Decision-Tree Approach to the Immunophenotype-BasedPrognosis of the B-Cell Chronic Lymphocytic Leukemia

Nikola Mas ic ,1 Alenka Gagro, 2 Sabina Rabatic´ ,2 Ante Sabioncello, 2 Gorana Das ic ,2Branimir Jaksˇ ic ,3 and Branko Vitale 1*

1Division of Molecular Medicine, Ruder Boskovic Institute, Zagreb, Croatia2Department of Cellular Immunology, Institute of Immunology, Zagreb, Croatia

3Department of Internal Medicine, Clinical Hospital Merkur, Zagreb, Croatia

Use of a nonlinear prediction method, such as machine learning, is a valuable choice inpredicting progression rate of disease when applied to the highly variable and correlatedbiological data such as those in patients with chronic lymphocytic leukemia (CLL). In thiswork, decision-tree approach to cell phenotype-based prognosis of CLL was adopted.The panel of 33 (32 different phenotypic features and serum concentration of sCD23)parameters was simultaneously presented to the C4.5 decision tree which extracted themost informative of them and subsequently performed classification of CLL patientsagainst the modified Rai staging system. It has been shown that substantial correlationbetween the percentage of expression of the CD23 molecule on CD19 + B-cells, the levelof sCD23, the percentage of CD45RA +, and the absolute number of CD4CD45RA +RO+

T-cells and the clinical stages, exists. The prediction vector, composed of their concat-enated values, was able to correctly associate 83% of the cases in the low-risk group (Raistage 0), 100% of the cases in the intermediate-risk group (Rai stage I and II), and 89% ofthe cases in the high-risk group (Rai stage III and IV) of CLL patients. Predictivity of thisvector was 100%, 95%, and 89%, respectively. In conclusion, from the described analysis,it may be inferred that two processes play important roles in the progression rate of CLL:1. deregulated function of the CD23 gene in B-cells accompanied by the appearance of itscleaved product sCD23 in the sera; and 2. functionally impaired and imbalanced CD4T-cell subpopulations found in the peripheral blood of CLL patients. Am. J. Hematol.59:143–148, 1998. © 1998 Wiley-Liss, Inc.

Key words: B-CLL; T- and B-cell immunophenotyping; prognosis; decision-tree analysis;sCD23

INTRODUCTION

Chronic lymphocytic leukemia (CLL) is a lymphopro-liferative disorder that can be defined as uncontrolledexpansion of a clone of B-lymphocytes that did not retainthe capacity to differentiate into functionally mature cellsand activate programmed cell death. In consequence,CLL syndrome could be as much a disease of uncon-trolled cell survival as of uncontrolled cell proliferation.In clinical terms, CLL syndrome is characterized by theremarkable variability in clinical presentation, course,and prognosis [1,2]. This underlines the importance ofprognostic factors and staging systems.

Although clinical stage is the strongest predictor ofsurvival, additional prognostic parameters, such as im-munophenotyping, have been identified recently. So far,

a number of immunological markers found on the leu-kemic cell surface and in the serum have contributedsignificantly to our knowledge about the pathogenesisand differential diagnosis of CLL.

The prognostic value of the immunophenotyping inB-cell CLL is rather vague, with many studies yieldingcontradictory results. We have studied the associationbetween 33 parameters derived from appropriate immu-nological cell markers on both B- and T-cells and clinicalstages in patients with CLL.

In this study, machine-learning approach to the pre-

*Correspondence to: Dr. Branko Vitale, Ruder Boskovic Institute,P.O. Box 1016, 10000 Zagreb, Croatia.

