Prognostic Significance of Actual Dose Intensity in Diffuse Large...

Prognostic Significance of Actual Dose Intensity in DiffuseLarge-Cell Lymphoma: Results of a Tree-Structured Survival

Analysis

By Larry W. Kwak, Jerry Halpern, Richard A. Olshen, and Sandra J. Horning

While diffuse large-cell lymphoma (DLCL) is consideredto be highly curable with current therapy, treatmentfailures are observed even with intensive combinationchemotherapy regimens. In order to study the prognos-tic significance of actual dose intensity of chemother-apy in DLCL, we retrospectively analyzed 115 previ-ously untreated patients treated at Stanford between1975 and 1986 with cyclophosphamide, Adriamycin(doxorubicin; Adria Laboratories, Columbus, OH), vin-cristine, and prednisone (CHOP), methotrexate, bleo-mycin, Adriamycin, cyclophosphamide, vincristine, anddexamethasone ([M]BACOD), or methotrexate, Adria-mycin, cyclophosphamide, vincristine, prednisone, andbleomycin (MACOP-B). The actual relative dose inten-sity (RDI), the amount of drug actually administered toeach patient during the first 12 weeks of therapy, wascalculated as standardized to CHOP and analyzed inaddition to clinical factors prognostic for survival byunivariate analysis. Multivariate recursive partitioning(tree-structured) survival analysis identified the actual

DIFFUSE large-cell lymphoma (DLCL) isthe most common histologic subtype among

the diffuse aggressive lymphomas, accounting forapproximately one third of all non-Hodgkin'slymphomas.' Once almost uniformly fatal, asmany as 50% to 70% of advanced-stage patientsand 80% of limited-stage patients remain alive onthe plateau phase of survival curves followingtreatment with modern combination chemother-apy regimens. 2"7 While a number of prognosticvariables have been identified by several groupsof investigators, reliable identification of patientsat high risk for treatment failure is made difficultby the heterogeneity of DLCL.8 -15

Recently, there has been widespread interestin the concept of dose intensity (DI), the amountof chemotherapeutic drug delivered per unittime, as a significant factor affecting outcome.The correlation between dose intensity of com-bination chemotherapy and treatment out-come has been demonstrated for a number oftumors, including breast cancer, ovarian cancer,Hodgkin's disease, small-cell lung cancer, andmultiple myeloma.' 6 21 DeVita et al have sug-gested that DI is also correlated with survival in

RDI of Adriamycin greater than 75% as the singlemost important predictor of survival. A model incorpo-rating the actual RDI of Adriamycin and performancestatus, in combination with serum lactate dehydroge-nase (LDH) and extranodal disease, defined threeoverall prognostic groups of patients with respective3-year survival rates of 89%, 63%, and 18%. Thethree prognostic groups remained distinct, even whenrestricted to complete responders. This model was alsopredictive of survival when dose intensity was ana-lyzed relative to the optimum dose defined for each ofthe three regimens and when applied to a subgroupof patients aged 50 years or younger. We concludethat actual RDI is an important prognostic factor forsurvival in DLCL and that analysis of RDI early in thecourse of treatment may allow modification of thetreatment plan.J Clin Oncol 8:963-977. o 1990 by American Societyof Clinical Oncology.

DLCL and that combinations of six or moredrugs are superior to cyclophosphamide, Adria-mycin (doxorubicin; Adria Laboratories, Colum-bus, OH), vincristine, and prednisone (CHOP).22

However, DI has been a poorly controlled vari-able in almost all studies of DLCL, and it is not

From the Division of Oncology, Department of Medicine,and the Division of Biostatistics, Department of HealthResearch and Policy, Stanford University Medical Center,Stanford, CA.

Submitted May 15, 1989; accepted January 25, 1990.Supported in part by National Institutes of Health Grants

CA 34233 and CA 41628, the John Simon GuggenheimMemorial Foundation, and National Science FoundationGrant Division of Mathematical Sciences 85-05609.

Larry W. Kwak, MD, PhD, was a Fellow of the AmericanCancer Society, and Richard A. Olshen, PhD, was a Re-search Scholar in Cancer of the American Cancer Society.

Presented in part at the 24th Annual Meeting of theAmerican Society of Clinical Oncology, May 22-24, 1988,New Orleans, LA.

Address reprint requests to Sandra J. Horning, MD,Division of Oncology, Room M211, Stanford UniversityMedical Center, Stanford, CA 94305.

© 1990 by American Society of Clinical Oncology.0732-183X/90/0806-0007$3.00/0

Journal of Clinical Oncology, Vol 8, No 6 (June), 1990: pp 963-977 963

171.65.37.237Information downloaded from jco.ascopubs.org and provided by at STANFORD UNIV MEDICAL CENTER on March 4, 2014 from

Copyright © 1990 American Society of Clinical Oncology. All rights reserved.

KWAK ET AL

clear that the more intensive multidrug regimensare superior to CHOP, delivered at full dosesevery 3 weeks, to comparable patients. For thesereasons, we undertook a survival analysis of 115patients with DLCL treated at Stanford between1975 and 1986 with three combination chemo-therapy regimens: CHOP, methotrexate, bleomy-cin, Adriamycin, cyclophosphamide, vincristine,and dexamethasone ([M]BACOD), and meth-otrexate, Adriamycin, cyclophosphamide, vincris-tine, prednisone, and bleomycin (MACOP-B).Using the multivariate statistical approach ofrecursive partitioning, or tree-structured, sur-vival analysis, we analyzed drug DI, based on theamount of drug actually delivered during the first12 weeks of therapy, in combination with pretreat-ment clinical prognostic factors. In the process ofthis analysis, a model was constructed that pro-duced three distinct prognostic groups with mark-edly different overall survival.

PATIENTS AND METHODS

Patient Characteristics

One hundred fifteen previously untreated patients with ahistologic diagnosis of DLCL, intermediate grade or immuno-blastic, as defined by the Working Formulation, who wereseen and treated at Stanford between 1975 and 1986,comprised the study group.' Patients were staged accordingto the Ann Arbor criteria.23 Initial staging procedures in allpatients included complete physical examination, chest radio-graph, routine blood counts and chemistry tests, bone marrowbiopsy, bipedal lymphangiography, and/or abdominal andpelvic computed tomography. Additional tests were per-formed as clinically indicated.

Patients in this study received CHOP, (M)BACOD, orMACOP-B treatment in standard doses in 3-week cycles forCHOP and (M)BACOD and weekly for MACOP-B.21

4'24 Thehigh-dose methotrexate in (M)BACOD was given as 1 g/m 2

on day 14 of each cycle with leucovorin rescue. Selection oftreatment regimen was individualized for each patient, but inmost cases, reflected the regimen with the best resultsreported at that time. Thus, the standard therapy for ad-vanced-stage patients with DLCL at Stanford was CHOPbefore 1981, (M)BACOD between 1981 and 1985, andMACOP-B beginning in 1985. CHOP continued to be usedafter 1981 for limited-stage patients, especially for those withnonbulky or asymptomatic disease. Patients who receivedradiation therapy were included in the analysis only if it wasgiven after the completion of chemotherapy.

