risk 14

54 3 1 ··

Figure 2: Time to first chronic lymphocytic leukaemia treatment according to the CLL-IPI risk groups (A) Watch-and-wait group in the full analysis dataset (CLL1 trial;16 403 patients). The external-validation datasets are (B) the MAYO cohort of 838 patients and (C) the SCAN cohort of 416 patients. The p values are the log-rank test values across all four groups.

for both Rai and Binet clinical staging were available together) from the full analysis dataset and 65 (8%) of 817 patients (for whom variables for both Rai and Binet clinical staging were available together) from the MAYO cohort would be switched to a different risk group (appendix p 42). Nevertheless, overall survival at 5 years

of the CLL-IPI risk groups was nearly the same with use of both the Rai or Binet clinical staging systems (appendix p 43). We also assessed the age cutoff used for the CLL-IPI and the performance of the CLL-IPI. Appropriate performance of the CLL-IPI in patients aged 70 years or

www.thelancet.com/oncology Published online May 13, 2016 http://dx.doi.org/10.1016/S1470-2045(16)30029-8

9

Articles

older for 5-year overall survival ranged from 21·7% to 88·6%; appendix pp 32–33). Additionally, multivariate analyses in younger and older patients confirmed the significance for all CLL-IPI risk factors (data not shown).

Discussion To our knowledge, our meta-analysis represents the largest treatment-naive population of patients with chronic lymphocytic leukaemia analysed for prognostic factors. We used individual patient data from eight randomised, phase 3 clinical trials to validate the currently known prognostic markers, including novel gene mutations. The results allowed us to propose an easily applied prognostic model—the CLL International Prognostic Index (CLL-IPI)—based on five widely available parameters defining four distinct groups of patients with significantly different prognoses. The Rai1 and Binet2 clinical staging systems were developed before current biological and genetic variables became available, and have formed the basis of clinical management for patients with chronic lymphocytic leukaemia for almost 40 years.25 Automated, more frequent blood counts have resulted in earlier diagnosis, and most patients are now diagnosed with early stage disease (Rai1 0; Binet2 A). Additionally, therapeutic advances in the past 10 years have greatly improved the outcome of patients with chronic lymphocytic leukaemia, partly abrogating the previously close relationship between clinical stage and overall survival.26,27 These advances have created a need for additional prognostic parameters to complement clinical staging. Although several new prognostic markers have been identified, no standard approach exists for integrating the available parameters in a unified prognostic system to be reported in clinical trials. The prognostic system proposed by Wierda and colleagues28 extended clinical staging by using biological variables, but did not incorporate major genetic factors and did not accurately predict the outcome of individual patients. The score developed by Pflug and colleagues29 is similar to our CLL-IPI and showed the potential of using a weighted combination of clinical, biological, and genetic markers for accurate prognostication. However, the use of this score29 seemed limited by the inclusion of serum thymidine kinase, because this factor is not routinely available in some countries.30 Rossi and colleagues31 created a score that exclusively relied on novel gene mutations and cytogenetics without considering clinical characteristics. To overcome these limitations, we pooled the data of 3472 treatment-naive patients who participated in eight international phase 3 clinical trials from five countries to develop a comprehensive prognostic model for chronic lymphocytic leukaemia by using a training-validation procedure. The analysis incorporated the most important prognostic markers currently known for chronic lymphocytic leukaemia, including novel gene 10

mutations such as NOTCH1 and SF3B1.13,14,32 Missing data also presented a challenge for our analysis because the requirements for missing data imputations were not fulfilled. Therefore, a complete-case analysis had to be consistently done. We took into account the extent of data completeness by applying a multistep procedure and using a training-validation approach. To further explore the external applicability, the CLL-IPI was externally validated in two independent, prospective cohorts of 1254 newly diagnosed patients from the USA and Europe. Finally, we developed a prognostic model based on five widely available parameters: TP53 aberrations (comprising del[17p] and TP53 mutation), IGHV mutational status, β2-microglobulin concentration, clinical stage, and age. The resulting CLL-IPI combines genetic and biological characteristics with traditional risk factors in an easily applied prognostic index for chronic lymphocytic leukaemia. This model will be available on the web and as a mobile app (Calculate by QxMD). The identified markers are of independent prognostic value and therefore allow for elimination of some potentially redundant prognostic tests (eg, for ZAP-70 and CD38). If broadly accepted, our CLL-IPI could help to focus resources on the prognostic tests of greatest value. Along these lines, all parameters of the CLL-IPI are now widely available and generally accepted, which should enable the broad use of the CLL-IPI at least in health-care systems with access to molecular testing (eg, FISH or IGHV) in chronic lymphocytic leukaemia. Furthermore, the modular composition of the CLL-IPI will facilitate the future integration of new markers with independent prognostic value. The type of first-line treatment was not identified as an independent factor in our analysis. This result might be caused by the fact that therapy has progressed during the period of our investigation and has improved the overall survival of all risk groups of the CLL-IPI. Our CLL-IPI defines four groups at low, intermediate, high, and very high risk, providing additional prognostic information regarding overall survival compared with conventional clinical staging. Moreover, the four risk groups could be reproduced within each Rai1 and Binet2 clinical stage criterion with a high replication to the overall model. Furthermore, our analysis confirms that not all patients with a del(17p) are equal because the CLL-IPI was able to distinguish between patients with del(17p) who were high risk and very high risk. Importantly, the CLL-IPI has the potential to support clinical patient management. In view of the very good outcome of the patients at low risk for 5-year overall survival, a watch-and-wait approach seems appropriate. Patients at intermediate risk might not be treated unless the disease presents with severe symptoms. Patients at high risk should start treatment except for those who are fully asymptomatic. Finally, those deemed to be at very high risk should be offered treatment in experimental

