UNDERSTANDING CLINICAL TRIALS
OSS9-85SS/OO $15.00
+ .OO
PROGNOSTIC FACTORS FOR HEMATOLOGIC CANCERS R. Rita Misra, PhD, MPH, Paul F. Pinsky, PhD, MPH, and Sudhir Srivastava, PhD, MPH
Cancer is the second leading cause of death in the United States and the fifth leading cause of death worldwide. It is generally accepted that as developing nations become more urbanized, their patterns of mortality become more similar to those of economically developed countries. Thus the worldwide burden of cancer is expected to increase dramatically in the coming years. In the United States, the average annual cancer incidence rate is estimated at 405 per 100,000, and the average annual cancer mortality rate is estimated at 170 per 100,000. For hematologic cancers (including leukemias, lymphomas, and myelomas), the average annual incidence rate is estimated at 33 per 100,000, and the average annual mortality rate is estimated at 17 per 100,000.48Although overall cancer incidence and mortality rates have been declining since 1992, incidence rates for certain cancers have risen markedly during the same period. For example, between 1973 and 1996 there was an 81% increase in the incidence of non-Hodgkin’s lymphoma, a 24% increase in the incidence of acute lymphocytic leukemia, and a 14% increase in the incidence of multiple myeloma. Between 1973 and 1996, mortality rates for non-Hodgkin’s lymphoma and multiple myeloma rose by 44% and 36%, respectively.4s It is clear that development of novel strategies to decrease both the incidence and mortality of these diseases is needed. Identification of good prognostic factors is likely to be an important part of the process. Prognostic factors can help a clinician choose the therapy that is From the Cancer Prevention Fellowship Program (a Early ) Detection , Research Group (PFP), and Cancer Biomarkers Research Group (SS), Division of Cancer Prevention, National Cancer Institute, Bethesda, Maryland HEMATOLOGY/ ONCOLOGY CLINICS OF NORTH AMERICA VOLUME 14 * NUMBER 4 . AUGUST 2000
907
best suited to an individual patient’s needs, with the goal of achieving the best possible clinical outcome. Prognostic factors may also be useful in idenhfymg critical steps in the process of carcinogenesis. A factor is considered prognostic if it is associated with tumor development or progression either in the absence or in the application of a specific therapy. Some researchers have made the distinction between factors that predict the natural history of the disease and outcomes under a standard treatment and those that predict response to a new therapy; they designate the former as prognosticfactors and the latter as predictive factors.34 In this article, the term prognostic factor is used for both. The term prognostic biomarker refers to a biochemical or genetic alteration detected in a body tissue or fluid that directly or indirectly reflects an important step in the disease process.34 This article reviews the molecular biology of hematologic cancers and the current understanding of prognostic factors for these cancers. Some specific molecular biomarkers that have potential as prognostic factors for various hematologic cancers are discussed. Finally, an overview of quantitative and statistical ways of evaluating the utility of prognostic factors is presented.
MOLECULAR PATHOGENESIS OF HEMATOLOGIC MALIGNANCIES
The process of carcinogenesis involves the accumulation of molecular changes which do not always result in clinical disease. Thus, the ability to detect molecular changes that have been associated with particular steps in disease progression is likely to be crucial for fine-tuning cancer prognoses. Although the molecular events involved in the progression of hematologic cancers are poorly understood, two main developmental pathways have been described.18The first pathway is believed to lead to the development of acute onset, clinically aggressive malignancies such as de novo acute leukemia or high-grade lymphoma, which are usually responsive to treatment and are potentially curable. The cytogenetic abnormalities in such malignancies are believed to be causally linked to the cancer, are often highly specific, and are usually balanced (i.e., formed by reciprocal chromosomal translocations or inversions without obvious gain or loss of genetic material). In contrast, the second pathway is believed to lead to malignancies such as secondary leukemia or myelodysplastic syndrome which are often unresponsive to therapy and tend to be relentlessly progressive. Malignancies of the second type are often characterized by the accumulation of multiple, nonspecific genetic lesions which are believed to contribute to disease progression rather than initiation. Typically, such events are imbalanced (i.e., result in a gain or loss of chromosomal material).