Sixty-eight patients were treated with CHOP, 31 with(M)BACOD, and 16 with MACOP-B. The clinical character-istics of these patients are shown in Table 1. There wereseveral differences in age and sex ratio among the groups,whereas the distribution of Eastern Cooperative OncologyGroup (ECOG) performance scores was essentially constantacross all three groups. Performance status was based on the

ECOG scale, in which 0 is associated with no symptoms, 1with symptoms but continued ability to ambulate, 2 withbedridden status less than 50% of the day, 3 with bedriddenstatus greater than 50% of the day, and 4 with chronicbedridden status and a requirement for assistance for dailymaintenance. Overall, approximately one third of patientspresented with B symptoms. The (M)BACOD group con-tained the highest proportion of advanced-stage patients(74%). Overall, the majority of patients were Ann Arborstage III or IV, and the (M)BACOD and MACOP-B groupscontained the highest proportion of patients with elevatedpretreatment lactate dehydrogenase (LDH), multiple extra-nodal sites of involvement, and bulk of disease, defined as thelargest dimension of a single tumor mass. The size of thelargest mass was determined by review of both the medicalrecord and relevant radiographic studies and included bothnodal and extranodal sites. Actual dimensions of masses a 3cm in diameter were used in the analysis. Sixty-four percentof patients presented with extranodal sites of disease of whichthe gastrointestinal tract was the most commonly involvedextranodal site (19%). Other sites of extranodal diseaseincluded pleura (13%), bone marrow (12%), bone (11%),lung (10%), testis (9%), skin (7%), pericardium (6%), kidney(6%), and liver (3%).

All patients underwent restaging with baseline stagingprocedures during treatment and at the completion of treat-ment. Bone marrow biopsies were, for the most part, repeatedonly in patients having positive biopsies on presentation.Complete response was defined as the disappearance of allclinical evidence of disease and either the normalization orstabilization of all laboratory and radiographic abnormali-ties. To be classified as a complete response, the response hadto be maintained for at least 30 days following the completionof therapy. Determination of the number of cycles to com-plete remission was based on restaging studies describedabove. A value of zero cycles was assigned to patients whohad undergone surgical resection of all gross disease beforethe initiation of chemotherapy (11 patients).

DI

Actual chemotherapy doses were available for all 115patients. The method of Hryniuk and Bush'6 was used tocalculate the DI of each drug actually administered to thepatient. For the purposes of this analysis, we considered theamount of each drug, normalized to body surface area(mg/m2), administered during the first 12 weeks of therapy.The period of 12 weeks was selected because (1) the MACOP-B regimen is of 12 weeks duration, (2) if found to beimportant, DI could then be used to make treatment deci-sions, including changes in therapy, relatively early in thecourse of treatment, and (3) a previous study has shown thatresponse during the first three to four cycles of chemotherapyis prognostically significant.2 5 Although this method mayintroduce bias in patients with early disease progression, theprimary regimen was changed in only three patients beforethe 12th week of treatment.

DI was expressed as a decimal fraction of the doseprescribed in a standard regimen over the same time frame(relative DI [RDI]). CHOP was selected as the standardregimen because the group treated with CHOP representedthe majority of our patients and also because CHOP remains

964



DOSE INTENSITY IN DIFFUSE LARGE-CELL LYMPHOMA

Table 1. Patient Characteristics by Regimen

Characteristic CHOP (M)BACOD MACOP-B Total (%)

No. of Patients 68 31 16 115

Age

c 50 years

51-65 years

> 65 years

Sex

Male

Female

ECOG performance status0-2

3-4

B symptoms

AbsentPresent

Stage

I, IIIII, IV

Hemoglobin<12

- 12

Lactate dehydrogenase

Normal range

Elevated

Bone marrow involvement

YesNo

Number of extranodal sites0-1

- 2

Tumor bulk

< 10cm

> 10cm

Histology

Immunoblastic

Nonimmunoblastic

Consolidative radiotherapy

YesNo

30

26

12

43

25

608

51

17

35

33

16

52

29

39

365

57

11

53

15

1256

2246

17

13

1

16

15

265

18

13

8

23

11

20

8

23

724

18

13

15

16

11

20

1219

11

5

0

6

10

14

2

7

9

79

5

11

412

4

12

10

6

10

6

610

106

58

4413

65

50

100

15

76

39

50

65

32

83

41

74

14

101

85

30

7837

29

86

4471

965

(51)

(38)(11)

(57)

(43)

(87)

(13)

(66)

(34)

(43)(57)

(28)

(72)

(36)

(64)

(12)

(88)

(74)(26)

(68)

(32)

(25)

(75)

(38)

(62)

the standard of comparison for the second and third genera-tion combination chemotherapy regimens. Average RDI forthe combination of drugs was calculated by taking thearithmetic mean of RDI of the four individual drugs for eachpatient. The projected RDIs for each drug and the four-drugaverage for each of the regimens standardized to CHOP areshown in Table 2. These are based on 100% of the calculateddose for that particular regimen without any delays intreatment. From these calculations, it is apparent that theregimen with the highest DI of cyclophosphamide is CHOP,while MACOP-B has the highest DI of Adriamycin. Thefollowing assumptions were made in calculating actual RDI:(1) a maximum value of 1.00 was allowed for the RDI of

prednisone/dexamethasone when calculating average RDI,since no clear dose-response relationship exists for steroidsgiven in therapeutic doses, (2) 1 mg of dexamethasone wasequivalent to 6.7 mg of prednisone, based on relative glucocor-ticoid activity, (3) the capping off of vincristine doses at 2.0mg for CHOP meant that the RDI of this drug for patientstreated with (M)BACOD varied with body surface areaabove 1.43 m2, (4) for two patients who had RDI of a singledrug greater than 1.00 (other than prednisone), a value of1.00 was assigned when calculating average RDI, and (5)three patients treated with (M)BACOD received higherdoses of methotrexate, between 1 and 3 g/m 2. A maximumvalue of 1.50 was assigned to the RDI of this drug when



966 KWAK ET AL

Table 2. Univariate Analysis of Actual DI for Survival

Relative Dose Intensities*

CHOP (M)BACOD MACOP-B

Median Median MedianRDI Projected Actual Projected Actual Projected Actual Pt

Four-drug average 1.00 0.96 0.84 0.84 1.18 1.17 .04

Cyclophosphamide 1.00 0.93 0.80 0.79 0.70 0.67 .03Adriamycin 1.00 0.95 0.90 0.86 1.50 1.45 .02Vincristine 1.00 1.00 0.85 0.90 1.50 1.44 .28

Prednisone 1.00 0.99 0.80 0.79 1.00 1.00 .34

NOTE. Actual RDI was analyzed for survival as a continuous variable across groups (115 patients), separately for each drug.