www.thelancet.com/oncology Published online May 13, 2016 http://dx.doi.org/10.1016/S1470-2045(16)30029-8

Articles

protocols, incorporating novel drugs (eg, kinase inhibitors or BCL-2 antagonists), but should avoid chemoimmunotherapy since this group is characterised by TP53 aberrations and does not appear to benefit from conventional chemoimmunotherapies. However, as these recommendations have not been validated independently, the criteria of disease activity according to the updated guidelines of the International Workshop on Chronic Lymphocytic Leukemia should continue to be used to define the start of therapy.25 We are aware of some limitations of our study. At the time of analysis, phase 3 trials of novel oral inhibitors (ie, idelalisib, ibrutinib, or venetoclax) did not have sufficiently long follow-up to be included. Therefore, a comprehensive re-evaluation will be completed based on the composite character of the CLL-IPI as soon as phase 3 trials testing these inhibitors reach sufficient maturity. However, factors such as TP53 abnormalities might retain their prognostic importance, because the 3-year follow-up results of trials testing ibrutinib showed that patients with del(17p) still had a comparatively poor outcome.33 Of note is that the CLL-IPI accurately described time to first treatment in the external-validation datasets, an endpoint that is not affected by the introduction of these novel compounds. Therefore, the CLL-IPI seems to be a necessary, important step towards harmonising prognostication in chronic lymphocytic leukaemia and allowing comparisons between clinical trials, which is of particular relevance in an era of dynamic therapeutic improvements. Finally, and most importantly, the CLL-IPI is able to identify patients with chronic lymphocytic leukaemia who usually do not benefit from conventional chemoimmunotherapies, in whom the testing of novel, non-cytotoxic treatment options is justified and urgently needed. Our model was developed based on a dataset where the median age was lower than the general median age of patients with chronic lymphocytic leukaemia at diagnosis (61 years vs 72 years34). However, the age cutoff used for the CLL-IPI is higher than the median (65 years), and appropriate performance of the CLL-IPI was shown for patients 70 years or older. Additionally, multivariate analyses in younger and older patients confirmed the significance for all CLL-IPI risk factors (data not shown). Of course, validations by use of datasets that also include elderly patients with coexisting medical conditions are warranted and already being planned. Another important strength of our index is the integration of clinical staging, which will allow a comparison with historical study results. However, the CLL-IPI might vary slightly on the basis of the clinical staging system used (Rai1 vs Binet2), and might result in slightly different risk groups, because the Rai1 and Binet2 systems are not identical. However, the overall survival of the CLL-IPI risk groups was shown being comparable with the use of both these clinical staging systems.