Received for publication 12 February 1998; Accepted 10 June 1998

American Journal of Hematology 59:143–148 (1998)

© 1998 Wiley-Liss, Inc.

diction task was adopted. C4.5 decision tree was used topredict expression of various markers on B- and T-cellsin CLL patients against the modified Rai staging system.A learning system is a computer program that makesdecisions based on accumulated experience contained insuccessfully solved cases. One of the main reasons it is agood candidate to be applied in biomedicine is its abilityto present solutions in a form compatible with typicalhuman reasoning. In this way, production rules, that issolutions to the system, can be carefully evaluated as towhat is particularly important in situations where deci-sions are likely to have critical consequences.

MATERIALS AND METHODSPatient Characteristics

Thirty-four nontreated patients diagnosed with B-CLL—20 males and 14 females with median age of 66years—formed the data basis of this work. The range was46–93 years (17% of the patients were under 50 years).The diagnosis was based on commonly accepted criteria,and subsequent staging was done according to the Na-tional Cancer Institute working group (WG) recom-mended ‘‘three-risk group’’ modification of the originalfive-stage Rai staging system [2,3]. The original fivestages were reduced into three groups: the low-risk group(Stage 0), with only lymphocytosis (blood and marrow);the intermediate-risk group, with lymphocytosis and en-larged nodes (Stage I) and/or enlarged spleen/liver (StageII); and the high-risk group, with lymphocytosis and ane-mia, (defined as hemoglobin <11 mg% [Stage III], and/orplatelets <100,000 mm3 [stage IV]). The median life ex-pectancy of patients in the Rai low-risk group is 14 years,intermediate risk, eight years, and high risk, about fouryears. Following the staging of the patients, heparinizedvenous peripheral blood and serum samples were taken.

Laboratory Analyses

The level of the sCD23, expressed inmg/L, was mea-sured using an immunoenzyme assay sCD23 kit (BindingSite, Birmingham, England).

Three-color immunofluorescence with monoclonal an-tibodies (mAbs) (Becton-Dickinson, ImmunocytometrySystem, San Jose, CA) was performed on whole bloodsamples as described earlier [4]. A panel of used fluoro-chrome-conjugated mAbs combinations is presented inTable I.

Measurements were performed with a FACScant cy-tometer (Becton Dickinson, Mountain View, CA) equippedwith a 488 nm Argon ion laser. Data were acquired andanalyzed by CELLQuest software. At least 20,000 events/sample were collected and lymphocyte gate was set on thebasis of forward vs. side scatter parameters.

The results were expressed as percentages of singlepositive (CD19+, CD20+, CD21+, CD23+, CD72+, CD5+,

CD3+, CD4+) or double positive (CD19+CD5+,CD19+CD23+, CD19+CD21+) cells in the lymphocytegate. Analyses were further made for percentages ofCD19+ cells coexpressing CD5, CD23, CD21, or CD72;percentages of CD4+ cells coexpressing CD45RO orCD45RA; percentages of CD4+ cells with phenotypeCD45RA+RO−, CD45RA+RO+, CD45RA−RO+; percent-ages of CD19+ cells with phenotype CD21+CD23−,C D 2 1+ C D 2 3+ , C D 2 1− C D 2 3+ , C D 5+ C D 7 2− ,CD5+CD72+, CD5−CD72+; and percentages ofCD19+CD5+, CD19+CD5−, CD19+CD21+ orCD19+CD21− cells coexpressing CD23. Finally, the ab-solute number of CD4+CD45RO+RA+, as well asCD4+CD45RA+/CD4+CD45RO+ indexes were calcu-lated. With soluble sCD23 found in the serum, thesemade a total of 33 variables loaded into the C4.5 machinelearning program to generate a decision tree.

It should be noted that not every variable was deter-mined for each case, so we had to deal with missing dataas well.

Decision Tree

The program C4.5 generates a classifier in the form ofa decision tree, not necessary binary, a structure withelements that are either leaves or decision nodes [5]. Theleaf shows a class and the decision node specifies sometest to be implemented on an attribute value, with onebranch and subtree for each possible result of the test.The starting node is usually referred to as the root node.A decision tree is used to predict a case by starting at theroot of the tree and moving through it until a leaf isencountered.