*Standardized to CHOP.

"tP values were determined by an approximate likelihood analysis of a Cox proportional hazards model.

calculating average RDI (for each regimen as its ownstandard).

Statistical Methods

Survival curves were calculated from the date of initiationof treatment according to the actuarial method of Kaplan andMeier.2 6 The generalized Wilcoxon test of Gehan was used toassess the significance of differences between patient groups. 27

The ability of each variable alone to predict survival wastested by a univariate Cox analysis.2 8 Multivariate Coxanalyses with up to eight variables were also done. For aparticular model, each variable was treated separately, andup to five variables were tested in combination for theiradditional prognostic significance beyond that of the othervariables in the model.28,

29

A new and important aspect of our work is the use of a"tree-structured" survival analysis. The method, also termed"recursive partitioning," is based on extensions by Gordonand Olshen and others of classification and regression trees. 3

0-33

With this method, the entire patient group is divided into twosubgroups as defined by the patient characteristic (eg, sys-temic symptoms) that identifies the subgroups most differentin prognosis. These two subgroups are again partitioned(each subgroup being split on the same or other prognosticfactors that provide the greatest difference in prognosis), thuscreating a tree-structure. This process is continued until nofurther subdividing is worthwhile for prognostic purposes. Insum, the tree-growing paradigm for each group (1) examinesevery allowable split on each prognostic factor and (2) selects(creates subgroups) the best of these splits. The best splitswere selected by computing a log-rank statistic for thesurvival differences. The prognostic factors, their values, andthe subgroups for which the survival differences were greatestwere further selected, using a criterion that compared thehorizontal distance between the survival functions in eachsubgroup pair.30 Kaplan-Meier survival curves were made foreach distinct group identified by our model, and pairs of thesecurves were compared with Gehan's generalized Wilcoxontest.

Cross-validation, which estimates how well a tree will do onfuture data, was also used." This sample reuse methoddeletes a random patient group from the data, constructs atree on the remaining data, and uses the group left out as"new data" on which to test the tree. The technique has abuilt-in method for handling missing data, of which there is

very little in our study group (LDH was missing in a singlepatient).3

RESULTS

Response and Survival Data

Figure IA shows the Kaplan-Meier survivalcurve for the total group of 115 patients. Theactuarial 5-year survival is 58% with a medianfollow-up of survivors of 3 years (range, 8 monthsto 7.5 years). The survival of the patients isstratified by regimen in Fig 1B. As shown, whenthe three regimens were compared for survival inthis retrospective, nonrandomized analysis, nostatistically significant differences were ob-served. Complete response (CR) rates were highfor all three groups, although the (M)BACODregimen was associated with a somewhat lowerCR rate (68%) compared with the other tworegimens (CHOP, 87% and MACOP-B, 94%).The lack of observable differences in survivalmay relate to relatively small patient numbersand to the fact that the regimens, for the mostpart, were not used concurrently. Differencesamong the three groups in patient characteristicsthat may influence survival are noted in Table 1and further analyzed below.

Univariate Analysis of Clinical Features forSurvival

Eleven pretreatment clinical factors were eval-uated individually for the entire group of 115patients as prognostic indicators of survival (Ta-ble 3). ECOG performance status, number ofextranodal sites of disease, and Ann Arbor stage,all analyzed as continuous or ordered variables,were significant (P < .05), with poor perfor-mance status, greater number of extranodal sites,and more advanced stage each associated with an




A 00oo

>-

0moa.

80

60

40

20

00 1 2 3 4

TIME (YEARS)

B 0 CHOP

0 (M)BACOD

( MACOP-B,

5 6 7

87% CR

68% CR

94% CR

100

I--J

m

Q

ca

80

60

40

20

Fig 1. (A) Actuarial survivalof 115 patients with DLCL. (B)Actuarial survival and completeremission rates of 115 patientswith DLCL by regimen.

0

adverse effect on survival. Bone marrow involve-ment and the presence of B symptoms were alsoassociated with a significantly adverse effect onsurvival. Hemoglobin and LDH, both analyzedas continuous variables, were of borderline statis-tical significance, as was sex, in favor of females.Single factors not predictive for survival includedage and bulk, analyzed as continuous variables,histology (intermediate grade v immunoblastic),and specific nodal or extranodal sites of involve-ment other than bone marrow. The small numberof patients involved in each specific site of involve-ment may have precluded the full analysis of itsinfluence on survival.

The effect of consolidative radiation therapy

Gehan P-values1 vs. 2 0.968

0 1 2 3 4 5 6 7

TIME (YEARS)

on survival was also evaluated. Thirty patientshad been selected to receive radiation therapy tosites of initial bulky disease following the achieve-ment of CR and completion of chemotherapy.When they were compared with patients havingan initial mass > 5 cm who had not receivedconsolidative radiation therapy in CR (n = 23) asignificant difference in survival was observed infavor of radiotherapy (Table 3), suggesting abeneficial role for radiation therapy as an adjunctto chemotherapy in this setting.

The time interval required to achieve CR wasalso examined as a variable for survival. Amongthe 96 patients (83%) who achieved a CR, thenumber of cycles of chemotherapy required to

967



KWAK ET AL

Table 3. Factors Prognostic for Survival

Variable P*

Performance status < 0.01t

No. extranodal sites < 0.01-Stage 0.02t-Marrow involvement 0.02B symptoms 0.04

Hemoglobin 0.06tSex 0.07LDH 0.08"f

Age 0.17t

Histology 0.68Bulk 0.98-

Consolidative radiotherapy (n = 53) 0.01:1

No. cycles to CR (n = 96) 0.61t

*P values were determined by an approximate likelihood ratio

statistic for a Cox proportional hazards model.

tlndicates when the corresponding factor was analyzed as acontinuous or ordered variable.

*Analysis included patients with initial bulky mass > 5 cm who

did (n = 30) or did not (n = 23) receive radiation therapy

following achievement of complete remission.

achieve CR, analyzed as a continuous variable,was not of prognostic significance (Table 3).These results should be interpreted with caution,however, as not all patients were restaged at thesame time points in their therapy, and somepatients may have achieved CR at time pointssomewhat earlier than indicated by formal restag-ing.