Together, the results of our meta-analysis show that the CLL-IPI might improve physicians’ ability to counsel patients regarding potential future implications of their diagnosis. Moreover, the CLL-IPI should now be further tested in prospective clinical trials, particularly with novel, more targeted therapies for chronic lymphocytic leukaemia. Contributors JBa, NK, and MH wrote the first draft of the manuscript. JBa did the statistical analyses. JBa, NK, KB, MAB, JBy, KGC, HD, BFE, ME, CG, MG, SL, DN, DO, TR, RR, TDS, SS, and MH did the research and approved the manuscript and the concept of the paper. Members of The International CLL-IPI working group are listed in the appendix. Declaration of interests KB has received grants from the German Federal Ministry of Education and Research. BFE has received honorarium for advisory boards from Janssen, Gilead, Mundipharma, and GlaxoSmithKline (GSK); has received honorarium from Roche, Mundipharma, GSK, Gilead, and Janssen; and has received scientific grants from Roche and Mundipharma. CG has received personal fees from Roche, Janssen, Gilead, Celgene, Novartis, and AbbVie, outside the submitted work. MG has received grants from the National Cancer Institute during the conduct of the study; and has received grants, personal fees, and non-financial support from Pharmacyclics and Acerta, outside the submitted work. TDS has received grants from Genentech, Pharmacyclics Janssen, GSK, Celgene, Cephalon, Hospira, and Polyphenon E International, outside the submitted work. SS has received grants, personal fees, and other from AbbVie, Amgen, Boehringer Ingelheim, Celgene, Genentech, Genzyme, Gilead, GSK, Janssen, Mundipharma, Novartis, Pharmacyclics, Hoffmann La-Roche, and Sanofi, during the conduct of the study. JBa, NK, MAB, JBy, KGC, HD, ME, SL, DN, DO, TR, RR, and MH declare no competing interests. Members of The International CLL-IPI working group are listed in the appendix. Acknowledgments JBa, NK, and MH received funding from the José Carreras Leukaemia Foundation; MH received funding from the DFG (German Research Council; clinical research unit 286); ME was supported by the Arbib Charitable Fund. References 1 Rai KR, Sawitsky A, Cronkite EP, Chanana AD, Levy RN, Pasternack BS. Clinical staging of chronic lymphocytic leukemia. Blood 1975; 46: 219–34. 2 Binet JL, Auquier A, Dighiero G, et al. A new prognostic classification of chronic lymphocytic leukemia derived from a multivariate survival analysis. Cancer 1981; 48: 198–206. 3 Landau DA, Tausch E, Taylor-Weiner AN, et al. Mutations driving CLL and their evolution in progression and relapse. Nature 2015; 526: 525–30. 4 Puente XS, Bea S, Valdes-Mas R, et al. Non-coding recurrent mutations in chronic lymphocytic leukaemia. Nature 2015; 526: 519–24. 5 Döhner H, Stilgenbauer S, Benner A, et al. Genomic aberrations and survival in chronic lymphocytic leukemia. N Engl J Med 2000; 343: 1910–16. 6 Hallek M, Fischer K, Fingerle-Rowson G, et al, on behalf of an international group of investigators, the German Chronic Lymphocytic Leukaemia Study Group. Addition of rituximab to fludarabine and cyclophosphamide in patients with chronic lymphocytic leukaemia: a randomised, open-label, phase 3 trial. Lancet 2010; 376: 1164–74. 7 Zenz T, Eichhorst B, Busch R, et al. TP53 mutation and survival in chronic lymphocytic leukemia. J Clin Oncol 2010; 28: 4473–79. 8 Damle RN, Wasil T, Fais F, et al. Ig V gene mutation status and CD38 expression as novel prognostic indicators in chronic lymphocytic leukemia. Blood 1999; 94: 1840–47. 9 Hamblin TJ, Davis Z, Gardiner A, Oscier DG, Stevenson FK. Unmutated Ig V-H genes are associated with a more aggressive form of chronic lymphocytic leukemia. Blood 1999; 94: 1848–54. 10 Crespo M, Bosch F, Villamor N, et al. ZAP-70 expression as a surrogate for immunoglobulin-variable-region mutations in chronic lymphocytic leukemia. N Engl J Med 2003; 348: 1764–75.

www.thelancet.com/oncology Published online May 13, 2016 http://dx.doi.org/10.1016/S1470-2045(16)30029-8

