Trends in the Risks and Benefits to Patients With Cancer Participating in Phase 1 Clinical Trials FREE

Since the National Cancer Act was signed in 1971, research into discovering and developing new cancer therapies has risen to the level of national priority. Despite the availability of more than 100 cancer drugs, treatments for most patients with solid tumors remain suboptimal. This year, more than 550 000 Americans will die from cancer,^1- 2 and this number is expected to increase as the population ages. In line with the burden cancer places on society, efforts to develop new therapies have never been greater. The number of cancer drugs under development now exceeds the combined total of the next 2 most-represented therapeutic classes, ie, anti-infective and digestive disease drugs.^3- 4

Phase 1 clinical trials are critical to the process of developing new cancer drugs. These trials represent the first testing of an investigational agent in humans and act as a point of translation of years of laboratory research into the clinic. The major objectives during phase 1 are to characterize the agent’s toxicity profile and to determine a dose and schedule appropriate for phase 2 testing. Under the standard phase 1 design, investigators enroll successive cohorts of 3 to 6 patients at increasing doses of an experimental therapy with the primary goal of determining dose-limiting toxicity and then backing down to a dose appropriate for phase 2 testing. Whereas phase 1 trials in other areas of medicine enroll healthy participants, phase 1 trials in oncology typically enroll patients who have cancer and who have exhausted standard treatment options. The dilemma of initiating trials in cancer patients results from the potentially lethal toxicity of many cancer drugs. A second difference among cancer phase 1 trials is that investigators also seek to realize therapeutic benefit, usually as a secondary end point. Treating physicians almost always enroll patients with therapeutic intent, and patients often expect to benefit.^5- 6 There are more than 550 phase 1 trials open to cancer patients in the United States at any given time,⁷ and the numbers are steadily increasing as more products of the biotech industry reach the clinic.

The ethical basis of phase 1 cancer trials has been questioned, in part because these trials involve potentially vulnerable patients at the end of life.^8- 16 Patients who choose to participate may experience significant risks with limited chance to benefit.^17- 20 Another concern is that those who participate may have unrealistic expectations about their chance to benefit, despite having participated in the informed consent process.^{5- 6,21- 23} Additional concerns relate to the uncertainty regarding risks and benefits of cancer drugs in their first stage of human testing. Agrawal and Emanuel²⁴ recently noted that these limitations principally arise from outdated studies. The most recent comprehensive analysis of phase 1 trials derived its data from studies published from 1972-1987.^18,24 Since then, there has been a shift in focus toward more targeted drugs, an introduction of better supportive care, and enhanced scrutiny of clinical research.^25- 27 Over the last decade in particular, a combination of factors, including several high-profile deaths of research participants, well-publicized problems with monitoring of research at leading medical centers, and concern over financial conflicts of interest among investigators, has prompted enhanced levels of oversight by institutional review boards (IRBs) and more vigorous enforcement of federal regulations protecting human participants.^25- 26 The impact of these changes on the safety of clinical cancer research has yet to be quantified.

We undertook the present study to determine if the levels of risk and benefit in phase 1 cancer trials have changed over time. Our primary objectives for this analysis were to determine trends in the rates of objective response, treatment-related death, and serious toxicity, and to identify factors such as drug class, sponsorship, and trial location that influence these outcomes.

The data for this analysis came from 2 sources: abstracts reporting on phase 1 cancer treatment trials drawn from a systematic review of the proceedings of annual meetings of the American Society of Clinical Oncology (ASCO)²⁸ for the years 1991 through 2002, and journal articles that resulted from those trials. We chose to create our own database because we found no preexisting database suitable for analysis.²⁴ We initiated our search with meeting abstracts because of concern about selection bias in published trials.²⁹ We chose the ASCO as it is the primary specialty society for medical oncologists and hosts the world’s largest oncology meetings.²⁸

We searched indexes of the published proceedings as well as individual abstracts to identify relevant studies. When there were multiple reports of the same study presented in successive years, we used the year of initial submission to analyze time trends. By comparing our capture rates to the number of phase 1 clinical trials listed with the National Library of Medicine,⁷ we estimated that our strategy would capture approximately two thirds of phase 1 oncology trials initiated over the study period.

We searched for journal articles to extract detailed information on clinical outcomes. In October 2002, we conducted a structured search of MEDLINE based on each author name listed on a given abstract in combination with all known names of the experimental agent studied. In order to minimize study heterogeneity, we limited our review to published studies of single agents not yet approved by the US Food and Drug Administration (FDA) for any indication at the time of the ASCO submission. We excluded trials that used radiation as part of the treatment to isolate the outcomes of investigational drugs. We also excluded trials that enrolled patients with hematologic malignancies (eg, leukemia and lymphoma), because these diseases have heterogeneous response criteria. Our selection process focused our review on the earliest trials of investigational cancer agents—studies that would be expected to have the least-favorable risk-to-benefit ratio in medical oncology and that would be expected to undergo high levels of scrutiny by IRBs and the FDA.

Five abstractors with advanced pharmacology training were assigned abstracts and journal articles for review with overlapping assignments to evaluate interobserver agreement. Abstractors received training for at least 40 hours prior to initiating their assignments and used a Microsoft Access interface (Microsoft Corp, Redmond, Wash) to record the structured information presented in the Box. Information from the abstracts was used to select studies for our detailed analysis. We relied exclusively on data from rel evant journal publications for clinical outcomes, since abstracts of phase 1 trials often present only preliminary results. Two investigators (T.G.R., B.H.G.) reviewed all data and resolved conflicts by discussion with a third investigator (J.W.C.).

Administrative Data

Trial Descriptors

Experimental Agent Descriptors

Clinical Outcomes

Abbreviations: ASCO, American Society of Clinical Oncology; FDA, US Food and Drug Administration.

*These data were obtained only from journal articles. Data for all other variables were obtained from both abstracts and journal articles.

†“Phase 1/phase 2” refers to those trials that included a phase 2 component (ie, efficacy testing) as part of the study design.

‡Studies using intrapatient dose escalation allowed individual patients to receive successively higher doses of an investigational agent. Studies using interpatient dose escalation administered successively higher doses to different groups of patients.

§Names not found in the publications were obtained from a commercially available drug database.

∥Cytotoxic agents were defined as those that cause cell death at standard doses, typically by interfering with DNA or cell division (eg, platinum analogs). Targeted/biologic agents were defined as those that have specific molecular targets and that alone are not toxic to cells at standard doses, including agents produced using recombinant technologies (eg, monoclonal antibodies or vaccines).

An objective response was defined in a patient when his or her target tumor(s) regressed on radiological evaluation by more than 50% in the sum of the products of the maximal perpendicular diameters without progression in any lesion and without any new lesion. This definition was based on World Health Organization criteria,^30- 31 which had near universal application in the trials and time period we reviewed. Responses were typically assessed every 2 cycles of treatment by computed tomography scans and were required to be maintained for at least 1 month. For patients with androgen-independent prostate cancer, we also allowed a sustained 50% reduction in the level of prostate-specific antigen to indicate a response. The inclusion of this definition is customary in prostate cancer trials, has been recommended by a consensus conference,³² and had little impact on the overall findings of our study. Treatment-related deaths were defined as those due to fatal toxicity that occurred directly as a result of trial participation. Non–treatment-related deaths were defined as those due to disease progression or any other competing risk (eg, an accident). We defined serious nonhematologic toxicity as any grade 3 or 4 neurologic, hepatic, renal, or pulmonary toxicity as reported in the publications; if the publication did not clearly state the toxicity grade, we used National Cancer Institute Common Toxicity Criteria (version 2.0).³³ These organ systems were chosen because of the relative difficulty of managing their toxicity. For each outcome, we individually reviewed investigator descriptions when available. Response, death, and toxicity rates were calculated as the number of events divided by the total number of enrolled patients.

In addition to our primary goal of examining time trends, we analyzed the association of response and treatment-related death with multiple factors. These variables included the source of funding and location of the trial, the dose escalation scheme and number of tumor types treated, and the route of administration and therapeutic class of the investigational agent (Box). We selected these explanatory variables as the factors most likely to influence clinical outcomes, based on structured interviews that were held with members of the Experimental Therapeutics Program of the Dana-Farber Cancer Institute/Harvard Cancer Center. For therapeutic class, we designated agents as either cytotoxic or targeted/biologic. Cytotoxic agents were defined as those that cause cell death at standard doses, typically by interfering with DNA or cell division (eg, platinum analogs). Targeted/biologic agents were defined as those that have specific molecular targets and that alone are not toxic to cells at standard doses, including those that are produced using recombinant technologies (eg, monoclonal antibodies).

We used χ² analysis to test for differences in categorical patient characteristics over time. To analyze trends in rates of response, treatment-related death, and toxicity, we first examined differences using the trial as the unit of analysis weighted by accrual size. To compare response and serious toxicity rates, we performed analysis of variance; for death rates, we used the Kruskal-Wallis test since death rates were highly skewed. To examine trends over time at the patient level, we performed univariate and multivariate logistic regression with an individual patient response or treatment-related death representing an event. For the multivariate models, we used stepwise logistic regression with P<.05 required for inclusion in the model. Since our primary objective was to determine changes in outcomes over time, we first created multivariate models excluding time, then tested for the independent association of time with response or treatment-related death. To analyze trends in response rates of individual tumor types over time, we used the Cochran-Armitage test for linear trend. We used the t test to compare response rates of agents that achieved FDA approval with rates for those that did not achieve approval. We stratified the 12-year study period into 4-year tertiles for reporting purposes; the findings were unchanged using time as a continuous variable except where noted. Planned analyses were prespecified in a protocol prior to data review. We made no adjustment for multiple testing in the univariate analyses, so interpretations of the univariate results must be judged carefully and significant results must be viewed as hypotheses requiring further validation. Statistical analyses were performed using SAS version 8.1 (SAS Institute Inc, Cary, NC) with a P value <.05 used to specify significance.

Our initial search identified 2460 phase 1 treatment trials (ie, the complete group). Figure 1 summarizes the trial flow resulting from our strategy to identify the relevant subset for detailed analysis. After excluding those trials with abstracts indicating the testing of multiple agents, the use of radiation, or the presence of a phase 2 component, 711 remained. We then excluded an additional 165 trials because the agents under testing were already approved by the FDA for at least 1 indication at the time of their ASCO presentation. Of the remaining potentially appropriate 546 studies, we successfully identified 242 (44%) as the subject of articles in the published literature, a rate of full publication similar to what has been seen in other fields.^34- 35 Twenty-nine trials were subsequently excluded because they violated 1 of our inclusion criteria upon further review (Figure 1). The remaining 213 published trials (ie, the restricted group), involving 6474 patients, provided the basis for our detailed analysis.

Table 1 summarizes characteristics of the 213 published trials in the restricted group investigating the effects of 149 unique experimental agents on 6474 enrolled patients. One hundred twenty-four (58%) of the studies were published in 2 oncology journals, which made the reporting of trial data relatively uniform. The mean number of patients enrolled per trial was 30 (SD, 15), with most trials escalating drug doses in 4 to 8 cohorts of patients prior to establishing a maximum tolerated dose, prior to stopping the trial, or both. Approximately half (100 [47%]) of the trials involved the testing of targeted/biologic agents, and most (168 [79%]) used the parenteral route for administration (ie, intravenous injection).

The characteristics of the patients were broadly comparable over the 3 study periods (period 1 [1991-1994], period 2 [1995-1998], and period 3 [1999-2002). The patients had a median age of 57 years in each of the periods; approximately 75% of the patients were between the ages of 50 and 70 years, with similar distributions of trial median ages across time periods. Table 2 describes additional baseline characteristics of the 6474 enrolled patients. The overall proportion of men was 56% (3346/5935), which decreased slightly over the 3 time periods. The majority (89% [4111/4612]) of patients with available data for performance status had a good baseline performance status score as determined by Eastern Cooperative Oncology Group (ECOG) or Karnofsky criteria (ie, an ECOG performance score of 0 or 1 or Karnofsky score ≥70). The median ECOG performance status score in each period was 1 (scale from 0-4, where 0 represents no symptoms and 4 represents complete disability).An ECOG performance score of 1 characterizes patients as “restricted in physically strenuous activity but ambulatory and able to carry out work of a light or sedentary nature.”³⁶ The composition of cancers among the trial participants remained relatively stable over time. More than 75 distinct tumor types were included in the analysis, but patients with cancers of the colon or rectum, lung, kidney, breast, and prostate were most common and together accounted for 56% of all patients. Among these 5 most frequently encountered tumor types, the proportions of patients with lung and breast cancer were similar over time, the proportions of patients with colorectal and kidney cancer increased, and the proportion with prostate cancer decreased.

There were 243 confirmed objective responses among 6474 patients enrolled in the 213 studies in the restricted group, producing an overall objective response rate of 3.8%. We found no relationship between response rates and time to article publication. At the trial level, average response rates decreased significantly over the 12-year study period, from 6.2% (SE, 1.7) in period 1, to 2.6% (SE, 0.5) in period 2, and to 2.5% (SE, 0.5) in period 3 (P<.01 between all periods) (Figure 2).

Table 3 presents univariate and multivariate predictors of response at the patient level according to trial characteristics. Based on the univariate analysis, the period in which the study was submitted to the ASCO, the location of the trial, number of tumor types treated, dose escalation scheme, source of funding, and route of administration were all significantly associated with response to therapy. Higher response rates were noted in trials conducted outside of the United States compared with those conducted within the United States (4.8% vs 3.2%; odds ratio [OR], 1.5; 95% confidence interval [CI], 1.2-2.0; P = .002). Trials without identified industry sponsorship had higher response rates compared with those with an identified industry sponsor (4.4% vs 3.2%; OR, 1.4; 95% CI, 1.1-1.8; P = .008). Trials allowing 3 or more types of tumors had lower response rates compared with those trials focusing on 1 or 2 types of tumors (2.9% vs 7.4%; OR, 0.38; 95% CI, 0.29-0.49; P<.001). Finally, patients enrolled in trials that allowed dose escalation for individual patients had higher response rates compared with those that did not allow such intrapatient dose escalation (5.3% vs 3.2%; OR 1.7; 95% CI, 1.3-2.2; P<.001).

Predictors of response in the multivariate model at the patient level included the period in which the study was submitted to the ASCO, the location of the trial, number of tumor types treated, and dose escalation scheme. Even after adjusting for all other significant predictors, the period in which the study was submitted to the ASCO remained an independent determinant of response to therapy. The adjusted odds of a patient responding to therapy in a study submitted during the most recent period (1999-2002) were less than half those for a patient participating in a study submitted in the earliest period (1991-1994) (OR, 0.46; 95% CI, 0.32-0.66; P<.001).

Table 4 lists response rates by period for each of the 10 most frequently encountered tumor types. We performed sensitivity analyses on subgroups of these 10 tumor types to evaluate whether significant declines in response rates for any individual or small group of tumors could have accounted for the overall effect, yet the findings proved to be robust. For example, among the 10 most frequently encountered tumor types, response rates declined significantly only in patients with colorectal, breast, and prostate cancer. Even after excluding the 2393 patients with these types of cancer, the response rates among the remaining 4081 patients continued to show a significant decrease over time, from 5.5% in period 1 (1991-1994), to 3.2% in period 2 (1995-1998), and to 2.9% in period 3 (1999-2002) (P<.01 between all periods). Additional analyses indicated that overall trends were driven by declining response rates across a broad array of tumor types, including lung, ovary, sarcoma, and head and neck cancer, and particularly among the combined group of less-common cancers.

Response rates in phase 1 trials predicted whether the experimental agent obtained FDA approval. Response rates in the phase 1 trials of the 11 agents in our study that had achieved FDA approval as of July 2004 (bortezomib, cetuximab, docetaxel, exemestane, gemcitabine, gifitinib, irinotecan, temozolomide, pemetrexed, topotecan, and tretinoin) were significantly higher compared with the trials of the 138 agents that did not achieve approval (8.2% vs 3.1%; P = .02).

There were 137 deaths from any cause among 6474 patients enrolled in the 213 studies in the restricted group, producing an overall death rate of 2.1%. The total number of deaths (and death rates) decreased over the 3 periods, from 63 out of 2113 (3.0%) in period 1, to 56 out of 2544 (2.2%) in period 2, and to 18 out of 1817 (1.0%) in period 3. Of the 137 deaths from any cause, 35 (26%) were classified as treatment-related (ie, fatal toxicity), leading to an overall toxic death rate of 0.54%. The types of fatal toxicities were diverse: representative examples included a fatal pulmonary embolism from an agent suspected of contributing to thrombosis and a fatal bacterial infection that developed in the setting of drug-related myelosuppression. At the trial level, treatment-related death rates decreased significantly over the study period, from an average of 1.1% (SE, 0.5) during period 1, to 0.39% (SE, 0.15) during period 2, and to 0.06% (SE, 0.06) during period 3 (P = .02 between all periods) (Figure 2).

Table 5 presents univariate and multivariate predictors of treatment-related death at the patient level according to trial characteristics. Based on the univariate analysis, the period in which the study was submitted to the ASCO, the source of funding, and the drug class were all significantly associated with treatment-related death. Trials without an industry sponsor had higher treatment-related death rates compared with those with an identified industry sponsor (1.0% vs 0.14%; OR, 7.1; 95% CI, 2.7-18.3; P<.001). Targeted/biologic agents had lower treatment-related death rates compared with cytotoxic agents (0.19% vs 0.80%; OR, 0.23; 95% CI, 0.09-0.60; P = .003).

Predictors of treatment-related death in the multivariate model included the period in which the study was submitted to the ASCO, the source of funding, and the drug class. Even after adjusting for all other significant factors, the period in which the study was submitted to the ASCO remained an independent predictor of treatment-related death. The adjusted odds of a patient dying from an experimental treatment in a study submitted during the most recent period (1999-2002) were less than one tenth those of a patient participating in a study submitted during the earliest period (1991-1994) (OR, 0.09; 95% CI, 0.01-0.67; P = .009). Of note, the adjusted odds of a patient dying from a targeted/biologic agent were one fourth those of a patient dying from a cytotoxic agent (OR, 0.25; 95% CI, 0.10-0.65; P = .005).

Even though overall death rates decreased during our study, we were concerned that changes over time in investigator classification of death as either treatment-related or disease-related may have biased our findings. We therefore performed sensitivity analyses to determine whether the period of study continued to be an independent predictor of treatment-related death after purposefully misclassifying the proportion of deaths recorded as treatment-related vs disease-related. We found that the period of study remained an independent predictor of death in the multivariate model even after artificially increasing the proportion of treatment-related deaths in the most recent period from 6% to 55%.

Our third outcome of interest was the trend in the rate of serious treatment-related toxicity. At the trial level, we found a decrease over time in the frequency with which trials were stopped on the basis of toxicity. At the patient level, there were 670 nonfatal serious (grade 3 or 4) toxic events meeting our definition among 6474 patients, leading to an overall serious toxicity rate of 10.3%. Approximately 85% of the events were reported as partly or completely reversible. Table 6 presents the occurrence of these events in the 4 organ systems on which we focused. Clinical signs and symptoms provided the basis for 247 (37%) of the total number, while laboratory or imaging tests provided the basis for 423 (63%). Rates of serious toxicities based on clinical signs and symptoms decreased from 4.2% in period 1, to 3.9% in period 2, and to 3.1% in period 3 (P<.01 between all periods). However, overall rates of serious toxicity (based on either signs and symptoms or laboratory/imaging tests) did not follow a linear pattern, as they peaked near the middle of the study (1994-1996) and then trended downward over the remaining 6 years.

We had also planned to collect data on cardiac toxicity, but there were very few documented instances of serious cardiac events within the trials we reviewed. The low rate of cardiac toxicity may relate in part to the difficulty of delineating cardiotoxicity of experimental agents from underlying heart disease, the infrequent performance of routine follow-up electrocardiography testing after a baseline examination, and the possible failure to detect cardiac events (eg, cardiomyopathy) if their clinical manifestations were delayed beyond the usual period of toxicity monitoring (ie, 1 month after trial completion).

Phase 1 studies that achieved publication by our search date had characteristics similar to those that did not, according to data obtained from respective abstract reports. Specifically, we found no significant differences between published and unpublished studies in terms of their study location (United Sates vs other), types of tumors treated, source of funding (industry vs other), class of agent under testing (targeted/biologic vs cytotoxic), or frequency in which pharmacokinetic studies were included. However, authors of published studies were almost twice as likely as authors of unpublished studies to indicate that the agent under investigation had completed phase 1 testing by the time of the ASCO presentation (27% vs 16%; OR, 1.9; 95% CI, 1.3-3.0; P = .002). This difference suggests that investigators who published their studies by the time of our literature search were more likely to have reported mature trials at the ASCO meetings. These investigators would have therefore required comparatively less additional time to complete their studies and publish the results.

We report here on the first analysis of phase 1 cancer trials in the era of targeted therapy. The odds of a patient with cancer dying from an experimental treatment while participating in a phase 1 trial have decreased over our 12-year study period. Coincident with this decreased rate of toxic death, the odds that a patient with cancer will respond to a single-agent experimental treatment in a phase 1 trial have also decreased. The overall ratio of benefit to risk may have improved, however, because death rates have decreased more rapidly than have response rates (Figure 2). The adjusted odds of a patient dying from an experimental treatment in a study submitted during the earliest period (1991-1994) were more than 10 times those of a patient participating in a study submitted during the most recent period (1999-2002).

There are several potential explanations for the sharp decline in treatment-related deaths. First, almost half (47%) of the trials in our detailed analysis involved the testing of targeted/biologic agents, which tend to have more favorable toxicity profiles compared with the cytotoxic drugs that have dominated cancer drug development in the past. However, drug class accounted for only part of the decline in treatment-related death rates in the multivariate model, suggesting that other factors have also played a role. The introduction of better supportive care over the study period, including effective antiemetic agents and hematologic growth factors, are likely to have played a positive role. Additionally, cancer trials in general have undergone more oversight by IRBs and have used stricter enrollment criteria, factors that are likely to have resulted in more conservative trial designs and more intense screening of potential patients. We found, for example, an increase over time in the proportion of trials that were stopped prior to achieving a maximum tolerated dose, and a modest but significant increase in the proportion of patients with good baseline performance status. In a growing number of instances, investigators relied on end points other than toxicity to define a dose appropriate for phase 2 testing. Finally, we cannot exclude the presence of publication bias. Although we found no relationship between response rates and times to article publication, the numbers of toxic deaths in the studies were too small to perform statistical tests for detecting publication bias based on this outcome. If trials with higher death rates achieved publication more slowly or with less frequency, then death rates in the later periods of the analysis would be artificially low. However, it is unlikely that publication bias alone could invalidate the direction of our findings given the strength of the trends, the large data set, the plausibility of alternative explanations, and the similar characteristics of published and unpublished studies.

We were surprised to see response rates decrease over our 12-year study period, yet there are several potential explanations. First, there may be stricter adherence to reporting criteria and expanded use of independent radiology review. We made efforts to review descriptions of each response to ensure that it conformed to standard World Health Organization criteria. However, because we did not have access to primary radiology films, we cannot exclude reporting bias. Second, because the number of standard treatments available to patients expanded over the study period, patients enrolling in trials during more recent years tended to have had more prior treatment, possibly contributing to drug resistance. Third, our exclusion of trials that enrolled patients with hematologic malignancies is likely to have biased downward the overall estimates of response rate, because recently approved highly effective drugs such as imatinib mesylate, used to treat chronic myelogenous leukemia, were excluded from the analysis. Likewise, our exclusion of trials testing combinations of agents as well as those testing approved drugs in new schedules is likely to have biased downward the estimates of response rate.

One interpretation of these data could be that patients with advanced cancer should not participate in phase 1 treatment trials because of the modest chance for benefit. This interpretation has several limitations. First, the overall response rate of 3.8% must be viewed in the context of the patients who participate in phase 1 treatment trials. Most of the participants have been heavily pretreated with chemotherapy and would have little chance to respond to standard therapies. For example, in a recent analysis, the response rate for patients with non–small cell lung cancer who received standard chemotherapy after treatment with 2 prior regimens was only 2.3%,³⁷ which is similar to the 4.2% overall response rate for lung cancer patients in our study (Table 4). Second, some phase 1 trials produce rates of objective response that are as high or higher than those used to support FDA approvals.²⁴ For example, FDA approvals for topotecan in ovarian cancer and irinotecan in colon cancer were based in part on phase 2 trials with response rates of 10% to 15%.³⁸ In comparison, the response rates among trials in the top decile of our restricted group exceeded 13%. Third, objective tumor response (ie, shrinkage) may not be an accurate estimate of benefit. A growing number of agents (eg, angiogenesis inhibitors) may inhibit tumor growth without producing tumor shrinkage; for these agents, alternative end points such as time to tumor progression may provide a better assessment of benefit.³⁹

Our principal finding—that phase 1 cancer trials appear to be safer, coincident with declining response rates—has important policy implications. A patient with cancer who is considering enrollment in the types of trials we analyzed may expect an improved level of safety compared with that expected by a patient enrolling just 10 years ago. Despite a possible improvement in the ratio of risk to benefit, rates of objective response have declined over time. It is possible that patients with advanced cancer would accept higher risk if it were accompanied by higher likelihood of response. In an influential article promoting a role for patients to participate in the planning of phase 1 trials, a trial participant, Professor George Zimmer, wrote that “the enemy is not pain or even death, which will come for us in any eventuality. The enemy is cancer, and we want it defeated and destroyed. This is how I wanted to die—not a suicide and not passively accepting, but eagerly in the struggle.”⁴⁰ Patients confronting death may have higher tolerances for risk than the agencies overseeing their clinical research, a point that was articulated persuasively in the 1980s by AIDS activists lobbying for earlier access to experimental drugs.^41- 42

A second implication of our principal finding relates to recent efforts to improve safety in clinical research. Because phase 1 cancer trials are considered among the most risky in all of medicine, the outcomes of these trials over time may serve as a relatively sensitive indicator of changes in the clinical research environment. Federal regulations governing research in human participants have not changed materially over the last decade,⁴³ but oversight of clinical research and the enforcement of the regulations have increased.^25- 27 For example, 7 of the 11 medical schools with the largest funding from the National Institutes of Health (NIH) established supplemental IRBs over the period from 1995 through 2002, doubling the total number of review boards at these schools. Many institutions have also developed educational programs for investigators, hired additional staff to oversee research, and placed senior officials in charge of programs for the protection of research participants.²⁵ The number of oversight steps required to initiate a clinical trial has increased at most institutions over the last decade.

At the federal level, the Office of Human Research Protection has been elevated from part of the NIH to exist as a separate Department of Health and Human Services agency and has received a larger budget. The NIH itself has also increased funding directed toward strengthening the oversight of clinical research.²⁵ There was even a Clinton administration proposal to levy civil penalties onto investigators who violated informed consent or other major research practices. Our study was not able to determine which aspect of trial review and management was responsible for the safety trends we observed, yet we do provide the first empirical data demonstrating a temporal association between improved safety in clinical trials and enhanced oversight of clinical research.

In summary, changes in the types of cancer drugs under study and in the clinical research environment have made participation in single-agent phase 1 cancer trials safer, particularly with respect to the probability of experiencing a treatment-related death. The enhanced safety may have also driven improvements in the ratio of risk to benefit. Future work should focus on applying new phase 1 designs that would treat fewer patients at subtherapeutic drug levels, especially for trials testing targeted agents with little expectation of toxicity. Investigators should also focus on how to include end points other than toxicity in phase 1 trials.⁴⁴ There is the growing realization that optimal dosing for targeted agents may exist at drug levels well below the maximum tolerated dose. Counter to the experience with cytotoxic chemotherapy, toxicity may not be a prerequisite for optimal antitumor activity for the new cohort of agents. Finally, the outcomes of experimental trials outside the cancer field should be studied in order to determine if the favorable safety trends that we observed apply to other areas of medicine.