The C4.5 algorithm presumes existence of an appro-priate number of learning examples described by a set ofattributes and by classes representing conditions. Itsearches for the most informative attribute according tothe gain criterion and constructs a decision tree. Thissearch is based on Shannon’s measure of information.Pruning is used to reduce the decision tree i.e., producinga more comprehensible structure without compromisingaccuracy on unseen cases.

On the Reporting of Results of Classification

The objective of learning classifications from sampledata is to classify and predict successfully on new unseendata. A commonly used measure of performance good-

TABLE I. Panel of Monoclonal Antibodies Used forImmunophenotypic Analysis by Three-Color Flow Cytometry

IgG1 FITC IgG2a PE IgG1 PerCPCD5 FITC CD72 PE CD19 PerCPCD5 FITC CD23 PE CD19 PerCPCD21 FITC CD23 PE CD19 PerCPCD5 FITC CD3 PE CD20 PerCPCD45RA FITC CD45RO PE CD4 PerCP

144 Masic et al.

ness is a classifier’s error rate. If distinguishing amongerror types is important, then a confusion matrix can beused to lay out the different errors. It lists the correctclassification against the predicted classification for eachclass. In medical applications, classifier performance isusually measured by computing frequency ratios fromthe elements of a confusion matrix. Relative accuracy orsensitivity can be defined as the ratio of the number oftrue predictions to the size of the appropriate true class.It measures ability to classify correctly the case that ac-tually belongs to the appropriate true class. Predictivitycan be defined as a ratio of the number of true predictionsto the size of the appropriate prediction class. It measuresreliability of the classifier predictions.

The best method for evaluating classifier performanceon small samples is the leaving-one-out error estimationtechnique. For a given method and sample size (n), aclassifier is generated using (n-1) cases and tested on thesingle remaining case. This is repeated n times, each timedesigning a new classifier by leaving one out. This way,each case in the sample is used for testing, and each timenearly all the cases are used for learning (i.e., designinga classifier). The error rate is averaged out.

For any learning system, the performance on data mustexceed that of simply choosing the largest, most fre-quently occurring class. Otherwise we must concludethat we are dealing with random noisy data.

RESULTS

To evaluate the degree of association between surface-membrane immunophenotype(s) and clinical stage(s) inpatients with B-CLL, we studied prediction of concate-nated expression of various markers on B- and T-cellsand presence of soluble sCD23 in their sera against themodified Rai staging system. The results obtained bymeans of the C4.5 algorithm applied to the panel of 33variables are shown in Figure 1.

The decision rules that are thus produced can be readas follows. If the concentration of soluble sCD23 in thesera is equal or less than 69mg/L, then the patient is inthe low-risk group. If this rule is not satisfied, then thepercentage of CD23+ B-cells has to be taken into con-sideration. Thus, if the percentage of CD23+ B-cells ishigher than 92 and the concentration of sCD23 is higherthan 69mg/L, then the patient is in the high-risk group.However, when the percentage of CD23+ B-cells is be-low 92 in CLL patients, then the percentage of CD4+

T-cells coexpressing the CD45RA molecule also needsto be taken into consideration. Therefore, other rules pro-duced by the decision tree can be deduced as well. Thus,if the sCD23 serum concentration is higher than 69mg/L,the percentage of CD23+ B-cells is lower than 92, andthe percentage of CD4+ T-cells coexpressing CD45RA+

is equal to or less than 19, then patient is again in the

high-risk group. In addition, if the sCD23 serum concen-tration is higher than 69mg/L, the percentage of CD23+

B-cells is 92 or lower, the percentage of CD4+ T-cellscoexpressing CD45RA+ is higher than 19, and the abso-lute number of CD4+CD45RO+RA+ T-cells is higherthan 1.2 × 109 L−1, then patient is also at high risk.However, when the sCD23 serum concentration is higherthan 69mg/L, the percentage of CD23+ B-cells is equalto or lower than 92, the percentage of CD4+ T-cells co-expressing CD45RA+ is higher than 19, and the absolutenumber of CD4+CD45RO+RA+ T-cells is equal to orlower than 1.2 × 109 L−1, then the CLL patient is in theintermediate-risk stage. Needless to say, such predictionresults cannot be obtained based on any of the parametersalone.