PROGNOSTIC FACTORS FOR HEMATOLOGIC CANCERS
909
Leukemia
The leukemias are a family of malignancies resulting from a genetic alteration in one or more cells in the bone marrow. As the leukemic cells expand in number, the growth of normal blood cells is suppressed, resulting in clinical disease.36The five main types of leukemia are acute lymphocytic leukemia (ALL), chronic myelocytic or granulocytic leukemia (CML or CGL), acute myelocytic or acute nonlymphocytic leukemia (AML or ANLL), chronic lymphocytic leukemia (CLL), and adult T-cell leukemia (ATL).45Almost all information regarding the pathogenesis of leukemia has come from studies of AML. Based on certain clinical and cytogenetic features and the patient's history of exposure to specific carcinogens, AML cases may be classified as de novo or secondary. The most common chromosomal rearrangements in de novo AML, t(8;21), t(3;21), inv(l6), t(16;16), and t(12;21), involve chromosomes 21 or 16 and affect the acute myeloid leukemia 1 (AML1)-core binding factor P (CBFP) heterodimeric transcription factor complex. The AML1-CBFP complex seems to function as an important regulator of early hematopoie s i ~ .6,~37, Similarly, specific chromosomal translocations involving the retinoic acid receptor (RAR-a)locus on chromosome 17, t(15;17)(q22;qll12), t(ll;l7)(q23;qll-12), t(5;17)(q35;qll-12) and t(11;17)(q13;q21), have been associated with acute promyelocytic leukemia.** Most of the knowledge about the pathogenesis of secondary AML comes from studies of cancer survivors who were exposed to large doses of hematotoxic chemotherapeutic agents. More than 90% of patients with classic chemotherapy-related AML exhibit loss of the long arm or all of chromosome 5 or 7.". 43 It has been assumed that chemotherapy for the primary tumor results in the loss of only one allele of a suppressor gene on chromosome 5 or 7, and that the mutation of a second allele occurs during the characteristically long (3- to 10-year) latency period. The putative tumor suppressor gene on chromosome 5 or 7 has not yet been identified, however. The mechanism of leukemogenesis in AML after treatment with topoisomerase I1 inhibitors is thought to differ from those involved in classic chemotherapy-related AML. The predominant aberration in this subtype seems to involve a balanced translocation rather than a deletion. Topoisomerase 11-induced AML is most often associated with rearrangements of the MLL gene located at position 923 of chromosome 11. Translocations of the AMLl gene at position 922 of chromosome 21 have also been reported.4I Leukemia is the most common cancer diagnosed in children, and ALL alone accounts for about 76% of all childhood leukemias. Like topoisomerase 11-induced adult AML, infant ALL is characterized (> 80% of cases) by chromosomal rearrangement of the MLL gene on chromosome 11q23.9 For this reason, maternal exposure to a number of natural and synthetic topoisomerase I1 inhibitors is associated with the disease.55 Several syndromes are characterized by an increased incidence of leukemia, including Down syndrome, which is characterized by trisomy
of the whole of chromosome 21 or its distal and9; Fanconi’s anemia, a hereditary DNA-repair disorder associated with increased susceptibility to chemical cross-linking agents13;and childhood monosomy 7, which is characterized by the loss of part or all of chromosome 7.40 Although pure familial leukemia is considered to be extremely rare, linkage to different chromosomal sites has been rep0rted.3~ Interestingly two of the loci believed to contain the putative familial leukemia susceptibility gene also contain genes involved in de novo leukemia. These are the CBFol gene, which maps to 21422, and the CBFP gene, which maps to 16q22. Adult T-cell leukemia (ATL) is a unique malignancy of T cells first described in Japan in 1977. Although the mechanism of oncogenesis in ATL is still poorly understood, the TAX protein, one of the proteins encoded by the human T-cell leukemia virus-1 (HTLV-I)provirus, seems to play a critical role in the It is thought that TAX alters the expression of various genes involved in cell growth?*, Although it does not bind directly to the enhancer regions of target genes, TAX seems to interact with several proteins which do bind to enhancer sites. Such intermediate proteins include CREB, CREM, and ~ 6 7 sp~ 6. F is involved in the serum activation pathway of the nuclear proto-oncogenes c-fos and c-eg. In addition, TAX may be indirectly involved in the activation of the nuclear transcription factor NF-KBand directly involved in the inactivation of p16. p16 is a known regulator of the retinoblastoma (RB) tumor suppressor gene. Other factors such as the host immune capacity, hereditary predisposition, and accumulation of secondary oncogenic events are also thought to be important in the development of ATL. Hodgkin’s Disease
Hodgkin’s disease is a form of cancer involving the lymphatic system. There are several histopathologic subtypes of this disease that suggest different causalities. Childhood Hodgkin’s disease is usually of the mixed cellularity or lymphocyte depletion subtype and is associated with a higher frequency of Reed-Sternberg cells, a feature that is generally associated with poorer prognosis. (Reed-Sternberg cells are large, binucleated cells which contain large eosinophilic nucleoli. They are believed to originate from activated or progenitor B cells.=) Early descriptions of Hodgkin’s disease noted the frequent occurrence of fever and enlarged cervical lymph nodes which suggested an infectious cause for this disease. Additional evidence of an infectious basis for young-adult Hodgkin’s disease include the geographic patterns of this disease.lo Thus far, two viruses have been linked to Hodgkin’s disease: Epstein-Barr virus (EBV) and human herpesvirus-6 (HHV6). Of particular interest is the association between Hodgkin’s disease and infectious mononucleosis, a disease known to be caused by EBV and, rarely, by other herpesviruses such as HHv6.29It is not known, however, whether these associations are a direct result of viral infection or of the
PROGNOSTIC FACTORS FOR HEMATOLOGIC CANCERS
911
accompanying immunodepression. Although certain genetic markers have been associated with Hodgkin’s disease,20genetic susceptibility does not seem to be of major i m p ~ r t a n c e . ~ ~
Non-Hodgkin’s Lymphoma
Like Hodgkin’s disease, non-Hodgkin’s lymphomas (NHL) are cancers which develop in lymphocytes. Non-Hodgkin’s lymphomas are characterized by abnormal growth of lymphocytes in the lymph nodes, spleen, thymus, bone marrow, tonsils, stomach, small intestine, or skin.66 Immunodeficiency is a well-established predisposing factor for NHL. Persons with a high risk of developing NHL include those with hereditary immunodeficiency disorders, such as ataxia-telangiectasia, Wiskott-Aldrich syndrome, and severe combined immunodeficiency, and those with acquired immunodeficiency states associated with HIV infection and immunosuppressive chemotherapies.18The two viruses strongly implicated in lymphomagenesis associated with HIV infection are EBV and human herpesvirus-8 (HHV8). Cytogenetic abnormalities are present in almost all cases of NHL. In fact, acquisition of chromosomal aberrations is thought to be necessary for the malignant phenotype. The nonrandom aberrations seen in NHL are strongly associated with specific histologies and immunophenotypes. Activation of various cellular oncogenes seems to be the most common molecular event.18 This association is in marked contrast with solid tumors and myeloid malignancies, in which the loss of tumor suppressor genes or the creation of chimeric genes are more commonly observed. In NHL, oncogene activation is believed to occur through somatic mutation, chromosomal translocation, or incorporation of viral oncogenes or oncogene enhancers into the host genome. In this context, the two best characterized oncogenes involved in NHL are c-myc and bcZ-2. The c-myc gene, located at 8q24, is often involved in translocations with immunoglobulin-encoding genes such as an immunoglobulin heavychain gene at 14q32; a K light-chain gene at 2p12; and a A light-chain gene at 22pll. These translocations can lead to c-myc dysregulation and are consistently seen in high-grade, small, non-cleaved-cell lymphomas.18Although the precise mechanism is poorly understood, it is believed that the tumorigenic activity of c-myc may be related to translocation-induced changes in the level of expression of the c-myc gene product.44BcZ-2 is an apotosis regulator gene located at 18q21. The most common translocation in NHL (found in 85% of follicular lymphomas) places bcZ-2 next to the heavy-chain immunoglobulin locus at 14q32. This translocation results in altered bcZ-2 expression. It is believed that overexpression of bcZ-2 blocks apotosis of affected cells, which renders them more susceptible to secondary genetic instability and ultimately results in complete malignant transformati~n.~~
Multiple Myeloma
Multiple myeloma is a cancer affecting plasma cells in the bone marrow. The malignant plasma cells are derived from a single B cell that has undergone antigen selection and somatic hypermutation in the lymph node but has not yet undergone immunoglobulin isotype class Once they reach the bone marrow, the primary characteristic of these malignant cells is excess production of a specific immunoglobulin type or subunit. This excess production eventually leads to bone destruction and interferes with the production of red blood cells, white blood cells, and platelets. Although both genetic and environmental factors seem to be involved,% little is known about the cause of multiple myeloma. It is thought that prolonged stimulation of the immune system by repeated infections, allergic conditions, or autoimmune disorder may increase the risk of this disease, but convincing evidence for such mechanisms is lacking. Like most cancers, the clinical characteristics of multiple myeloma (a long progressive phase followed by a final aggressive one) suggest that a multistep process is involved in disease progression. Recent experimental results also suggest that chromosomal rearrangement at 14q32.33, involving a translocation of various protooncogenes into the IgH gene, may be a critical event in the pathogenesis of multiple myel0rna.4~ PROGNOSTIC FACTORS FOR HEMATOLOGIC MALIGNANCIES
The classification scheme for a particular cancer must cover all the characteristics associated with the various stages of tumor growth and development. The American Joint Committee on Cancer (AJCC) and Union Internationale Contre le Cancer (UICC) classification systems are based on the premise that cancers of the same anatomic site and histology share similar patterns of growth and extension? In most cases, the primary tumor increases in size, then spreads into the lymph nodes that drain the area of the tumor or invades other anatomic sites by infiltrating blood vessels. These potential events in the life history of a cancer-local tumor growth (T), spread to regional lymph nodes (N), and metastasis (M)-are used in clinical examination to assess the extent or stage of the cancer. Additional factors can be used to predict disease progression if they are clearly associated with any one of these events. Because spread to the lymph nodes and metastasis usually occur before they are clinically observable, examination during surgical procedures and histologic examination of surgically removed tissues are often necessary to identrfy additional prognostic factors. The pathologic classification (pTNM) is recorded with the clinical classification, and both are maintained in the patient's permanent medical record. In general, clinical staging is used to select the primary therapy, and pathologic staging is
PROGNOSTIC FACTORS FOR HEMATOLOGIC CANCERS
913
used to select adjuvant therapy, to estimate prognosis, and for reporting end results. Therapeutic procedures may alter the course and life history of the disease. Cancers that recur after therapy may be restaged using criteria similar to those used in pretreatment staging, but the clinical significance of these criteria may not be the same. Restage classification (rTNM) of recurrent cancer is usually considered separately for therapeutic guidance, re-estimation of prognosis, and reporting end results at a particular time in the patient’s clinical course. For most cancer sites, staging is based only on the anatomic extent of disease, but in some cases histologic grade and patient age are factors that significantly influence prognosis and choice of treatment. In the future, introduction of new therapeutic interventions or well-evaluated prognostic factors may require further modifications of classification systems. For example, serum protein biomarkers have already been introduced as prognostic factors in the staging of testicular ~ a n c e r . ~ For hematologic malignancies, AJCC staging recommendations are available only for Hodgkin’s disease and non-Hodgkin’s lymphomas. Staging of Hodgkin’s disease is not based on the local extent of the disease but rather on its anatomic distribution and symptomatology. Because a patient usually presents with widely disseminated disease, it is usually not possible to determine the primary tumor site. The evaluation of many organs and groups of lymph nodes thus becomes extremely important. Pathologic staging depends on one or more lymph node biopsies, a bone marrow biopsy, and in some cases a laparotomy (which should include liver biopsy, splenectomy, and multiple nodal biopsies to assess the distribution of abdominal disease). Hodgkin’s disease is divided into four major histologic subtypes (nodular sclerosis, lymphocyte predominance, mixed cellularity, lymphocyte depletion) and a fifth ”unclassified As mentioned earlier, such subtyping may have prognostic significance. Long-term survival or cure can often be achieved in patients with NHL. Correct classification and staging can be critical for guiding initial therapy. Because the site of origin for NHL is often unclear, recognizing the histologic pattern of node involvement (follicular or diffuse) and the extent of disease at particular anatomic sites is extremely important. Clinical staging usually includes a physical examination, chest radiographs, blood evaluation, bilateral bone marrow biopsies, and CT scans. Diagnosis of malignant NHL also requires a biopsy and histologic analysis of lymph nodes or of an extranodal lymphoid tumor to determine the subtype. With new immunologic, cytogenetic, and molecular tools, new subtypes of lymphomas have recently been re~ognized.~ POTENTIAL PROGNOSTIC BIOMARKERS FOR HEMATOLOGIC MALIGNANCIES
Although staging of most cancers is based primarily on the anatomic distribution of disease, novel molecular markers hold promise for re-
fining cancer treatment and prevention. At present, newly emerging prognostic biomarkers for hematologic malignancies fall into four main categories: protein antigens, cytogenetic abnormalities, genetic polymorphisms, and gene expression alterations. Criteria for evaluating the potential usefulness of these prognostic biomarkers include the reproducibility, sensitivity, specificity, and cost of the assay procedure; how well the assay results predict the biologic endpoint being considered; and whether use of the marker will lead to more favorable clinical outcomes. Examples of each type of marker are listed in Table 1 and are described in detail here. Protein Antigens
Two protein antigens which may eventually be useful in managing AML are Tdt and CD7. It was recently reported that detection of Tdt and CD7 on the surface of mononuclear cells obtained from the peripheral blood or bone marrow of AML patients corresponded well with (9;22) translocation, defects involving chromosome 5 or 7, and less favorable outcomes as suggested by a lower rate of remission and a shorter duration of These markers may have prognostic value and their association with specific mechanisms of tumor progression may also facilitate development of more effective cancer therapies. Cytogenetic Abnormalities
New evidence suggests that complete karyotypic evaluation of blood or bone marrow samples from patients with CML may be useful in the prognosis of this disease. In one study the occurrence of two
Table 1. POTENTIAL PROGNOSTIC BIOMARKERS FOR HEMATOLOGIC MALIGNANCIES TYPe
Molecular Target
rLPe of Malignancy
Protein antigen
Tdt, cD7
Acute myelocytic leukemia
Cytogenetic abnormality
Trisomy 8 Isochrome 17q
Chronic myelocytic leukemia
Genetic polymorphism
fnf promoter
Non-Hodgkin’s lymphoma
Gene expression alterations
17856 genes
Non-Hodgkin’s lymphoma
Association
Lower rate of remission and shorter duration of survival Shorter duration of survival and resistance to therapy Shorter duration of sunrival and resistance to therapy Duration of survival
PROGNOSTIC FACTORS FOR HEMATOLOGIC CANCERS
915
Philadelphia chromosomes, a trisomy 8 or an isochromosome 17q, was associated with shorter median survival times and increased resistance to therapy.27Again, such information may be particularly valuable in individual case-management decisions. Genetic Polymorphisms
A specific, single base-pair alteration in the promoter region of the tumor necrosis factor (TNF) gene is one example of a polymorphic marker which may prove useful in improving prognosis of malignant NHL. Recently, an allele-specific polymerase chain reaction (PCR) assay was used to demonstrate an association between the presence of the TNF allele involved in increased TNF gene transcription and significantly shorter survival times and first-line treatment failure in NHL.64 Gene Expression Alterations
In the past few years, new technologies have emerged that can be used to measure the expression of thousands of genes sim~ltaneously.'~ Most of these methods involve the immobilization of various cDNA sequences onto a solid support in an ordered array. The total amount of mRNA in a cell is then amplified, fluorescently labeled, hybridized to the array, and quantified. It is generally assumed that the signal detected for a given gene sequence on the array is roughly linear to the proportion of mRNA corresponding to that gene in the starting mixture. Information about the differences in gene expression patterns between normal and transformed cells, or between cells from early stage and later stage tumors, can also be efficiently obtained using a slightly modified version of this technique. In the modified technique, amplified mRNA from the two cells of interest are labeled with different fluorescent probes, mixed, and then hybridized to the microarray. The relative expression of each gene in the two samples is ultimately provided by the ratio of the two fluorescent signa1s.l This method has recently been used in the search for improved methods for determining the prognosis and best treatment of Hodgkin's disease and NHL.2,l1 It is currently believed that neoplastic Reed-Sternberg cells may be the underlying cause of Hodgkin's disease. Such cells secrete peptides that elicit a systemic inflammatory response and are usually outnumbered by surrounding non-neoplastic cells by a factor of 1000 to 1. Thus far, the elusive nature of the Reed-Sternberg cells has made it extremely difficult to obtain information concerning the association of Hodgkin's disease with specific genes or cytogenetic abnormalities. By preparing cDNA libraries from individual Reed-Sternberg cells selected by micropipette from cell suspensions of primary tissue and analyzing them using an array technique, investigators have been able to generate gene expression profiles for 2666 known genes in Reed-Sternberg cells." It is hoped
that such information will help identdy the gene or genes responsible for Hodgkin’s disease and eventually lead to better prognosis and treatment. Similarly, diffuse large B-cell lymphoma (DLBCL), the most common subtype of NHL,, is known to be clinically heterogeneous. Although most DLBCL patients initially respond to chemotherapy less than half achieve long-term remission. Unfortunately attempts to define morphologic subgroups of DLBCL have largely failed. Using DNA microarray technology, investigators have recently demonstrated a diversity in gene expression profiles that seems to correspond with DLBCL subgroups. The gene expression profile of tumors that were associated with better post-therapy survival seemed similar to that of germinal center B cells, whereas the gene expression profile of the subgroup with poorer prognosis seemed similar to that of activated peripheral blood B cells? These results represent one of the first examples of the use of gene expression profiling to identify previously indistinguishable but clinically s i g h cant subtypes of cancer. STATISTICAL METHODS FOR EVALUATING PROGNOSTIC FACTORS
The evaluation of putative prognostic factors is a multistep process. As seen in Figure 1, evaluation proceeds from demonstrating an association of the factor with an outcome of interest, to building and testing
Process
Build predictive model
Test predictive model
Evaluate utility for treatment
4
StudiesNeeded
Large observational
4
b
statistical Tools
Neutral networks Decision trees
cox regression l~gisticregmsion
Independent validatitation
Clinical vials
Rate of change analysis Error r?te wmpson
cox regressim
Figure 1. Stages in validation of putative prognostic markers.
PROGNOSTIC FACTORS FOR HEMATOLOGIC CANCERS
917
predictive models incorporating the factor, and finally, to analyzing the use of the factor in treatment choice. Some statistical procedures that can be used at each step of the evaluation are reviewed here.
Demonstrating Associations with Survival or Other Outcomes
The initial phase of the assessment of a putative prognostic factor should determine whether the factor is associated with a particular outcome of interest (usually, in the cancer literature, overall survival or recurrence-free survival) and should provide an estimate of the magnitude of the association, if one exists. A factor may seem to be associated with an outcome when analyzed separately (i.e., with a univariate analysis) but this association may disappear if the factor is correlated with other known prognostic factors and these other factors are controlled for in an appropriate multivariate analysis. Kaplan-Meier survival curves and Cox proportional hazards regression are two common statistical methods used to examine associations with survival. The use of Kaplan-Meier survival curves amounts to a univariate analysis of a prognostic factor with two different levels and, as such, does not indicate whether the factor would be independently associated with the outcome after controlling for other known prognostic factors.I5Cox proportional hazards regression accommodates a multivariate analysis and can be performed on continuous as well as on binary or categorical prognostic variables; as suggested by the name, this method assumes that the hazards (i.e., risk of death or relapse) associated with different levels of a prognostic factor are proportional through time.” Methods exist to test this underlying assumption of proportionality, and various remedies have been proposed to deal with nonproportionality, including adding a time-varying interaction term and fitting a piecewise A Cox multivariate regression produces a proportional hazards P value indicating whether one can reject the null hypothesis of no independent association of the prognostic factor with the outcome of interest. If the null hypothesis of no association is rejected, one is then interested in the magnitude of the association; in Cox regression this magnitude is assessed by the degree to which the hazard is increased or decreased at different levels of the prognostic factor. If the null hypothesis is not rejected, it is important for researchers to specify the power of the study (i.e., the probability that the null hypothesis would have been rejected if an association of a given magnitude had existed). If this power is low, inability to reject the null hypothesis should not be taken as strong evidence against an association but rather indicates that the study is noninformative as to the presence of an association of the given magnitude. Note that both Cox regression and Kaplan-Meier analysis do not require specifying an arbitrary cutoff of survival time (e.g., 5year survival) but use survival time as a continuous measure. They
are also designed to accommodate censoring of the data (e.g., loss to follow-up). It may be of scientific interest to identify prognostic factors that have independent association with an outcome of interest even if it is not currently feasible to use the factor for clinical prediction. For example, the presence or absence of an association may give clues about the pathologic processes associated with the disease. Often, however, there is a specific interest in using a prognostic factor, or a set of factors, for clinical prediction. Models for clinical prediction are discussed in the next section. Building and Testing Predictive Models
The types of models commonly used for clinical prediction include the standard statistical models of logistic regression and Cox regression, decision-tree methods (e.g., classification and regression trees, or CART), and artificial intelligence systems such as neural networks." 8, 26, 47, 51, 57, 58 Each of these models can be constructed to give a probability of an outcome of interest for each subject (e.g., probability of 5-year survival) or to classify each subject according to the most likely outcome status. Another class of statistical models currently being used for prediction in cancers that are potentially curable are cure-rate models, which attempt to predict the cure rate as a function of prognostic fa~tors.2~ Several measures are commonly used to assess the predictive ability of these models. When the output of the model is subject classification, commonly used measures of predictive ability include predictive accuracy (i.e., the proportion of subjects correctly classified) and the percentage increase in accuracy over that of the null model (i.e., the model in which all subjects are classified as having the same, most likely, outcome)?O When the model output is a probability, receiver-operator characteristic (ROC) analysis is often utilized to assess predictive power.31 The ROC curve plots the true-positive versus false-positive rates at different cutpoints of a continuous variable; in the current context, the continuous variable is the probability level assigned by the model, the true-positive rate can be taken as the proportion of those subjects with the outcome (e.g., surviving 5 years) who have assigned probabilities above the cutoff, and the false-positive rate can be taken as the proportion of subjects without the outcome with assigned probability above the cutoff. The area under the ROC curve is the probability that the model will assign a higher probability of survival to a randomly chosen surviving subject than to a randomly chosen nonsurviving subject; its maximum value of 1 indicates perfect prediction, whereas its minimum value of 0.5 indicates no better than chance prediction. Another measure of predictability, which derives from information theory, is known as mutual information.35 Mutual information compares the uncertainty about a subject's outcome before and after the subject's prognostic factors are identified.
PROGNOSTIC FACTORS FOR HEMATOLOGIC CANCERS
919
In general, a model can be made sufficiently complex to fit any particular data set well; however, the resulting model may perform poorly when applied to another, independent data set. This phenomenon is termed overfitting, and, roughly speaking, indicates that the model has been fit to the noise in the data rather than the underlying structure.62 To guard against overfitting, it is critical that any predictive model be validated. The most rigorous validation of a model is external validation, that is, the application of the model, in its final ("frozen") form, to an independent validation data When true external validation for a predictive model is reported in the literature, the text should contain a clear statement to the effect that the validation data set played no part in model development. If external validation is not feasible (e.g., because no appropriate alternate data set is available), internal validation may be used. Methods of internal validation include data splitting, crossvalidation and boot-strapping.16,17,52 In theory, properly applied internal validation methods can give relatively accurate estimates of true predictive ability, although they are often computationally extensive. In practice, however, internal validation is often subject to subtle, or notso-subtle, manipulation on the part of the modeler. A number of studies have been performed to try to assess which of these models gives the best predictive performance^.^, 46, 59 In general, there is no clear consensus on an optimal method. Some methods do perform better under certain circumstances; neural networks, for instance, are particularly suited to large data sets with large numbers of potential prognostic factors.15For a specific modeling problem, statistical methods exist to compare the predictive performance of different models or to compare the performance of a model with that of a standard clinical staging system.32 As discussed in the next section, predictive models are often used on a clinical level to inform treatment choice. Predictive models may also have clinical value, in guiding patients and their families in planning for the future. Predictive models can also be used outside a clinical setting, for example, to help assess quality of care across institutions, where survival from a given disease is used as a surrogate for quality of care.3o Because the distribution of important underlying prognostic factors may vary across institutions, direct comparisons of survival rates may be invalid. A model, however, allows the comparison of the observed survival rate for any institution with a projected survival rate based on the prognostic factor profiles of the institution's patients. Utility of Predictive Factors in Treatment Choice The final, and perhaps most clinically useful, assessment of the utility of prognostic factors lies in evaluating their usefulness in choosing optimal treatments. It is rare for there to be a definitive study, in the form of a clinical trial, for example, which examines whether assigning treatments based on a new set of prognostic factors leads to better
outcomes than assigning treatments in the standard manner. In the context of a standard clinical trial comparing alternative treatments, however, evidence of the usefulness of a prognostic factor in choosing treatments may exist in the form of an interaction of treatment with the factor. A factor (or covariate) by treatment interaction s i d e s that the treatment effect varies according to the level of the factor. In the most salient form of an interaction, the treatment effect is seen at only some levels of the factor, whereas at other levels there is no treatment effect, or the treatment effect is even reversed. For example, in a study of gastric cancer, Saji et a1 showed that patients with high levels of immunosuppressive acidic protein (IAP) had improved survival if splenectomy was performed in conjunction with immunotherapy but patients with low levels of IAP had improved survival with immunotherapy if the spleen was preserved.” Because clinical trials are designed to detect primary treatment effects, however, their power to detect sigruficant interactions is usually quite low; thus, interactions of the type discussed here are rare in the clinical trial literature. A more common method of treatment choice involves the use of predictive models. Although a predictive model may contain information about multiple treatments, direct inference about the relative merits of these treatments for a person with a given set of prognostic factors is problematic at best, because various biases may exist in the way in which the treatments were assigned. A model based on a single (standard) treatment can, however, be used to identify subsets of patients who have poor (or good) prognosis when given that treatment, and this information can in turn be used in treatment choice. Patients with poor prognosis when given a standard treatment, for instance, may be suggested as candidates for new experimental treatments. Because these treatments may be costly potentially dangerous, or associated with diminished quality of life, it is not advisable to give these treatments to all patients. Wirth et al, for example, evaluated prognostic factors for clinical stage I and I1 Hodgkin’s disease in a group of patients treated with mantle radiotherapy (MRT) alone as the initial therapy.ffiThey identified a subset of patients with relatively good prognosis for whom MRT alone would be recommended and suggested that the remainder of the patients would require more intensive treatment. As suggested by Hayes et al, knowledge of prognostic factors should ultimately lead to clinical decisions that result in more favorable outcomes.34Such favorable outcomes may take the form of improved survival, increased quality of life, or decreased cost. As suggested by this discussion, there are several pathways, some more direct than others, by which the identification and use of prognostic factors can help achieve more favorable outcomes. AREASFORFUTURERESEARCH
The identification of novel molecular markers is key to improving diagnostic and prognostic capabilities with respect to hematologic can-
PROGNOSTIC FACTORS FOR HEMATOLOGIC CANCERS
921
cers. To use molecular markers most effectively, it is necessary to be able to separate the signal from the noise, that is, to distinguish the truly important markers from spurious ones. On the biologic level, a more thorough understanding of the underlying molecular and pathologic processes involved in these cancers and of the interactions between markers of interest and other environmental or genetic factors is critical. Additional research is also needed on statistical techniques; for example, the issue of how to analyze and interpret data on multiple markers generated from certain molecular technologies (e.g., gene-expression arrays) is far from resolved. Finally, ingenuity, effort, and cooperation from the scientific community are needed to develop methods for incorporating new types of molecular information into the cancer-staging process. Integration of all relevant sources of information will be essential for refining cancer prognostics and treatment.
References 1. Alizadeh A, Eisen M, Botstein D, et al: Probing lymphocte biology by genomic-scale gene expression analysis. J Clin Immunol 18:373-379, 1998 2. Alizadeh AA, Eisen MB, Davis RE, et al: Distinct types of diffuse large B-cell lymphoma identified by gene expression profiling. Nature 403:503-511, 2000 3. American Joint Committee on Cancer: AJCC Cancer Staging Manual, ed 5. Philadelphia, Lippincott-Raven, 1992 4. Anderson K Advances in the biology of multiple myeloma: Therapeutic applications. Semin Oncol26(suppl 13):lO-22, 1999 5. Bitter MA, Le Beau MM, Rowley JD, et al: Associations between morphology, karyotype and clinical features in myeloid leukemias. Hum Pathol 18:211-225, 1987 6. Bottaci L, Drew P, Hartley J: Artificial neural networks applied to outcome prediction for colorectal cancer patients in separate institutions. Lancet 350:469472, 1997 7. Brenn T, Amesen E: Selecting risk factors: A comparison of discriminant analysis, logistic regression and Cox’s regression model using data from the Tromso heart study. Stat Med 441M23, 1985 8. Burke H, Goodman P, Rosen D, et al: Artificial neural networks improve the accuracy of cancer survival prediction. Cancer 79557-862, 1997 9. Chen CS, Sorenson PH, Domer PH, et al: Molecular rearrangements on chromosome 11923 predominate in infant acute lymphoblastic leukemia and are associated with specific biologic variables and poor outcome. Blood 81:2386-2393, 1993 10. Cole P, MacMahon B, Aisenberg A: Mortality from Hodgkin’s disease in the U.S.: Evidence for the multiple aetiology hypothesis. Lancet 583:1371-1376, 1968 11. Cossman J, Annunziata CM, Barash S, et al: Reed-Stemberg cell genome expression supports a B-cell lineage. Blood 94:411416, 1999 12. C ~ XDR - Regression models and life tables. Joumal of the Royal Statistical Society 34187-220, 1972 13. DAndrea AD, Grompe M: Molecular biology of Fanconi anemia: Implications for diagnosis and therapy. Blood 90:17251736, 1997 14. DeRisi JL, Iyer VR, Brown PO: Exploring the metabolic and genetic control of gene expression on a genomic scale. Science 278:680-686, 1997 15. Drew P, Ilstrup D, Kerin M, et al: Prognostic factors: Guidelines for investigation design and state of the art analytical methods. Surg Oncol 771-76, 1999 16. Efron B: Estimating the error rate of a prediction rule: Improvement on cross-validation. Joumal of the American Statistical Association 78:316-331, 1983
17. Efron B, Tibshirani R An Introduction to the Bootstrap. New York, Chapman and Hall, 1993 18. Finiewicz KJ, Olopade 01Pathogenesis of early leukemia and lymphoma. In Srivastava S, Henson DE, Gazdar A (eds): Molecular Pathology of Early Cancer. Washington, DC, 10s Press, 1999, pp 233-253 19. Fong C-T, Brodeur GM Down’s syndrome and leukemia: Epidemiology, genetics, cytogenetics, and mechanisms of leukemogenesis. Cancer Genet Cytogenet 2855-76, 1987 20. Forbes JF,Moms PJ: Leukocyte antigens in Hodgkin’s disease. Lancet 678849-851,1970 21. Franchini G: Molecular mechanisms of human T-cell leukemia/lymphotropic virus type I infection. Blood 863619-3639, 1995 22. Fraumeni JL,Li FP: Hodglun’s disease in childhood: An epidemiologic study. J Natl Cancer Inst 42:681-691, 1969 23. Frizzera G, Wu CD, Inghnami G: The usefulness of immunotypic and genotypic studies in the diagnosis and classification of hematopoietic and lymphoid neoplasms. Am J Clin Pathol lll:S13-S39, 1999 24. Gieser PW, Chang MN, Rao PV, et ak Modelling cure rates using the Gompertz model with covariate information. Stat Med 178314339,1998 25. Gilpin E, Olshen R, Chatterjee K, et ak Predicting 1-year outcome following acute myocardial infarction: Physicians versus computers. Comput Biomed Res 23:46-63, 1990 26. Gonzalez-Vela M, Garijo M, Femandez F, et al: Predictors of d a r y lymph node metastases in patients with invasive breast carcinoma by a combination of classical and biological prognostic factors. Path01 Res Pract 195:611418, 1999 27. Griesshammer M, Heinze B, Bangerter M, et al: Kqotype abnormalities and their clinical significance in blast crisis of chronic myeloid leukemia. J Mol Med 7 5 8 3 6 838, 1997 28. Grignani F, Ferrucci PF, Testa U,et al: The acute promyelocytic leukemia specificPML/ RARa fusion protein inhibits differentiation and promotes survival of myeloid precursor cells. Cell 74:423431, 1993 29. Grufferman S Hodgkin’s disease. In Schottenfeld D, Fraumeni JF (eds):Cancer Epidemiology and Prevention. Philadelphia, WEl Saunders, 1982, pp 739-753 30. Hadom D, Draper D, Rogers W, et ak Cross-validation performance of mortality prediction models. Stat Med 1L475-489, 1992 31. Hanley J, McNeil B: The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology 143:29-36,1982 32. Hanley J, McNeil B: A method for comparing areas under receiver operating curves derived from the same cases. Radiology 1488394343, 1983 33. Harrell F, Lee K, Mark D Multivariable prognostic models: Issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Stat Med 15:361-387, 1996 34. Hayes D, Bast R, Desch C, et ak Tumor marker utility grading system: A framework to evaluate clinical utility of tumor markers. J Natl Cancer Inst 88:1456-1466, 1996 35. Heckerling P:Information content of diagnostic tests in the medical literature. Methods M Med 29:6146, 1990 36. Henderson E S Definition and classification. In Henderson ES, Lister TA (eds): Leukemia. Philadelphia, WB Saunders, 1990, pp 13-15 et al: The t(12;21) converts AML-1B from an activator 37. Hiebert SW, Sun W-H, Davis JN, to a repressor of transadivation. Mol Cell Biol 161349-1355, 1996 38. Hiscott J, Petropoulos L, Lacoste J: Molecular interactionsbetween HTLV-1 Tax protein and the NF-kappa B/Kappa B transcription complex. Virology 2143-11,1995 39. Horowitz M The genetics of familial leukemia. Leukemia 113347-1359, 1997 40. Johnson E, Cotter FE: Monosomy 7 and 7q-associated with myeloid malignancy. Blood Rev 11:46-55, 1997 41. Larson RA, LeBeau MM, Ratain h4J, et ak Balanced translocations involving chromosome bands 11q23 and 21422 in therapy-related leukemia. Blood 791892-1893, 1992 42. Le Beau MM, Espinosa R, Davis EM, et ak Cytogenetic and molecular delineation of