Relative DI

Shown in Table 2 are the median actual RDIand the corresponding projected RDI of eachdrug and of the four-drug average for each of theregimens, standardized to CHOP, based on dosesof chemotherapy received by each of the 115patients during the first 12 weeks of therapy asdescribed above. With the exception of six pa-tients, there were no arbitrary dose reductions foradvanced age or poor performance status. Themost frequent reason for dose attenuation ortreatment delay was leukopenia (8%), or anepisode of fever with absolute neutropenia associ-ated with a prior cycle of treatment (12%). Liverdysfunction necessitated dose attenuation ofAdriamycin in two patients, and prednisone wasremoved from the regimen of one patient becauseof severe mental status changes and in anotherpatient because of severe muscle weakness. Neu-rotoxicity associated with vincristine resulted indose attenuations or deletions of this drug fromthe regimen of 11% of patients. However, pa-

tients in all three treatment groups received closeto full projected doses as shown by the smalldifferences between projected and actual RDI foreach of the drugs within each regimen.

When evaluated separately as continuous vari-ables for the entire group of 115 patients, averageRDI, as well as the RDI for cyclophosphamideand Adriamycin were significant (P < .05) forsurvival (Table 2). The RDIs for vincristine andprednisone were not significant for survival inthis analysis. Similarly, the RDI for methotrex-ate and bleomycin evaluated for the 47 patientstreated with (M)BACOD or MACOP-B had nosignificant effect on survival (P > .05, data notshown).

Tree-Structured Survival Analysis(Multivariate Recursive Partitioning)

The covariates selected for this analysis were(1) the pretreatment clinical factors that wereprognostically significant in the univariate analy-sis (ECOG performance status, number of extra-nodal sites, stage, marrow involvement, B symp-toms, and LDH), (2) age, sex, and bulk, whichwere not statistically significant in the univariateanalysis but were included because other retro-spective analyses have found these factors to beprognostically important, (3) average RDI andthe RDI for cyclophosphamide and Adriamycin,and (4) treatment regimen.

Tree-structured survival analysis identified theRDI of Adriamycin greater than 75% as the firstsplit point (Fig 2). This divided the patients intogroups of 92 and 23 patients; superior survivalwas associated with the group who receivedhigher doses of Adriamycin (Fig 3A). The signif-icance of actual DI as a determinant of survivalwas further supported by the identification ofaverage RDI (all four drugs) and RDI of cyclo-phosphamide as the surrogate covariates. Theterm surrogate covariate refers to the covariatethat best predicts the best split. 32

The less favorable group (those with RDI ofAdriamycin < 75%) of patients was best furthersubdivided by pretreatment LDH resulting intwo terminal subgroups (Fig 2, subgroups I andII). Patients with RDI of Adriamycin < 75% andpretreatment LDH greater than 1.4 x normalformed a particularly poor prognostic group (Fig3B). The surrogate covariate for this node wasstage.

968




RDI

ADRIAMYCIN

> .75

ECOG PERFORMANCE

0-2 3-4

NUMBER EXTRANODAL IIISITES

IV V 1

IV V

Fig 2. Results of multivariate recursive partitioning (tree-structured) survival analysis of 115 patients with DLCL. LDH1.4 x upper limit of normal value. ECOG Performance,Eastern Cooperative Oncology Group performance status.Numbers in squares indicate number of patients in eachsubgroup. Roman numerals designate terminal subgroups.

A( RDI ADRIAMYCIN >.75 (92 PTS)( RDI ADRIAMYCIN <.75 (23 PTS)

DH.75

LDH

Gehan P-values1 vs. 2 0001

~·--............

2 3 4 5 6 7

TIME (YEARS)

M

.0aM

100

80

60

40

20

0


0 1 2 3 4

TIME (YEARS)

RDI ADRIAMYCIN > 75 AND:1 ECOG PERFORMANCE STATUS 0-2 (82 PTS)

C ECOG PERFORMANCE STATUS 3-4 (10 PTS)

0aa0.

IL

==

100

80

60

40

20

0

TIME (YEARS)

D DI ADRIAMYCIN > 75 AND ECOG PERFORMANCE STATUS 0-2 AND:

( NUMBER OF EXTRANODAL SITES = 0 (35 PTS)( NUMBER OF EXTRANODAL SITES= 1-4 (47 PTS)

2 3 4 5

TIME (YEARS)

Fig 3. Actuarial survival curves of the subgroups of patients generated at each sequential split point in the "tree" shown inFig 2. Split points: (A) RDI of Adriamycin, (B) LDH, (C) ECOG performance status, (D) number of extranodal sites of disease.

Figure 2 shows that patients who receivedgreater than 75% RDI of Adriamycin were bestfurther subdivided by ECOG performance sta-tus. Patients with poor initial performance scores(ECOG 3 or 4) despite greater than 75% RDI ofAdriamycin (subgroup III), had a very poorprognosis as demonstrated in Fig 3C. The surro-gate covariates for this node were number ofextranodal sites and age.

The favorable group of patients with RDIAdriamycin greater than 75% and good perfor-mance status was then further subdivided by thenumber of extranodal sites of disease. As shownin Fig 2, this generated the most favorablesubgroup (IV) of all patients, and a secondsubgroup (V) with an intermediate prognosis.These results are shown in Figure 3D. Thesurrogate covariates for this node were age andthe RDI of Adriamycin and cyclophosphamide.

The resulting five subgroups of patients ateach terminal node divide naturally into three

BRDI ADRIAMYCIN 1 75 AND:

() LDH < 1.4 x NORMAL (11 PTS)( LD H >1.4 x NORMAL (12 PTSI

969

<1.4 X

I

> 1II4 X

II

100

80

60

40

o-

a.0

201

0

C

100

80

S60

` 40

20

Gehan P-values1 vs 2 0D030

0D

1L0

1



970 KWAK ET AL

overall prognostic groups. The survival curves ofthese three groups are shown in Fig 4. Thefavorable group (3-year survival, 89%) consistsof those patients with RDI of Adriamycin greaterthan 75%, ECOG performance status 0 to 2, andno extranodal disease (subgroup IV). The poor-prognostic group (3-year survival, 18%) consistsof those patients with RDI of Adriamycin < 75%together with significantly elevated LDH (sub-group II), or poor performance status, regardlessof DI (subgroup III). The intermediate group(3-year survival, 63%) consists of the remainingpatients: those with RDI of Adriamycin _ 75%without significantly elevated LDH (subgroupI), and those with RDI Adriamycin greater than75% and good performance status, but at leastone extranodal site of disease (subgroup V). TheCR rates for the three groups were 97% for thefavorable, 84% for the intermediate, and 60% forthe poor prognostic groups. It is important tonote that 60% of the patients in the poor prognos-tic group achieved CR, as it suggests that theeffect of the variables we have identified onsurvival is not solely related to the ability toachieve a CR. This point is further illustrated bythe survival curves in Fig 5, which show that thethree prognostic-group comparisons maintain sta-tistical significance when this model is appliedonly to the 96 patients who achieved a CR.