Performance of this decision tree was estimated oncross validation and the leaving-one-out method. Overallprediction results (predictivity and sensitivity for predic-tion vector composed of concatenated sCD23, the per-centage of CD23+ B-cells and CD4+CD45RA+ T-cells,and the absolute number of CD4+CD45RO+CD45RA+

T-cells) derived from the confusion matrix are summa-rized in Table II.

DISCUSSION

Staging systems that appeared in the late seventies andeighties based on the multivariate Cox regression analy-sis of various conventional clinical parameters havemade it possible to predict the survival time of CLLpatients with different stages of the disease [6–8]. Nev-ertheless, all developed staging systems, which haveproved to be of great practical value, are still limited intheir ability to predict the progression rate of the disease.Besides methodological limitations, there are othersources of poor classification and prediction performancethat may further restrict their validity. Some of them aredata errors, poor features, mislabeled classes, and non-representative samples. A special source of poor predic-tion performance can be attributed to the missing data aswell.

In recent years significant progress has been made inimmunological and molecular biology methodologiesthat opened new vistas on the complex CLL pathogen-esis. These breakthroughs resulted in the appearance oftwo newly proposed concepts of CLL. One is based onthe functionally impaired and imbalanced T-cell sub-populations accompanied by uneven profiles of cytokinesecretion that may lead to the unequally matched activa-tion of various oncogenes, defective differentiation, andblockade of programmed cell death that may end up inthe eventual emergence of a malignant B-cell clone [9],and a second one that is based on the assumption that theCLL syndrome represents an anti-self B-cell malignancy[10].

Decision-Tree Approach in B-CLL Prognosis 145

A search for a cellular and/or molecular event (orevents) that could be responsible for a low- or high-riskform of CLL is a basic prerequisite for better understand-ing of its pathogenesis. The most appropriate way toelucidate possible significance of these events is to de-

termine their prognostic value; namely, an event that isassociated with prognosis, is also very likely to be asso-ciated with the pathogenesis of the disease.

Due to the variability of the clinical and laboratorydata that could stem from the measurement technique

Fig. 1. Decision tree of (sCD23, CD23 +, CD4+CD45RA+, a CD4+CD45RO+RA+) prognostic vector; numbers associated witheach branch represent test values used with the decision rule and applied to the appropriate decision node. Prognosis wasbased on the modified Rai clinical staging system. The prognostic vector component values are given as: concentrationof sCD23; percentage of CD23 + lymphocytes; percentage of CD4 + cells coexpressing CD45RA + (CD4+CD45RA+); andabsolute number of CD4 + cells coexpressing CD45RO +RA+ (aCD4+CD45RO+RA+). The inserted plots show the distributionof results for respective variables with the attribute values indicated by horizontal lines. The nonlinear relation between testresults and modified Rai staging is evident.

146 Masic et al.

and from the heterogeneity of the disease itself, determi-nation of the degree of association between various im-munobiological and clinical stages in patients with CLLhas been poor. However, inconsistent or even contradic-tory results published in the literature came as no surprisebecause obviously classical regression analysis was notable to tackle the heterogeneous and nonlinear pattern ofdisease.

In an attempt to overcome methodological limitationsof classical linear models and the fact that the CLL pat-tern is probably nonlinear, we used a nonlinear machine-learning approach to the analysis and prediction of clini-cal and laboratory CLL data. Having performed thoroughdata quality control, we presumed that prediction perfor-mance depends on predictive capabilities of the chosenprediction vector only. If the particular prediction vectorenables good predictive performance of the appropriatedecision tree, it can be said that it is a good prognosticparameter.