PROGNOSTIC FACTORS FOR HEMATOLOGIC CANCERS
923
a region of chromosome 7 commonly deleted in myeloid diseases. Blood 88:19301935, 1996 43. Le Beau MM, Espinosa R, Neuman WL, et al: Cytogenetic and molecular delineation of the smallest commonly deleted segment of chromosome 5 in malignant myeloid diseases. Proc Natl Acad Sci U S A 90:5484-5488, 1993 44. Leder A, Pattengale PK, Kuo A, et al: Consequences of wide-spread deregulation of the c-myc gene in transgenic mice: Multiple neoplasms and normal development. Cell 45:485495, 1986 45. Linet M S Leukemias. In Haras A (ed): Cancer Rates and Risks. Bethesda, MD, National Institutes of Health, 1996, NIH Publication No. 96-691, pp 148-154 46. Long W, Griffith J, Selker H, et al: A comparison of logistic regression to decision-tree induction in a medical domain. Computers and Biomedical Research 26374-97, 1993 47. Motzer R, Mazumdar M, Bacik J, et al: Survival and prognostic stratification of 670 patients with advanced renal cell carcinoma. J Clin Oncol 172530-2535, 1999 48. National Cancer Institute: SEER Cancer Statistic Review 1973-1996. Available at: http: / / www.seer.ims.nci.nih.gov. Accessed January 5, 2000 49. Nishida K, Tamura A, Nakazawa N, et al: The Ig heavy chain gene is frequently involved in chromosomal translocations in multiple myeloma and plasma cell leukemia as detected by in situ hybridization. Blood 90:526-534, 1997 50. Nunez G, Seto M, Seremetis S, et al: Growth and tumor promoting effects of deregulated bcl-2 in human B-lymphoblastoid cells. Proc Natl Acad Sci U S A 86:45894593, 1989 51. Otto F, Goldmann T, Biess B, et al: Prognostic classification of malignant melanomas by combining clinical, histological, and immunohistochemical parameters. Oncology 56~208-214,1999 52. Picard R Data splitting. American Statistician 44:140-147, 199053. Quantin C, Abrahamowicz M, Moreau T, et al: Variation over time of the effects of prognostic factors in a population based study of colon cancer: Comparison of statistical models. Am J Epidemiol 150:1188-1200, 1999 53. Quantin C, Abrahamowicz M, Moreau T, et al: Variation over time of the effects of prognostic factors in a population-based study of colon cancer: Comparison of statistical models. Am J Epidemiol 150:1188-1200, 1999 54. Riedel D, Pottern LM: The epidemiology of multiple myeloma. Hematol Oncol Clin North Am 6:22!%247, 1992 55. Ross JA, Potter JD, Robison LL: Infant leukemia, topoisomerase I1 inhibitors, and the MLL gene. J Natl Cancer Inst 861678-1680, 1994 56. Saji S, Sakamoto J, Teramukai S, et al: Impact of splenectomy and immunochemotherapy on survival following gastrectomy for carcinoma: Covariate interaction with immunosuppressive acidic protein, a serum marker for the host system. Surg Today 293504510, 1999 57. Sauerbrei W, Royston P, Boja H, et al: Modelling the effects of standard prognostic factors in node-positive breast cancer. Br J Cancer 79:1752-1760, 1999 58. Segal M: Regression trees for censored data. Biometrics 443547, 1988 59. Segal M, Bloch D: A comparison of estimated proportional hazards models and regression trees. Stat Med 8:539-550, 1989 60. Shurtleff SA, Meyers S, Hiebert SW, et al: Heterogeneity in CBFb/MYHIl fusion messages encoded by inv(l6) and t(16;16)(p13;q22) in acute myelogenous leukemia. Blood 85:3695-3703, 1995 61. Smith MR, Green WC: Type I human T-cell leukemia virus tax protein transforms rat fibroblasts through the cyclic adenosine monophosphate response element binding protein/ activating transcription factor pathway. J Clin Invest 88:1038-1042, 1991 62. Tu J: Advantages and disadvantages of using artificial neural networks versus logistic regression for predicting medical outcomes. J Clin Epidemiol 49:1225-1231, 1996 63. Venditti A, Del Poeta GBF, Tamburini A, et al: Prognostic relevance of the expression of Tdt and CD7 in 335 cases of acute myeloid leukemia. Leukemia 12:1056-1063, 1998 64.Warzocha K, Ribeiro P, Bienvenu J, et al: Genetic polymorphisms in the tumor necrosis
factor locus influence non-Hodgkin’s lymphoma outcome. Blood 913574-3581, 1998 65. Wirth A, Chao M, Corry J, et ak Mantle irradiation alone for clinical stage I-II Hodgkin’s disease: Long term follow-up and analysis of prognostic factors in 261 patients. J Clin Oncol 17230-240, 2000 66. Zahm S Non-Hodgkin’s lymphoma. In Harras A (ed): Cancer Rates and Risks. Bethesda, MD, National Institutes of Health, 1996,NIH Publication 96-691,pp 170-174 Address reprint requests to Sudhir Srivastava, PhD, MPH Cancer Biomarkers Research Group Division of Cancer Prevention National Cancer Institute 6130 Executive Boulevard, Suite 330F Bethesda, MD 20892
e-mail:
[email protected]