Overall survival in oncology phase III trials: an individual patient level meta-analysis (OPTIMA)

Abstract

Designing phase III oncology trials to demonstrate overall survival (OS) superiority may be challenging due to sample size, follow-up time, resources, and costs as compared to trials evaluating intermediate endpoints, which may or may not be reliable surrogates of OS. In this study, we investigate at scale the relative gains in OS achieved in modern phase III oncology trials using OS primary endpoints in OPTIMA, a prospectively registered, pre-specified meta-analysis leveraging a large assembly of reconstructed individual patient-level data (PROSPERO: CRD420251162778). After screening 1184 trials and excluding 937 without an OS primary endpoint, 159 trials enrolling 104 740 patients met the inclusion criteria. In the two-stage Cox model, randomization to the experimental arm was associated with improvement in OS (random effects HR, 0.87; 95% CI, 0.85–0.89; P < .0001). In the one-stage Cox model stratified by trial identity, OS was also longer with the experimental arm (HR, 0.88; 95% CI, 0.85–0.90; P < .0001). Taken together, the data from OPTIMA should encourage renewed enthusiasm among trialists, funders, regulatory agencies, and policymakers for prioritizing phase III oncology trials evaluating OS primary endpoints.

Introduction

The primary endpoints of phase III oncology trials have shifted over time from overall survival (OS) to alternative endpoints such as progression-free survival.¹ Such intermediate endpoints, which are most often claimed to be surrogates for OS, are attractive due to faster time to events and higher event rates, enabling trials to reduce enrollment targets and trial costs and facilitating faster trial readouts and regulatory approvals. In a previous work, we reviewed the outcomes of 791 modern phase III oncology trials and found that only 28% (n = 221) demonstrated OS superiority, despite 55% (n = 434) claiming superior treatment effects via alternative endpoints.² Importantly, OS was a primary or co-primary endpoint in 35% of trials (n = 420), a foundational design feature that may partly explain why OS superiority was only infrequently observed.³ Here, using the largest assembly of individual patient-level data to our knowledge from phase III trials, we evaluate the relative gains in OS at the individual patient level across the body of modern phase III oncology trials with OS primary endpoints. The objective of the “Overall survival in oncology Phase III Trials: an Individual patient level Meta-Analysis” (OPTIMA) study was to assess the overall effect of experimental vs control arm randomization on OS in phase III oncology.

Methods

OPTIMA was a prospectively registered, pre-specified meta-analysis of phase III oncology trials (PROSPERO: CRD420251162778). Eligible trials screened from ClinicalTrials.gov were two-arm superiority-design phase III interventional oncology studies with a matured reported primary endpoint of OS and a published reconstructable Kaplan–Meier plot. Two-arm studies were chosen to simplify the interpretation and avoid multiple comparisons of the same control group. OS was chosen rather than disease-specific survival, as OS is much more commonly used as a primary endpoint in phase III trials, which is due, among other factors, to the subjectivity introduced by attribution assessment. We selected only studies with OS primary endpoints—that is, those powered and designed to evaluate OS—to reduce meta-analysis heterogeneity. Following screening, individual patient-level data were reconstructed using previously reported methods.^4,5 Briefly, survival probabilities over time were obtained by manually annotating Kaplan–Meier curves using WebPlotDigitizer. Following integration of survival probabilities over time together with the number at risk over time, individual patient-level data were then estimated using Guyot’s approach for inversed Kaplan–Meier equations.⁶ Reconstruction quality was verified by comparing the HR point estimate from the Cox regression of the reconstructed data to the published HR as previously described.⁴

Following screening, data collection, and verification of reconstruction quality, both a two-stage and a one-stage meta-analysis were performed. Given the clinical heterogeneity across included trials (varying tumor types, disease stages, and treatment regimens), we a priori specified a random effects model for the two-stage meta-analysis. In the one-stage model, the Cox regression was stratified and clustered by trial identity to account for varying baseline hazards, and no assumption was made for a common treatment effect across studies.⁴ Subgroup analyses in the one-stage approach were also stratified and clustered by trial identity. Statistical significance was defined by P < .05. Analyses were performed using R v4.5.1.