Each result obtained with the classifier/predictorshould be compared to the default value, otherwise theresult achieved would not give much information aboutpredictivity but rather about the class distribution. De-fault classifier/predictor means the predictor that wouldclassify all cases as the majority class cases or accordingto the initial class distribution of cases.

In this study comprising 32 different phenotypic pa-rameters, only three, along with soluble sCD23, emergedas having prognostic value. As can be seen in Table 2, theprediction vector, consisting of concatenated parameters(sCD23, CD23+, CD4+CD45RA+, aCD4+CD45RO+RA+),was able to correctly associate 83% of the cases in thelow-risk group, 100% of the cases in the intermediate-risk group, and 89% of the cases in the high-risk group ofCLL patients. Its predictivity for the low-, intermediate-

and high-risk groups was 100%, 95%, and 89%, respec-tively. Overall prediction accuracy was 94%. Quantita-tive threshold values can be obtained from Figure 1 asappropriate decision node test values associated with re-spective branches.

To date, several studies have been done that attemptedto elucidate the importance of the membrane CD23 mol-ecule expressed on the B-cell surface and its soluble form(sCD23) in the sera of CLL patients as a disease markerwith possible prognostic value [11–16]. Due to gene de-regulation, the CD23 surface molecule, active in signaltransduction, is overexpressed on virtually all leukemicB-cells [17]. In addition, the existence of two isoforms ofthe CD23 molecule, which seem to be regulated differ-ently, have been shown by other studies of CLL patients.Thus, it was found that interferon-alpha (IFN-a) and in-terleukin (IL)-2 selectively up-regulate the expression ofthe CD23b isoform, which may stimulate B-cell growth,in contrast to IL-4 and IFN-g which selectively up-regulate the CD23a isoform which in turn suppressesB-cell apoptosis [18]. These data further illustrate howthe imbalance among Th1- and Th2-like cytokine profilesmay influence the biological features of leukemic B-cells.

At present, the prognostic value of the CD23 moleculeis unclear. In one study, a high percentage of CD23+

B-cells in CLL patients was associated with a good prog-nosis [12], whereas in other studies, the opposite conclu-sion was reached [13–15]. Our results show clearly thatboth the B-cell CD23 marker expression and solublesCD23 have good prognostic capabilities. However, ourfindings only partly agree with other findings that showthat sCD23 concentration in sera of CLL patients in-creases with advancing stages of the disease [16]. Usingdecision tree generated with sCD23 as a prediction vec-tor, we can say that sCD23 sera values less then 69mg/lare hallmark of the low-risk group of CLL. Thus, a lowerconcentration of sCD23 is related to low risk, whereashigher concentration of sCD23, together with a high fre-quency of the membrane CD23 molecule expression onB-cells, is associated with the high-risk group of CLL.However, if the high concentration of sCD23, togetherwith a lower-than-threshold percentage of CD23+ B-cellsis encountered, then the percentage of CD45 moleculeisoforms on CD4+ T-lymphocytes should be added forfurther classification.

Two restricted (R) isoforms of the CD45 molecule,CD45RA and CD45RO, are expressed on ‘‘naive’’ and‘‘memory’’ T-lymphocytes respectively. By in vitrostimulation, CD45RA+ T-cells mature into CD45RO+ T-cells, achieving not only the phenotype but also the func-tion of memory cells [19–21]. Activated CD45RA+ cellsproduce IL-2, whereas the CD45RA+Ro+ subset, appear-ing three days after activation, acquire a lymphokine pro-file similar to that of CD45RO+ T-cells, e.g., IL-4,

TABLE II. Classification Results Obtained With a (sCD23,CD23+, CD4+CD45RA+, aCD4+CD45RO+RA+)Prognostic Vector*

(%)

Prognostic vector

(sCD23, CD23+, CD4+CD45RA+,CD4+CD45RO+RA+)

[6,19,9]

Accuracy 94Low-risk sensitivity 83Medium-risk sensitivity 100High-risk sensitivity 89Low-risk predictivity 100Medium-risk predictivity 95High-risk predictivity 89

*Prognosis was based on the modified Rai clinical staging system. Thefield in the brackets designates distribution of cases among classes i.e., themodified Rai stages. The prognostic vector component values are given as:concentration of sCD23; percentage of CD23+ lymphocytes; percentage ofCD4+ cells coexpressing CD45RA+ (CD4+CD45RA+); and absolute num-ber of CD4+ cells coexpressing CD45RO+RA+ (aCD4+CD45RO+RA+).

Decision-Tree Approach in B-CLL Prognosis 147

IFN-g, IL-5, and IL-2 [19,21]. In the case of a low pro-portion (ø19%) of naive, CD45RA+ CD4+ T-lymphocytes in CLL patients with lower than thresholdCD23 molecule expression on B-lymphocytes, higherproduction of IL-4 and IFN-g by CD45RO+ T-cells couldbe accompanied by suppression of apoptosis [18], andthis is related to high-risk CLL. If more than 19% of theCD45RA+ CD4+ T-lymphocytes producing IL-2 are en-countered, then the absolute number of transientCD45RA+RO+, CD4+, T-lymphocytes is decisive. In thecase of their lower number (ø1.2 × 109/L), IL-2-promoted proliferation could prevail over IFN-g-promoted suppression of apoptosis, and the intermediate-risk CLL would be the consequence. However, in thecase of a high proportion of CD45RA+ CD4+ T-lymphocytes accompanied by a high number ofCD45RA+RO+ CD4+ T-lymphocytes, IL-2 and IL-4 mayact synergistically, and high-risk CLL would be the con-sequence.

In conclusion, irrespective of the fact that only 34 CLLpatients were included in this study, the nonlinear deci-sion-tree method was able to extract four of 33 offeredvariables as being relevant for disease prognosis. A pre-diction vector composed of four concatenated variableswas able to correctly associate 83% of the cases in thelow-risk group, 100% of the cases in the intermediate-risk group, and 89% of the cases in the high-risk group.From this type of analysis two processes emerged asbeing crucial in the pathogenesis of CLL: 1. deregulatedfunction of the CD23 gene active in signal transductionin B-cells accompanied by the appearance of its cleavedproduct(s) (sCD23) in the sera; and 2. functionally anddevelopmentally impaired and imbalanced CD4+ T-cellsubpopulations.

REFERENCES1. Foon KA, Rai KR, Gale RP: Chronic lymphocytic leukemia: New

insights into biology and therapy. Ann Intern Med 113:529–539, 1990.2. Cheson BD, Bennett JM, Rai KR, Grever MR, Kay NE, Schiffer CA,

Oken MM, Keating MJ, Bold DH, Kempin SJ: Guidelines for clinicalprotocols for chronic lymphocytic leukemia: Recommendations of theNational Cancer Institute sponsored working group. Am J Hematol29:152–163, 1988.

3. Rai KR: A critical analysis of staging in CLL. In: Gale RP, Rai KR,eds. Chronic Lymphocytic Leukemia: Recent Progress and Future Di-rections. New York: Alan R. Liss, 1987, pp 252–264.

4. Rabatic´ S, Gagro A, Medar-Lasic´ M: CD21-CD23 ligand pair expres-sion in children with allergic asthma. Clin Exp Immunol 94:337–340,1993.

5. Quinlan JR: C4.5: Programs for Machine Learning. San Francisco:Morgan Kaufman Publishers, 1993.

6. Binet JL, Auquier A, Dighiero G, Chastang C, Piguet H, Goasguen J,Vaugier G, Potron G, Colona P, Oberling F, Thomas M, Tchernia G,Jacquillat C, Boivin P, Lesty C, Cuault MT, Monconduit M, BelabbesS, Gremy F: A new prognostic classification of chronic lymphocyticleukemia derived from a multivariate survival analysis. Cancer 48:198–206, 1981.

7. International Workshop on CLL: Chronic lymphocytic leukaemia:Proposal for a revised prognostic system. Br J Haematol 48:365–367,1981.

8. Jaksˇic B, Vitale B: Total tumor mass score (TTM): A new parameterin chronic lymphocytic leukaemia. Br J Haematol 49:405–413, 1981.

9. Vitale B, Kusic B, Antica M, Poljak LJ, Gagro A, Rabatic´ S, Jakšic´ B:Chronic lymphocytic leukemia (CLL): A paradigm for developmentalimbalance in B-cell and T-cell communication. Period Biol 97:273–280, 1995.

10. Caligaris-Cappio F: B-chronic lymphocytic leukemia: A malignancyof anti-self B cells. Blood 87:2615–2620, 1996.

11. Geisler CH, Larsen JK, Hansen NE, Hansen MM, Christensen BE,Lund B, Nielsen H, Plesner T, Thorling K, Andersen E, Andersen PK:Prognostic importance of flow cytometric immunophenotyping of 540consecutive patients with B-cell chronic lymphocytic leukemia. Blood78: 1795–1802, 1991.

12. Dadmarz R, Cawley JC: Heterogeneity of CLL: High CD23 antigenanda-IFN receptor expression are features of favorable disease and ofcell activation. Br J Haematol 68:279–282, 1989.

13. Barnett D, Reilly JT: Lack of correlation between cell surface activa-tion antigen expression and clinical stage in B-CLL. Br J Haematol73:572, 1989.

14. Gibson J, Neville S, Joshua D, Kroninberg H: CD23 antigen expres-sion in CLL. Br J Haematol 72:598, 1989.

15. Newman RA, Peterson B, Davey FR, Brabyn C, Collins H, BrunettoVL, Duggan D, Weiss RB, Royston I, Millard FE, Miller AA, Bloom-field CD: Phenotypic markers and bcl-1 gene rearrangements in B-cellchronic lymphocytic leukemia: A Cancer and Leukemia Group BStudy. Blood 82:1239–1245, 1993.

16. Reinisch W, Willheim M, Hilgarth M, Gasche C, Mader R, SzepfalusiS, Steger G, Berger R, Lechner K, Boltz-Nitulescu G, SchwartzmeierJD: Soluble CD23 reliably reflects disease activity in B-cell chroniclymphocytic leukemia. J Clin Oncol 12:2146–2152, 1994.

17. Fournier S, Rubio M, Delespesse G, Sarfati M: Role of low-affinityreceptor for IgE (CD23) in normal and leukemic B-cell proliferation.Blood 84:1881–1886, 1984.

18. Bonnefoy JY, Lecoanet-Henchoz S, Aubry JP, Gaushat JF, Graber P:CD23 and B-cell activation. Curr Opin Immunol 7:355–359, 1995.

19. Kristensson K, Borrebaeck CAK, Carlsson R: Human CD4+ T cellsexpressing CD45RA acquire the lymphokine gene expression ofCD45RO+ T-helper cells after activation in vitro. Immunology 76:103–109, 1992.

20. Frolova EA, Scott SC, Jones RA: CD45RO+ T-cells immunoregulatespontaneous in vitro immunoglobulin production by normal andchronic lymphocytic leukaemia B-cells. Leuk Lymphoma 18:103–111,1995.

21. Hamann D, Baars PA, Hooibrink B, van Lier RW: Heterogeneity ofthe human CD4+ T-cell population: Two distinct CD4+ T-cell subsetscharacterized by coexpression of CD45RA and CD45RO isoforms.Blood 88:3513–3521, 1996.

148 Masic et al.