One of the aims of this study was to identify a

(T) FAVORABLE (35 PTS)

© INTERMEDIATE (58 PTS)

( POOR (22 PTS)

100

subgroup of younger patients with poor prognosiswho may be candidates for very aggressive orinvestigational therapy, such as bone marrowtransplantation. To this end, we applied themodel to the subgroup of patients aged s 50years (n = 58). The analysis for this subset ofpatients was not as extensive, in part because ofthe smaller number of patients, and the majorityof the patients fell into a favorable prognosticgroup; that is, high RDI of Adriamycin and goodperformance status (Fig 6A). However, using thesame covariates with slightly different split points,it was possible to identify a small subgroup ofpatients at high risk for failure (Fig 6B). Thecharacteristics of this poor prognostic group(actuarial 2-year survival, approximately 20%)were (1) RDI of Adriamycin < 85% and LDHgreater than 1.1 times normal or (2) ECOGperformance status 3 or 4, despite a high RDI ofAdriamycin.

An alternative method of analyzing DI wouldbe to consider the amount of drug deliveredduring the first 12 weeks of therapy relative tothe optimum dose defined for each regimen. Suchan approach would allow more general applica-tion to other regimens used in the treatment ofDLCL. While it is understood that 75% of theoptimum Adriamycin dose in one regimen maynot be equivalent to that in another regimen,considering each regimen as its own standard

97% CR

84% CR

60% CR

Fig 4. Actuarial survivalcurves and CR rates for favor-able-, intermediate-, and poor-prognosis patients identified bytree-structured survival analysisin Fig 2. The respective 3-yearsurvival rates are 89%, 63%,and 18%. The favorable groupconsists of terminal subgroupIV, the intermediate group con-sists of subgroups I and V, andthe poor prognosis group con-sists of subgroups II and III.

I-CJ0a

Q.

80

60

40

20

00 1 2 3 4 5 6 7

TIME (YEARS)



DOSE INTENSITY IN DIFFUSE LARGE-CELL LYMPHOMA 971

( FAVORABLE (34 PTS)

) INTERMEDIATE (49 PTS)

( POOR (13 PTS)

o-

>-

m

Occ0-

100

80

60

40

20

Fig 5. Actuarial survivalcurves of the three prognosticgroups of patients restricted topatients who achieved CR (96patients).

0


0 1 2 3 4 5 6 7

TIME (YEARS)

RDI

ADRIAMYCIN

.85 .85

LDH ECOG PERFORMANCE

<1.1 x 1.1 X 02 3-

( RDI-A > 85% AND ECOG 0-2 OR RDI-A <85% AND LDH <1.1 X

( RDI-A 5 85% AND LDH > 1.1 X OR RDI-A > 85% AND ECOG 3-4

Gehan P-valuesi ,. ) - n nni

5 6

TIME (YEARS)

Fig 6. (A) Recursive parti-tioning survival analysis appliedto patients aged s 50 years (58patients). Numbers in squaresindicate number of patients ineach subgroup. (B) Actuarial sur-vival curves of the four terminalsubgroups grouped by progno-sis. The majority of patients ex-hibited favorable prognosis (49patients). Characteristics of thepoor-prognosis group (nine pa-tients) were RDI of Adriamy-cin 5 85% together with LDH >1.1 times normal or ECOG per-formance status 3-4 despite RDIof Adriamycin > 85%.

A

B

1

:3-J

m

rrp

0

.0

1



KWAK ET AL

also would reflect indirectly the impact of addi-tional drugs in the regimen not included inCHOP (eg, bleomycin, methotrexate). Whenthis approach was used in calculating DI, theRDI for Adriamycin, cyclophosphamide, and theaverage RDI (four or six drugs) were againhighly significant for survival in univariate analy-sis (P < .01, data not shown). Moreover, whenthese values of RDI were incorporated into themultivariate model, again three overall prognos-tic groups were generated (Fig 7).

In order to more fully assess the confoundingof RDI with other prognostic variables, theclinical features of the 23 patients with RDIAdriamycin less than 75% were separately evalu-ated. Six (26%) were older than 65 years of age,five (22%) had an ECOG performance status of 3or 4, 14 (61%) had stage IV disease, 11 (48%)had pretreatment LDH greater than 1.5 x nor-mal, four (17%) had more than two extranodalsites of disease, and six (26%) had tumor bulkgreater than 10 cm in diameter. However, whenevaluated individually, less than one half ofpatients (11 of 23) had more than two adverseprognostic clinical factors, as defined by theresults of the univariate analysis in Table 3. Thereasons for Adriamycin dose attenuation or delayin therapy in these patients were leukopenia orfever with neutropenia associated with a priorcycle of therapy (nine patients), cholestasis, which

© FAVORABLE (36 PTS)

( INTERMEDIATE (58 PTS)

POOR (21 PTS)

was idiopathic or associated with chronic activehepatitis (two patients), patient preference (threepatients), and arbitrary dose reduction dueto advanced age (one patient). The remainingeight patients received fewer than 12 weeks oftheir primary chemotherapy regimen because ofdisease progression (three patients) or earlytermination of chemotherapy following CR(four patients), and one patient received cyclo-phosphamide, vincristine, and prednisone (CVP)as his first cycle of chemotherapy.

Because of the suggestion from our data thatlower doses of Adriamycin tended to be associ-ated with M-BACOD and this therapy wasadministered to patients with poor-prognosticfeatures, the model was applied to the populationexcluding those treated with M-BACOD. Poorperformance status or RDI of Adriamycin lessthan 75% together with LDH greater than 1.4 xnormal continued to define a prognostic groupwith significantly worse survival (P < .001, datanot shown).

Multivariate Proportional Hazards RegressionAnalysis of Survival

The same covariates were also included inmultivariate Cox regression analyses for sur-vival. Only a single patient had to be excludedfrom these analyses because of missing data(pretreatment LDH was not available). Table 4

DI RELATIVE TO OWN REGIMEN


Fig 7. Actuarial survivalcurves of the three prognosticgroups in Fig 4 with DI calcu-lated relative to the optimumdose defined for each regimen.The three groups remain distinctwhen DI is standardized to regi-mens other than CHOP.