Results

After screening 1184 registered phase III trials and excluding 984 for design reasons (including n = 937 that did not have an OS primary endpoint) and 41 that could not be reconstructed, 159 trials met inclusion criteria (Figure S1). Most trials were sponsored by industry (n = 139, 87%) and studied systemic therapy (n = 154, 97%) among patients with metastatic solid tumors (n = 126, 79%) (Table 1). The OS primary endpoint was met in 42% of studies (n = 67).

Table 1.

Characteristics of trials included in the meta-analysis.

	Incidence (%)
Site
Gastrointestinal	53 (33)
Genitourinary	27 (17)
Hematologic	15 (9)
Thoracic	35 (22)
Othera	29 (18)
Stage
Non-metastatic	15 (9)
Metastatic	126 (79)
Hematologic	15 (9)
Mixed or NS	3 (2)
Sponsor
Industry	139 (87)
Cooperative group	24 (15)
Treatment type
Systemic therapy	156 (98)
Local therapy	3 (2)
Enrollment, median (IQR)	608 (432-859)
Publication year, median (IQR)	2017 (2013-2021)

Abbreviation: NS, not specified.

^aOther: breast, central nervous system, gynecologic, head and neck, skin, and endocrine.

A total of 104 740 patients were included in the meta-analysis. In the two-stage model, randomization to the experimental arm was associated with improvement in OS (random effects HR, 0.87; 95% CI, 0.85-0.89; P < .0001). As anticipated, there was between-study heterogeneity (I² = 64%; τ² = 0.02; Cochran’s Q = 437, P < .0001) consistent with the a priori choice of random effects modeling. In the one-stage meta-analysis, OS was also longer (HR, 0.88; 95% CI, 0.85-0.90; P < .0001). OS superiority was also seen among multiple subgroups, including according to disease site, stage, and sponsor (Figure 1).

Figure 1.

Results of one-stage meta-analysis Cox regression, stratified and clustered by trial identity, overall and across selected trial subgroups.

Discussion

While the prevalence of OS superiority is infrequent among all phase III oncology trials, among the subset with an OS primary endpoint, the present meta-analysis demonstrates a statistically significant OS benefit associated with randomization to the experimental arm. The data from OPTIMA should encourage renewed enthusiasm for developing and accruing phase III oncology trials prioritizing OS evaluations. Limitations of this study include lack of individual patient-level covariates precluding adjustment for prognostic factors. Additionally, while we accounted for between-study heterogeneity through stratification and clustering by trial, residual heterogeneity likely remains. We omitted formal risk-of-bias assessment given that all included studies were randomized phase III trials with OS as the endpoint, which inherently limits concerns about selective reporting and outcome ascertainment.

Author contributions

Conception and design: Alexander D. Sherry, Pavlos Msaouel, Ethan B. Ludmir. Data collection: Alexander D. Sherry, Noah Meimoun, Avital M. Miller, Jessica Liu, Ramez Kouzy, Timothy A. Lin, Ethan B. Ludmir. Data analysis: Alexander D. Sherry. Interpretation: all authors. Manuscript writing: Alexander D. Sherry wrote the first draft. All authors revised the paper for critical intellectual content. Final approval of manuscript: All authors. Accountable for all aspects of work: All authors.

Funding

This work was supported in part by the National Institutes of Health/National Cancer Institute through MD Anderson’s Cancer Center Support Grant P30CA016672. E.B.L. and P.M. are recipients of the Andrew Sabin Family Foundation Fellowship.

Conflicts of interest

Dr. Msaouel reports honoraria for scientific advisory board membership for Mirati Therapeutics, Bristol-Myers Squibb, and Exelixis; consulting fees from Axiom Healthcare; non-branded educational programs supported by DAVA Oncology, Exelixis, and Pfizer; leadership or fiduciary roles as a Medical Steering Committee Member for the Kidney Cancer Association and a Kidney Cancer Scientific Advisory Board Member for KCCure; and research funding from Regeneron Pharmaceuticals, