Surrogate endpoints in clinical trials: Cancer

STATISTICS IN MEDICINE, VOL. 8,405-413 (1989)

SURROGATE ENDPOINTS IN CLINICAL TRIALS: CANCER

SUSAN S . ELLENBERG* Biometric Research Branch, National Cancer Institute, Bethesda, M D 20892, U.S.A.

AND

J. MICHAEL HAMILTON Clinical Investigations Branch, National Cancer Institute, Bethesda, M D 20892, U.S.A.

SUMMARY Investigators use a surrogate endpoint when the endpoint of interest is too difficult and/or expensive to measure routinely and when they can define some other, more readily measurable, endpoint, that is sufficiently well correlated with the first to justify its use as a substitute. A surrogate endpoint is usually proposed on the basis of a biologic rationale. In cancer studies with survival time as the primary endpoint, surrogate endpoints frequently employed are tumour response, time to progression, or time to reappearance of disease, since these events occur earlier and are unaffected by use of secondary therapies. In early drug development studies, tumour response is often the true primary endpoint. We discuss the investigation of the validity of carcinoembryonic antigen (a tumour marker present in the blood) as a surrogate for tumour response. In considering the validity of surrogate endpoints, one must distinguish between study endpoints that provide a basis for reliable comparisons of therapeutic effect, and clinical endpoints that are useful for patient management but have insufficient sensitivity and/or specificity to provide reproducible assessments of the effects of particular therapies.

KEY WORDS Clinical trials Cancer Endpoint Surrogate

INTRODUCTION

In clinical research, the endpoint of greatest relevance to inferences concerning therapeutic efficacy is frequently not practical or even feasible to measure. Sometimes the determination of the true endpoint is difficult, requiring an expensive, invasive or uncomfortable procedure. Sometimes we find it unobservable for an impractically long interval. Occasionally the true endpoint is not directly measurable at all, at least with current technology. In these cases we must rely on alternative, or surrogate, endpoints. The following examples illustrate some problems of endpoint assessment in clinical cancer research.

1. Hairy-cell leukaemia is a rare form of cancer, with approximately 500 incident cases per year in the United States. It is a relatively indolent disease; patients live for a median of 7-8 years after diagnosis, even untreated, but most will eventually die of the disease.' Currently, two very promising new treatments for this disease (deoxycoformycin and alpha interferon) are undergoing evaluation in a randomized trial. There is real hope that either or both of these treatments will result in prolongation of survival for a large proportion of patients. The primary interest in treatment comparison is whether one treatment produces longer survival times, on average, than the other. The use of survival time as the primary endpoint, however,

* Present address: AIDS Program, National Institute of Allergy and Infectious Diseases, 6003 Executive Boulevard, Rockville, MD 20892, U.S.A.

0277471 5/89/040405-09$05~00 0 1989 by John Wiley & Sons, Ltd.

Received December 1987 Revised July 1988

406 S. S. ELLENBERG AND J. M. HAMILTON

presents problems. First, the comparison requires a very large number of patients followed for an extended period to observe the necessary number of deaths. Second, patients for whom the assigned treatment appears ineffective will most likely receive the other treatment as secondary therapy, confounding the survival comparison. Because of the rarity of the disease, very large sample sizes are unobtainable; because of the enthusiasm about the preliminary results obtained with the new agents, a very long evaluation time is undesirable. Alternative endpoints that are being considered are response rates to the initial and secondary treatments and the time to disease progression.

2. Many cancer patients suffer severe pain. One objective of therapy, even when cure is improbable, is palliation. Thus in some cases we wish to assess the palliative value of a therapy. However, while we may observe that patients experience pain, and that pain clearly may increase or decrease in magnitude, we have no way to measure the degree of pain directly. In these trials, we must rely on indirect measures such as amount of analgesic received or requested, or self-administered or observer-administered questionnaires that provide a subjective scoring of pain.

3. Malnutrition is a major problem in cancer patients. Of greatest concern is the condition known as cachexia, or ‘wasting’, which involves loss of muscle as well as fat tissue. The technique believed most sensitive to changes in muscle mass is the measurement of key elements of body composition - nitrogen, potassium and water. We can obtain these measurements only with highly specialized equipment that requires the patients to be immobile for long periods of time. The equipment is very expensive and is available at only a few selected sites. Clearly, this test is impractical for widespread use. Simpler anthropometric measures, such as skinfold thickness, are considered much less precise but are nevertheless commonly used.