100

80

60

40

0

0t~rn

20

02 3 4 5

TIME (YEARS)

972




Table 4. Multivariate Proportional Hazards Regression Analysis of Survival

Covariates (Significance Level)

Effect of clinical factors ECOG PS No. extranodal sites LDH Marrow Sex Stage Age

A 0.03 0.57 0.11 0.24 0.02 0.47 0.21

Effect of RDI ECOG PS No. extranodal sites LDH RDI RDI RDI

Adria Ctx Average

B 0.04 0.24 0.47 - - -

C 0.14 0.22 0.42 0.02 0.04 0.07

Effect of regimen ECOG PS No. extranodal sites LDH RDI CHOP M-BACOD MACOP-B

Adria

D 0.09 0.39 0.28 0.06 - - -

E 0.08 0.23 0.19 0.07 0.15 - -

F 0.07 0.24 0.20 0.01 - 0.07 -

G 0.09 0.50 0.32 0.07 0.51 0.18 -

Abbreviations: PS, performance status; Adria, Adriomycin; Ctx, Cytoxan (cyclophosphamide; Bristol-Myers Co, Evansville, IN).

describes the results of selected Cox regressionanalyses based upon the clinical factors that wereprognostically significant in the univariate analy-ses (Table 3), variables identified in the tree-structured analyses, and the three regimens used.Among the pretreatment clinical factors found tobe significant in the univariate analysis or tree-structured analyses, only ECOG performancestatus consistently retained additional signifi-cance beyond that of various combinations offactors (P < .05). Of interest, sex in favor offemales also was found to be significant incombination with a number of pretreatmentclinical factors (P = .02). The individual RDI ofAdriamycin and cyclophosphamide, as well asthe average RDI, also retained significance inde-pendently (P = .02, .04, .07, respectively), andin combination (P = .048 by the likelihood ratiotest), when analyzed together with ECOG perfor-mance status. These results with proportionalhazards regression support the tree-structuredanalysis model, especially with regard to theindependent significance of RDI of Adriamycin.Treatment regimen provided no significant addi-tional prognostic information, when analyzedwith RDI and performance status.

DISCUSSION

There has been increasing interest in theconcept of DI with respect to cancer chemother-apy. Retrospective studies of chemotherapeuticregimens for a variety of cancers have reaffirmedthe importance of dose in achieving a maximumtherapeutic effect in responsive tumors and havehighlighted the implications of this concept for

the design and analysis of future clinical tri-als. 16-21 '34 In our analysis of 115 patients withDLCL, we have shown that actual RDI, asmeasured during the first 12 weeks of therapy, isan important prognostic factor for survival. Us-ing a tree-structured analysis, the characteristicthat generated two patient groups most differentin prognosis was the actual RDI of Adriamycin.Univariate analysis and multivariate Cox regres-sion analysis also confirmed the impact of actualRDI on survival in DLCL.

DI has been shown to be important in thetreatment of lymphomas. Carde et al reportedthat DI during the first three cycles of mechlor-ethamine (nitrogen mustard), vincristine, pred-nisone, and procarbazine (MOPP) chemother-apy was significantly related to outcome inHodgkin's disease.' 9 Using a hypothetical nine-drug regimen as the standard of comparison,DeVita et al at the National Cancer Institute(NCI) used DI analysis to calculate projectedDIs for a variety of treatment programs used inthe management of diffuse aggressive lympho-mas.22 They found a strong correlation of averageRDI (based on nine drugs) to long-term survival,but no relationship of outcome to separatelycalculated two-drug average RDI for Adriamy-cin and cyclophosphamide. While our analysisconfirms a correlation between average RDI andoutcome, there are several important differencesbetween the NCI analysis and our study, whichmay explain the discrepancy with regard to theindividual RDI of Adriamycin and cyclophospha-mide. First, the availability of specific data onactual doses of drugs delivered to each of the 115patients in our study allowed precise calculation

973



KWAK ET AL

and subsequent analysis of actual DI. It isimportant to emphasize the difference betweenactual and projected DI. Actual DI may varysignificantly from projected DI for individualpatients because of the nature of dose adjust-ments based on drug toxicity or other factors.Clearly, actual DI is the datum of greater impor-tance and would be expected to correlate muchmore closely with outcome than DI calculatedfrom an intended protocol.3 5 Second, initial dosereductions up to 50% were allowed for patientsover age 65 treated with CHOP in the SouthwestOncology Group (SWOG) studies cited in thestudy by DeVita et al. The effect of these dosereductions on actual DI could not have beenreflected in calculations of projected DI, and thisfactor may, at least in part, explain the lack ofcorrelation of the two-drug average RDI withoutcome. Third, two assumptions made whenusing the DI method to calculate average RDI oftwo drugs, (1) that both drugs in a multidrugregimen are equally effective and (2) that theeffects of the individual drugs are additive, areavoided in a multivariate analysis in which theRDI of Adriamycin and cyclophosphamide aretreated as separate variables. Finally, calcula-tions of DI for CHOP in the NCI study werebased on 4-week cycles; at Stanford CHOP isroutinely administered every 3 weeks as origi-nally described by McKelvey et al.24

Nevertheless, several limitations, outlined inrecent reviews, remain with all studies using theDI approach. 35 37 Among these is the assumptionthat drug scheduling is relatively unimportant.In addition, it remains to be seen to what extentDI correlates with outcome independent of totaldose. For example, among 121 patients withDLCL treated with at least eight cycles of m- orM-BACOD chemotherapy, no relationship wasfound between survival and the percentage of thetotal prescribed dose of each drug actuallyadministered.'"

One of our goals in evaluating this series ofpatients with DLCL was the identification ofprognostic groups that would have implicationsfor the design of randomized therapeutic trialsand for determination of optimal therapy forindividual patients. The actual RDI of Adriamy-cin and performance status, in combination withpretreatment LDH and the number of extra-nodal sites of disease, defined three groups of

patients with markedly different survival. Theidentification of these prognostic groups under-scores the predictive power of the novel statisticalapproach of recursive partitioning, or tree-structured, survival analysis. In contrast to multi-variate proportional hazards regression analysis,this method only identifies prognostic factorsthat are present in actual patients, rather than ahypothetical patient who might possess multiplesignificant factors. The analysis presented hereillustrates the prospective use of tree-structuredanalysis and, once the prognostic groups havebeen identified, the ease with which a givenpatient may be properly assigned to one of them.Other advantages of this approach include thatvery few assumptions are necessary to use it (andhence its broad applicability), its automatic iden-tification of interactions (ie, synergistic effectsamong the variables included for analysis), andthe natural way that it handles missing data.30 ,32,38

The tree-growing paradigm of examiningnearly every allowable split on each variable andselecting the best of these splits at each point hasbeen successfully applied to a variety of classifi-cation data,39-44 and in several of these applica-tions, performance of tree-based classifiers ontest subjects has proved superior to that oflogistic regression. When directly compared withother methods, tree-based analyses have usuallybeen somewhat more accurate, and have onlyrarely performed less well. For instance, clinicalfactors predictive for myocardial infarction, out-come of hypoxic coma, multiple markers for lungcancer diagnosis, lymphocyte markers and as-says for immunosuppression classification, andpatient characteristics predictive of findings atlaparotomy in Hodgkin's disease have all beenidentified using this statistical method.