11

Articles

11

12

13

14

15

16

17

18

19

20

12

Rassenti LZ, Jain S, Keating MJ, et al. Relative value of ZAP-70, CD38, and immunoglobulin mutation status in predicting aggressive disease in chronic lymphocytic leukemia. Blood 2008; 112: 1923–30. Hallek M, Wanders L, Ostwald M, et al. Serum β₂-microglobulin and serum thymidine kinase are independent predictors of progression-free survival in chronic lymphocytic leukemia and immunocytoma. Leuk Lymphoma 1996; 22: 439–47. Puente XS, Pinyol M, Quesada V, et al. Whole-genome sequencing identifies recurrent mutations in chronic lymphocytic leukaemia. Nature 2011; 475: 101–05. Quesada V, Conde L, Villamor N, et al. Exome sequencing identifies recurrent mutations of the splicing factor SF3B1 gene in chronic lymphocytic leukemia. Nat Genet 2012; 44: 47–52. Lepretre S, Aurran T, Mahe B, et al. Excess mortality after treatment with fludarabine and cyclophosphamide in combination with alemtuzumab in previously untreated patients with chronic lymphocytic leukemia in a randomized phase 3 trial. Blood 2012; 119: 5104–10. Bergmann MA, Busch R, Eichhorst B, et al. Overall survival in early stage chronic lymphocytic leukemia patients with treatment indication due to disease progression: follow-up data of The CLL1 trial of the German CLL Study Group (GCLLSG). Blood 2013; 122: 4127. Eichhorst BF, Busch R, Hopfinger G, et al. Fludarabine plus cyclophosphamide versus fludarabine alone in first-line therapy of younger patients with chronic lymphocytic leukemia. Blood 2006; 107: 885–91. Eichhorst BF, Busch R, Stilgenbauer S, et al. First-line therapy with fludarabine compared with chlorambucil does not result in a major benefit for elderly patients with advanced chronic lymphocytic leukemia. Blood 2009; 114: 3382–91. Robak T, Jamroziak K, Gora-Tybor J, et al. Comparison of cladribine plus cyclophosphamide with fludarabine plus cyclophosphamide as first-line therapy for chronic lymphocytic leukemia: a phase III randomized study by the Polish Adult Leukemia Group (PALG-CLL3 Study). J Clin Oncol 2010; 28: 1863–69. Catovsky D, Richards S, Matutes E, et al, on behalf of the UK National Cancer Research Institute (NCRI) Haematological Oncology Clinical Studies Group; NCRI Chronic Lymphocytic Leukaemia Working Group. Assessment of fludarabine plus cyclophosphamide for patients with chronic lymphocytic leukaemia (the LRF CLL4 Trial): a randomised controlled trial. Lancet 2007; 370: 230–39.

21

22

23

24 25

26

27

28

29

30 31

32

33

34

Flinn IW, Neuberg DS, Grever MR, et al. Phase III trial of fludarabine plus cyclophosphamide compared with fludarabine for patients with previously untreated chronic lymphocytic leukemia: US Intergroup Trial E2997. J Clin Oncol 2007; 25: 793–98. Cheson BD, Bennett JM, Grever M, et al. National Cancer Institute-sponsored Working Group guidelines for chronic lymphocytic leukemia: revised guidelines for diagnosis and treatment. Blood 1996; 87: 4990–97. Mansouri L, Cahill N, Gunnarsson R, et al. NOTCH1 and SF3B1 mutations can be added to the hierarchical prognostic classification in chronic lymphocytic leukemia. Leukemia 2013; 27: 512–14. Little RJA. A test of missing completely at random for multivariate data with missing values. J Am Stat Assoc 1988; 83: 1198–202. Hallek M, Cheson BD, Catovsky D, et al. Guidelines for the diagnosis and treatment of chronic lymphocytic leukemia: a report from the International Workshop on Chronic Lymphocytic Leukemia updating the National Cancer Institute-Working Group 1996 guidelines. Blood 2008; 111: 5446–56. Byrd JC, Furman RR, Coutre SE, et al. Targeting BTK with ibrutinib in relapsed chronic lymphocytic leukemia. N Engl J Med 2013; 369: 32–42. Roberts AW, Davids MS, Pagel JM, et al. Targeting BCL2 with venetoclax in relapsed chronic lymphocytic leukemia. N Engl J Med 2016; 374: 311–22. Wierda WG, O’Brien S, Wang X, et al. Prognostic nomogram and index for overall survival in previously untreated patients with chronic lymphocytic leukemia. Blood 2007; 109: 4679–85. Pflug N, Bahlo J, Shanafelt TD, et al. Development of a comprehensive prognostic index for patients with chronic lymphocytic leukemia. Blood 2014; 124: 49–62. Tam CS, Seymour JF. A new prognostic score for CLL. Blood 2014; 124: 1–2. Rossi D, Rasi S, Spina V, et al. Integrated mutational and cytogenetic analysis identifies new prognostic subgroups in chronic lymphocytic leukemia. Blood 2013; 121: 1403–12. Wang L, Lawrence MS, Wan Y, et al. SF3B1 and other novel cancer genes in chronic lymphocytic leukemia. N Engl J Med 2011; 365: 2497–506. Byrd JC, Furman RR, Coutre SE, et al. Three-year follow-up of treatment-naive and previously treated patients with CLL and SLL receiving single-agent ibrutinib. Blood 2015; 125: 2497–506. Howlader N, Noone AM, Krapcho M, et al, eds. SEER clinical statistics review, 1975–2007. Bethesda, MD: National Cancer Institute, 2010.

www.thelancet.com/oncology Published online May 13, 2016 http://dx.doi.org/10.1016/S1470-2045(16)30029-8