4. Cancers of the head and neck occur in patients with an average age of 60 years. The clear etiologic factors for head and neck cancer are cigarette smoking and heavy alcohol intake, both of which also contribute to the development of malignant and non-malignant lung disease, arteriosclerosis, heart disease, and liver disease. As in the first example, the primary goal of treatment is improvement of survival time. However, the rate of intercurrent deaths (deaths not attributable to the primary head and neck cancer) is high in this population and reduces the sensitivity of survival time to treatment effects. There are obvious difficulties associated with determination of the actual cause of death. Is a new deposit of tumour in the lungs a metastasis from the head and neck tumour or an ‘unrelated’ second cancer? Is death from a stroke an intercurrent event or attributable to radiation-induced narrowing of the carotid artery? Is an ‘obviously’ accidental death an accident or the result of a suicide due to treatment/cancer related depression? Endpoints that relate specifically to cancer, such as ‘local control’ (prevention of tumour re-growth following surgical removal) are commonly used in trials that involve regionally applied therapies such as surgery and radiation. Such endpoints are problematic, however, in that their use implies consideration of other cancer- related outcomes (such as death from metastatic disease) as intercurrent rather than informative events.

5. The determination of endpoints for adjuvant trials in colon cancer presents a related problem: many patients die from other causes because they live a long time following their diagnosis of cancer. Although the average age of 60 years for patients who enter colorectal adjuvant trials is the same as for head and neck cancer, long-term survival rates even without treatment are quite high in certain subgroups and approach 50 per cent for the total population of patients eligible for such trials. Recurrences of the primary tumour seldom occur more than four years after initial surgery. Therefore, long term evaluation of the

SURROGATE ENDPOINTS CANCER 407

therapy again becomes clouded by increasing numbers of non-cancer deaths in the aging patient population, often at the five to eight year point at which one might declare a ‘cure’.

In these and many other instances in cancer studies, investigators establish surrogate endpoints as the primary basis for inference because it is not feasible to observe the ‘true’endpoint in most or all of the patients under study. Surrogate endpoints are generally proposed on the basis of a biologic rationale. When this rationale is less than compelling one must establish statistically that the correlation with the ‘true’ endpoint is sufficiently great to justify the surrogate endpoint as a basis for inference. We provide an example of such an investigation later in this paper.

TRIALS WITH SURVIVAL TIME AS THE PRIMARY ENDPOINT

The consideration of surrogate endpoints in cancer studies must begin with specification of the true endpoint of interest. For preliminary trials of investigational treatments, the endpoint (as we will discuss later) may be a pharmacologic-, toxicity-, or tumour-based criterion. The ultimate goal of new treatment development, however, is to prolong life, or to reduce morbidity without adversely affecting survival probabilities. The Food and Drug Administration requirement for approval of a new anti-cancer treatment is that, compared to no treatment, it either lengthens survival or improves ‘quality of life’.2 In the examples given earlier we suggested that in some cases we may have reasons to consider an endpoint that serves as a surrogate for survival. In fact, one could make a credible argument that survival itself is a ‘surrogate’ for the more specific endpoint, time to cancer-related death (or ultimately cancer cure). Since we frequently have difficulty determining whether or to what extent a patient’s malignant disease (or treatment for that disease) contributed to an apparently ‘non-cancer’ death, these determinations are inevitably subjective to some degree and are therefore susceptible to bias. For this reason, we generally accept overall survival as the primary endpoint of interest, even in diseases such as head and neck cancer and colon cancer in which we know the competing risks of other potential causes of death are high. If, however, we could know for certain which deaths werk completely independent of a patient’s cancer or treatment for cancer, we would certainly use a disease-specific approach to estimate survival rather than to include all deaths.

Survival is the simplest and most definitive of all possible endpoints. There are, however, two justifications cited for utilizing other, ‘surrogate’, endpoints in cancer clinical trials. First, a patient may survive for a long time following diagnosis of certain cancers; we find it undesirable to require many years of observation to determine the effectiveness of therapy, especially when multiple treatments are available for evaluation and/or patient numbers are limited. Second, we may have effective secondary or ‘salvage’ therapies that prolong survival independent of the effect of the primary treatment under study. For these reasons, cancer investigators have sought alternative endpoints that they believe may reflect in some sense the ‘potential’ survival attributable to the administered therapy.

Perhaps the most commonly used surrogate for survival is ‘tumour response’. Tumour response means the disappearance of tumour entirely (complete response) as documented by X-ray, CT scan, exploratory surgery, etc. or a large reduction in tumour size (partial response), sufficient to rule out measurement error as the cause of the observed shrinkage. (The usual criterion for partial response is a 50 per cent reduction in the maximum area of a tumour cross-section. There is controversy, however, about whether this criterion is sufficiently stringent to prevent measurement error from playing a large part in endpoint determinati~n.~. ’) The biologic rationale for the use of tumour response as an endpoint is straightforward. Natural history studies of untreated cancer show that cancer kills by growing, spreading and inhibiting growth of normal


tissue to the point of irreversible compromise of organ function. Thus, a reduction in the amount of tumour in a patient’s body would seem likely to predict for prolonged survival as compared with a patient whose tumour continued to grow.

An obvious advantage to using tumour response as an endpoint is that we can characterize all patients as responders or non-responders within a short interval following study entry, thereby increasing the efficiency of the study. The correlation between improved response rates and improved survival is not perfect, however. Occasional randomized studies have demonstrated differences in response rates without any apparent differences in survival The failure to improve survival in the face of an increased response rate may be due in some cases to effective secondary therapies; in other cases, while the response data may very well reflect the cytotoxic potential of therapy, the reduction of a large tumour burden by 50 per cent may not be sufficient to improve survival in a population having advanced disease. A third possibility is that the number of responding patients is insufficient to permit observation of an effect on survival. Complete response (total disappearance of tumour) has shown a more consistent association with prolonged survival. In paediatric acute lymphocytic leukaemia, the use of complete response as primary endpoint in stepwise development of treatment strategies eventually led to regimens that prolonged survival, even though survival did not appear to be affected in the initial studies.

The strong rationale for tumour shrinkage as a meaningful endpoint has led physicians beyond the ‘surrogate endpoint’ to the ‘surrogate analysis’, a comparison of the survival of responders and non-responders in an uncontrolled series. The demonstration that responders live longer than non-responders is often taken as proof of the effectiveness of treatment, with the non-responders implicitly serving as surrogates for untreated controls. The difficulties with this type of analysis have been well documented.’-’ Briefly, these difficulties include lead-time bias (to, become a responder one must survive at the minimum to the first re-evaluation) and potentially confounding variables (both response and survival may be related to performance status or other important prognostic factors not balanced between responders and non-responders).

The type of surrogate endpoint selected depends on the stage of disease in the patient population studied. Tumour response is a feasible surrogate endpoint only when all patients have measurable tumour. Many cancer study populations consist of patients whose disease has been apparently eradicated, either by surgery, radiation, or chemotherapy. In such patients, ‘cure’ is a valid goal of treatment; further therapeutic intervention (adjuvant therapy) would be designed to eliminate any remaining microscopic disease, or micro-metastases, which may have gone undetected following the initial ‘curative’ therapy. In such patient populations, investigators commonly employ endpoints such as ‘disease-free survival’ or ‘time to treatment failure’. The event which serves as a surrogate for death is reappearance of disease. In most common cancers (for example, lung cancer, colon cancer) there is no effective salvage therapy for patients, and death from cancer is the inevitable outcome following disease recurrence. Further, the detection of recurrent disease is fairly straightforward. In such cases, disease-free survival would seem to be an acceptable surrogate for survival; because the endpoint occurs sooner, one may reach conclusions earlier with analyses based on the surrogate endpoint. In other diseases, such as Hodgkin’s disease and testicular carcinoma, there are effective secondary therapies which may shrink or destroy the recurrent tumour. In this case, the disease-free survival may not correlate closely with absolute survival but may still constitute a valid indication of a therapy’s potential to delay cancer growth.

Those who study paediatric cancers are especially interested in cure rate as an ultimate endpoint. On can never assess cure rate directly since it is impossible to examine a patient with sufficient thoroughness to rule out the presence of any malignant cells. Observation of large patient series, however, may guide us to a reasonable surrogate endpoint. In Wilms’ tumour, long- term follow-up of large numbers of children has shown that relapse-free survival for two years


Figure 1. Relapse-free surival curves for National Wilms' Tumor Study 2, group I to I11 patients Numbers in parenthesis are alive and well and still under observation at each anniversary.

[Reprinted with permission of the Journal of Clinical Oncology]

following initial surgery is a reliable indicator of cure, as shown in Figure 1." Thus, the relapse- free survival appears to be an appropriate endpoint in this disease if one's real interest is in the proportion of children who are cured by a particular therapy. One would have to consider the survival comparison as well, however, to ensure that a therapy which produced a superior relapse- free survival did not somehow detract from the capacity of the relapsed patients to benefit from secondary therapy, or produce some long-term adverse effect which reduces or eliminates the survival advantage. Freil' discusses the concept of utilizing the 'point-of-plateau' of disease-free or relapse-free survival curves as an indication of cure.

We note here the frequent use of the five-year survival rate as a surrogate for cancer cure. For many cancers, this may be reasonable. For some cancers (such as Wilms' tumour) an endpoint based on an earlier point in time may be equally acceptable. For other cancers, patients continue to experience recurrence or progression of their cancers for many more than five years following initial diagnosis and treatment. An excellent example is breast cancer. Studies shows that the probability of cancer recurrence following 'curative' mastectomy remains non-negligible for 10 or more years following primary surgery. In such diseases, five-year survival represents no more than an arbitrary point on the survival curve and does not constitute a reasonable endpoint as the basis for comparisons of therapeutic strategies.

TRIALS WITH OTHER PRIMARY ENDPOINTS

While the overall intent of developmental trials for new anti-cancer agents is to prolong survival by curing the cancer, cure or even anti-tumour activity may not be the primary endpoint of interest in all the phases of drug development (Table I). Phase I trials of new drugs are the first clinical tests of new agents after test tube and animal screening. The intent of such studies, usually conducted on patients who have no further options for curative treatment, is to determine the spectrum of


Table I. Types of cancer clinical trials

Trial phase Purpose of trial Usual endpoints

Phase I drug toxicity; pharmacokinetics; grades of side effects; drug

Phase I1 identify clinical drug activity tumour response Phase 111 compare new treatment regimen to patient survival

establishing maximal tolerable dose levels

standard regimen

human toxicity and drug kinetics and to establish the maximum dose level one can deliver safely. Phase I1 trials involve the testing of drugs as single agents in a population of patients with advanced disease to screen for anti-tumour activity. One rarely expects single-drug therapy to have an impact on survival. The development of effective combinations of drugs, however, clearly depends on the identification of individual drugs that have the capacity to kill tumour cells. Therefore, the goal of phase I1 studies is to identify such agents, with the hope of eventual development of curative therapy based on active drugs. Tumour response, then, is the most logical endpoint in this type of trial.

Tumour response itself is an endpoint for which surrogates have been considered. There is good documentation that our usual tools to measure tumour are impre~ise,~. making it difficult to judge confidently whether a tumour is increasing or decreasing in size. Perhaps more importantly, many tumours are not readily measurable by any method. Thus, investigators have sought other indicators of tumour status. One such potential indicator is carcinoembryonic antigen (CEA), one of a number of ‘tumour markers’ which have been identified. The study of CEA as a potential surrogate for tumour response serves as an interesting example of an attempt to establish the validity of a surrogate endpoint by investigating the statistical correlation between the surrogate and the true endpoint. While the attempt was ultimately unsuccessful - CEA levels are currently considered valuable for individual patient monitoring but not as reliable and reproducible endpoints for clinical trials - the process is instructive.

EVALUATION O F A SURROGATE ENDPOINT EXAMPLE

CEA is a glycoprotein manufactured by human neoplasms and released into the blood. One determines the CEA concentration by means of radio-immunoaisay of blood serum. Since its discovery in 1965, CEA has undergone extensive study in many different patient populations. If we could show that it correlates sufficiently well with disease status in patients whose disease status we can reliably assess by standard clinical methods, we might find CEA acceptable as a surrogate for response, progression, or recurrence of malignant disease.

Analyses of CEA measurements in cancer patients have focused primarily on two issues: first, whether increases and decreases in CEA levels over time in patients with residual tumour were reliable indicators of tumour growth and tumour shrinkage, respectively; and second, whether one could establish cutoff values for CEA levels that would, in patients whose tumours had been surgically removed, distinguish between patients who remained free of tumour and those experiencing re-growth of tumour. For the most part these analyses have been very straightforward. Initially, investigators studied patterns of CEA concentration in individual patients according to clinically determined tumour status (Figure 212 and others’3. 14). The presentation of serial CEA values, apparently so well correlated with clinical events, encouraged

SURROGATE ENDPOINTS CANCER 41 1

w

28 - 24 - 20 - 16 - 12 - 8 -

4-

SlGMOlO CARCINOMA DUKES B /.-. i

I I 1 1 I 1 1 I 1 1 1 1 I

4 6 8 10 12 14 16 18 2 0 22 2 4 2 6 2 8

MONTMS FOLLOWING FIRST TUMOR RESECTION

Figure 2. Repeated CEA determinations in a patient following complete resection of Dukes' B colon cancer [Reprinted with permission of Cancer]

researchers to pursue more sophisticated studies. The variability of serial assays in patients with and without clinical evidence of changing disease status has been analysed in an attempt to quantitate the degree of change required for the characterization of shrinking or enlarging tumours. A study by Lokich et ~ 1 . ' ~ used the coefficient of variation calculated from serial assays to determine the degree of rise or fall in CEA that was necessary to reliably suggest changing disease status. There have been studies of recurrence patterns in surgically resected patients according to level of CEA observed prior to and following surgery.I6 Winkel' used time-series methods to investigate CEA as a marker for recurrence in breast cancer, and Gail'* utilized the Cox proportional hazards model with serial CEA values as time-dependent covariates to perform a similar investigation in colorectal cancer. The results of these and many other studies led to the conclusion that CEA was not a sufficiently reliable indicator of tumour status to serve as a surrogate for tumour response. Even in colorectal cancer, where the correlation with clinical measures was strongest, there were clear limitations: localized disease often did not produce elevated CEA levels, and conversely, certain non-malignant conditions (such as hepatitis and chronic smoking) could cause CEA elevation. While very high levels of CEA were almost always associated with the presence of tumour, absence of high CEA levels or stability of serial CEA values did not reliably imply disease stability.

In 1980, an NIH consensus development conference on the role of CEA in cancer management concluded that the CEA assay might complement (not replace) the standard clinical measures of tumour resp~nse . '~ Nevertheless, the conference participants affirmed its clinical utility in monitoring response to treatment. While these conclusions might at first seem inconsistent, one must recognize that the parameters to be considered in clinical decision-making regarding individual patients cannot always be limited to those reliable enough for use as endpoints of clinical trials. To give an extreme example, we cannot wait until a patient dies to make the determination that the patient needs some alteration in therapeutic approach. The treating

412 S. S. ELLENBERG A N D J. M. HAMILTON

physician must consider laboratory measures such as CEA and other tumour markers, tests of organ function, etc., as well as clinical observations such as decreased energy and weight loss, to develop therapeutic strategies, even though these measures may lack sufficient sensitivity and specificity to warrant their use as formal endpoints in clinical trials.

CONCLUSIONS

Surrogate endpoints are frequently used in cancer studies, especially when long-term survival or ‘cure’ is the primary endpoint of interest. Endpoints which are commonly considered surrogates for cancer death are tumour enlargement and reappearance of tumour following its apparent eradication by primary therapy. These endpoints are biologically reasonable; nevertheleq, their use should not preclude monitoring of survival patterns, as some therapies may appear effective in preventing or delaying tumour growth but may also have long-term adverse effects which will diminish or even eliminate any survival advantage. Potential surrogate endpoints with less of a biologic rationale must be extensively studied in conjunction with the ‘true’ endpoints to determine their reliability. Some endpoints may not be sufficiently accurate for use in formal assessments of therapeutic strategies, but may still be valuable for individual patient management.

REFERENCES

1 . Golomb, H. M., Ratain, M. J. and Vardiman, J. W. ‘Sequential treatment of hairy cell leukemia: a new role for interferon’, in DeVita, V. T., Hellman, S. and Rosenberg, S . A. (eds), Important Aduances in Oncology, 1986, J. B. Lippincott Co., New York, 1986, pp. 311-321.

2. Johnson, J. R. and Temple, R. ‘Food and Drug Administration requirements for approval of new anticancer drugs’, Cancer Treatment Reports, 69, 1155-1 157 (1985).

3. Moertel, C. G. and Hanley, J. A. ‘The effect of measuring error on the results of therapeutic trials in advanced cancer’, Cancer 38, 388-394 (1976).

4. Warr, D., McKinney, S. and Tannock, I. ‘Influence of measurement error on assessment of response to anticancer chemotherapy: proposal for new criteria of tumor response’, Journal of Clinical Oncology, 2, 1040-1046 (1984).

5. Ruckdeschel, J. C., Finkelstein, D. M., Ettinger, D. S. ‘A randomized trial of the four most active regimens for metastatic non-small-cell lung cancer’, Journal of Clinical Oncology, 4, 14-22 (1986).

6. Jacobs, C., Meyers, F., Hendrickson, L. ‘A randomized phase 111 study of cisplatin with or without methotrexate for recurrent squamous cell carcinoma of the head and neck‘, Cancer, 52,1563-1569 (1983).

7. Anderson, J. R., Cain, K. C. and Gelber, R. D. ‘Analysis of survival by tumor response’, Journal ofclinical Oncology, 1, 710-719 (1983).

8. Weiss, G. B., Bunce, H., 111, and Hokanson, J. A. ‘Comparing survival of responders and nonresponders after treatment: a potential source of confusion in interpreting cancer clinical trials’, Controlled Clinical Trials, 4, 43-52 (1983).

9. Mantel, N. ‘Responder versus nonresponder comparisons: daunorubicin plus prednisone in treatment of acute nonlymphocytic leukemia’, Cancer Treatment Reports, 67, 315-316 (1983).

10. Breslow, N., Churchill, G., Beckwith, J. B. ‘Prognosis for Wilms’ tumour patients with nonmetastatic disease at diagnosis-results of the second National Wilms’ Tumour Study’, Journal of Clinical Oncology, 3, 521-531 (1985).

1 1 . Frei, E. F. ‘Curative cancer chemotherapy’, Cancer Research, 45, 6523-6537 (1985). 12. Mach, J. P., Vienny, H., Jaeger, P. ‘Long-term follow-up of colorectal carcinoma by repeated CEA

radioimmunoassay’, Cancer, 42, (Supplement), 1439-1447 (1978). 13. Mayer, R. J., Garnick, M. B., Steele, G. D. Jr and Zamcheck, N. ‘Carcinoembryonic antigen (CEA) as a

monitor of chemotherapy in disseminated colorectal cancer’, Cancer, 42 (Supplement), 1428-1433 (1978). 14. McCartney, W. H. and Hoffer, P. B. ‘Carcinoembryonic antigen assay: an adjunct to liver scanning in

hepatic metastases detection’, Cancer, 42 (Supplement), 1457-1462 (1978). 15. Lokich, J., Ellenberg, S. and Gerson, B. ‘Criteria for monitoring carcinombryonic antigen: variability of

sequential assays at elevated levels’, Journal of Clinical Oncology, 2, 181-186 (1984).


16. Steele, G., Jr, Ellenberg, S., Ramming, K. ‘CEA monitoring among patients in multi-institutional adjuvant G. I. therapy protocols’, Annals of Surgery, 196, 162-169 (1982).

17. Winkel, P., Bentzon, M. W., Statland, B. E. ‘Predicting recurrence in patients with breast cancer from cumulative laboratory results: a new technique for the application of time series analysis’, Clinical Chemistry, 28, 2057-2067 (1982).

18. Gail, M. H. ‘Evaluating serial cancer marker studies in patients at risk of recurrent disease’, Biomefrics,

19. Summary of an NIH Consensus Statement. ‘Carcinoembryonic antigen: its role as a marker in the 37, 67-78 (1981).

management of cancer’, British Medical Journal, 282, 373-375 (1981).

Documents

Surrogate endpoints in clinical trials: Cancer