The three risk groups generated in the modelwere also prognostic when limited to the popula-tion of complete responders. The relatively highCR rate of 60% in the poor prognostic groupsuggests that the remissions among the three riskgroups were not equally durable. These resultsare in contrast to the observation that the im-provement in overall survival observed betweentwo generations of chemotherapy regimens re-lates solely to the ability of the regimens withgreater DI to induce higher CR rates.2 2 RDI andthe clinical prognostic factors we have identifiedmay not only exert their influence on the achieve-

974




ment of a CR, but may also be used to predictgroups at higher risk for relapse following CR.We have also shown in Fig 7 that the modelremains valid when the DI analysis is standard-ized to regimens other than CHOP. In fact, wheneach regimen was considered as its own standard,a range of values for RDI of Adriamycin wasfound over which three significantly different riskgroups were generated. The precise limits of thesplit point on RDI of Adriamycin for generalapplication to patients with DLCL must bedetermined by comparison with other retrospec-tive and future prospective studies of actual RDIof this drug in other regimens. Thus, in Fig 7 thesplit on RDI of Adriamycin greater than 75%refers to 75% of the optimal dose of Adriamycinin the particular regimen used.

The favorable prognostic group (3-year sur-vival, 89%) appears to be effectively treated withexisting chemotherapy regimens. Significant im-provements over existing survival rates for thisgroup of patients may be difficult to demonstrateand achieve without substantial additional toxic-ity. Conversely, the poor prognostic group (3-year survival, 18%) requires alternative therapy.Several reports have suggested a role for high-dose therapy and autologous bone marrow trans-plantation in selected patients with relapsednon-Hodgkin's lymphoma.45'46 Younger patientsat high risk for treatment failure may be candi-dates for primary bone marrow transplantationfollowing cytoreductive chemotherapy to CR orminimal disease status, if they can be confidentlyidentified early on. This approach as reported byGulati et al appears very promising.47 The obser-vation that 60% of our poor-risk patients achieveda CR suggests that the majority can be effec-tively cytoreduced with chemotherapy. Further-more, when we applied our model to patientsaged _ 50 years, a small subset of patients withsuitably poor prognosis (2-year survival, approxi-mately 20%) could be identified. Thus, analysisof DI within the first 12 weeks of treatment mayallow modification of therapy, including alterna-tive experimental regimens or bone marrow trans-plantation.

Multiple pretreatment clinical factors prognos-tic for survival have been reported in DLCL;however, many have failed to retain significancein subsequent multivariate analyses, and thevariability of factors has led to some confusion in

analyzing and comparing clinical trials.8s' 5 Age,performance status, and factors that reflect tu-mor burden and growth characteristics, such asstage, the number and specific sites of extranodaldisease (eg, marrow, gastrointestinal tract), bulk,and serum LDH, consistently emerge as signifi-cant. Our results are in agreement with thesestudies, although bulk, measured as the largestdiameter of a single mass, and analyzed as acontinuous variable, was not found to be signifi-cant in our analysis. Bulk, measured in this way,was not a significant variable in several otheranalyses. 5 ,'9,11,12, 5 This may relate to the difficultyin assessing this parameter when multiple sites ofbulk are involved. In this regard, a method ofassessing bulk, which takes into considerationmore than one site and accounts for both nodaland extranodal disease, such as that suggested byJagannath et al, may more accurately reflecttumor burden.' 4 In addition, the lack of a signifi-cant adverse effect of bulk on survival in ouranalysis may, in part, be due to the significantproportion of patients with initial bulky diseasewho received radiation therapy following thecompletion of chemotherapy (57%). The signifi-cant survival advantage demonstrated amongpatients who received consolidative radiationtherapy to sites of initial bulky disease supportsthis explanation (Table 3). Age, evaluated as acontinuous variable, also was not significant forsurvival in our analysis. A significant number ofpatients were age 60 or older (29%), althoughonly 11% of the patient group was older than age65. The relatively small number of elderly pa-tients may, in part, explain the lack of impact ofage. Nonetheless, age did emerge as a surrogatecovariate at two points in the tree-structuredsurvival analysis.

The comparison of nonconcurrent groups, vari-ation in length of follow-up, and differences inpatient characteristics among the three chemo-therapy treatment groups may obscure the im-pact of treatment on survival. In our study, nosignificant differences in survival among thethree treatment regimens could be demonstratedby univariate, multivariate recursive partitioningor Cox regression analysis. Valid comparison ofthese three regimens is the goal of a largeprospective randomized study being conductedby the Southwest Oncology Group.

975



KWAK ET AL

Our analysis supports the concept that drugdose is directly related to therapeutic effect, asDI measured during the first 12 weeks of therapywas a strong predictor of overall survival in ourpatients with DLCL. Despite the fact that this isa retrospective analysis that should be inter-preted accordingly, some of the P values in Figs 3to 7 are so striking that we feel that "real"differences have been identified that would be

validated by future data. If these results areconfirmed by others, then actual DI should beincluded in the analysis and reporting of futureclinical trials of DLCL, and considered in theselection of alternative or consolidative therapies.

ACKNOWLEDGMENT

The authors express their gratitude to Saul A. Rosenberg,MD, for critical review of the manuscript.

REFERENCES

1. The Non-Hodgkin's Lymphoma Classification Project:National Cancer Institute sponsored study of classificationsof non-Hodgkin's lymphoma. Summary and description of aWorking Formulation for clinical usage. Cancer 49:2112-2135, 1982

2. Skarin AT, Canellos GP, Rosenthal DS, et al: Improvedprognosis of diffuse histiocytic and undifferentiated lympho-mas by use of high dose methotrexate alternating withstandard agents (M-BACOD). J Clin Oncol 1:91-98, 1983

3. Fisher RI, DeVita VT Jr, Hubbard SM, et al: Diffuseaggressive lymphomas: Increased survival after alternatingflexible sequences of ProMACE and MOPP chemotherapy.Ann Intern Med 98:304-309, 1983

4. Klimo P, Connors JM: MACOP-B chemotherapy forthe treatment of diffuse large-cell lymphoma. Ann InternMed 102:596-602, 1985

5. Laurence J, Coleman M, Allen SL, et al: Combinationchemotherapy of advanced diffuse histiocytic lymphoma withthe six-drug COP-BLAM regimen. Ann Intern Med 97:190-195, 1982

6. Miller TP, Jones SE: Initial chemotherapy for clinicallylocalized lymphomas of unfavorable histology. Blood 62:413-418, 1983

7. Connors JM, Klimo P, Fairey RN, et al: Brief chemo-therapy and involved field radiation therapy for limited stage,histologically aggressive lymphoma. Ann Intern Med 107:25-30, 1987

8. Fisher RI, DeVita VT, Jr, Johnson BL, et al: Prognosticfactors for advanced diffuse histiocytic lymphoma followingtreatment with combination chemotherapy. Am J Med 63:177-182, 1977

9. Jagannath S, Valesquez WS, Tucker SL, et al: Stage IVdiffuse large-cell lymphoma: A long-term analysis. J ClinOncol 3:39-47, 1985

10. Dixon DO, Neilan B, Jones SE, et al: Effect of age ontherapeutic outcome in advanced diffuse histiocytic lym-phoma: The Southwest Oncology Group experience. J ClinOncol 4:295-305, 1986

11. Todd MB, Portlock CS, Farber LR, et al: Prognosticindicators in diffuse large-cell (histiocytic) lymphoma. Int JRadiat Oncol Biol Phys 12:593-601, 1986

12. Danieu L, Wong G, Koziner B, et al: Predictive modelfor prognosis in advanced diffuse histiocytic lymphoma.Cancer Res 46:5372-5379, 1986

13. Shipp MA, Harrington DP, Klatt MM, et al: Identifi-cation of major prognostic subgroups of patients with large-cell lymphoma treated with m-BACOD or M-BACOD. AnnIntern Med 104:757-765, 1986

14. Jagannath S, Valasquez WS, Tucker SL, et al: Tumor

burden assessment and its implication for a prognostic modelin advanced diffuse large-cell lymphoma. J Clin Oncol4:859-865, 1986

15. O'Connell MJ, Harrington DP, Earle JD, et al:Prospectively randomized clinical trial of three intensivechemotherapy regimens for the treatment of advanced unfa-vorable histology non-Hodgkin's lymphoma. J Clin Oncol5:1329-1339, 1987

16. Hryniuk W, Bush H: The importance of dose intensityin chemotherapy of metastatic breast cancer. J Clin Oncol2:1281-1288, 1984

17. Hryniuk W, Levine MN: Analysis of dose intensity foradjuvant chemotherapy trials in stage II breast cancer. J ClinOncol 4:1162-1170, 1986

18. Levin L, Hryniuk WM: Dose intensity analysis ofchemotherapy regimens in ovarian carcinoma. J Clin Oncol5:756-767, 1987

19. Carde P, MacKintosh FR, Rosenberg SA: A dose andtime response analysis of the treatment of Hodgkin's diseasewith MOPP chemotherapy. J Clin Oncol 1:146-153, 1983

20. Murray N: The importance of dose and dose intensityin lung cancer chemotherapy. Semin Oncol 14:20-28, 1987(suppl 4)

21. Palmer M, Belch A, Hanson J, et al: Dose intensityanalysis of melphalan and prednisone in multiple myeloma. JNatl Cancer Inst 80:414-418, 1988

22. DeVita VT, Jr, Hubbard SM, Longo DL: The chemo-therapy of lymphomas: Looking back, moving forward.Cancer Res 47:5810-5824, 1987

23. Carbone PP, Kaplan HS, Musshoff K, et al: Report ofthe committee on Hodgkin's disease staging classification.Cancer Res 31:1860-1861, 1971

24. McKelvey EM, Gottlieb JA, Wilson HE, et al: Hydrox-yldaunomycin (Adriamycin) combination chemotherapy inmalignant lymphoma. Cancer 38:1484-1493, 1976

25. Armitage JO, Weisenburger DD, Hutchins M, et al:Chemotherapy for diffuse large-cell lymphoma-Rapidlyresponding patients have more durable remissions. J ClinOncol 4:160-164, 1986

26. Kaplan EL, Meier P: Non-parametric estimation fromincomplete observation. J Am Stat Assoc 53:457-481, 1958

27. Gehan EA: A generalized Wilcoxon test for compar-ing arbitrarily singly censored samples. Biometrika 52:203-223, 1965

28. Cox DR: Regression models and life tables (withdiscussion). J R Stat Soc [B] 34:187-220, 1972

29. Tsiatis A: A large sample study of Cox's regressionmodel. Ann Stat 9:93-108, 1981

976




30. Gordon L, Olshen RA: Tree-structured survival analy-sis (with discussion). Cancer Treat Rep 69:1065-1069, 1985

31. Breiman L, Friedman JH, Olshen RA, et al: Classifi-cation and regression trees. Belmont, CA, Wadsworth, 1984,pp 75-78

32. Breiman L, Friedman JH, Olshen RA, et al: Classifi-cation and regression trees. Belmont, CA, Wadsworth, 1984,pp 142-143

33. Miller RG Jr: Survival analysis. New York, NY,Wiley 1981, pp 94-126

34. Hryniuk WM: Average relative dose intensity and theimpact on design of clinical trials. Semin Oncol 14:65-74,1987

35. DeVita VT Jr: Dose-response is alive and well. J ClinOncol 4:1157-1159, 1986 (editorial)

36. Dembo AJ: Time-dose factors in chemotherapy: Ex-panding the concept of dose intensity. J Clin Oncol 5:694-696, 1987 (editorial)

37. Hryniuk WM: The importance of dose intensity in theoutcome of chemotherapy, in DeVita VT Jr, Hellman S,Rosenberg SA (eds): Important Advances in Oncology 1988.Philadelphia, PA, Lippincott, 1988, pp 121-141

38. Breiman L, Friedman JH, Olshen RA, et al: Classifi-cation and regression trees. Belmont, CA, Wadsworth, 1984,p181

39. Goldman L, Weinberg M, Weisberg M, et al: Acomputer-derived protocol to aid in the diagnosis of emer-gency room patients with acute chest pain. N Engl J Med307:588-596, 1982

40. Goldman L, Cook EF, Brand DA, et al: A computerprotocol to predict myocardial infarction in emergency depart-ment patients with chest pain. N Engl J Med 318:797-803,1988

41. Levy DE, Caronna MD, Singer BH, et al: Predictingoutcome from hypoxic-ischemic coma. JAMA 253:1420-1426, 1985

42. Gail MH, Muenz L, McIntire KR, et al: Multiplemarkers for lung cancer diagnosis: Validation of models foradvanced lung cancer. J Natl Cancer Inst 76:805-816, 1986

43. Dillman RO, Koziol JA: Statistical approach to immu-nosuppression classification using lymphocyte surface mark-ers and functional assays. Cancer Res 43:417-421, 1983

44. Liebenhaut M, Hoppe RT, Efron B, et al: Prognosticindicators of laparotomy findings in clinical stage I-II supra-diaphragmatic Hodgkin's disease. J Clin Oncol 7:81-91, 1989

45. Philip T, Armitage JO, Spitzer G, et al: High-dosetherapy and autologous bone marrow transplantation afterfailure of conventional chemotherapy in adults with interme-diate grade or high grade NHL. N Engl J Med 316:1493-1498, 1987

46. Takvorian T, Canellos GP, Ritz J, et al: Prolongeddisease-free survival after autologous bone marrow transplan-tation in patients with non-Hodgkin's lymphoma with a poorprognosis. N Engl J Med 316:1499-1505, 1987

47. Gulati SC, Shank B, Black P, et al: Autologous bonemarrow transplantation for patients with poor-prognosislymphoma. J Clin Oncol 6:1303-1313, 1988

977



Prognostic Significance of Actual Dose Intensity in Diffuse Large...

Documents

Transcript of Prognostic Significance of Actual Dose Intensity in Diffuse Large...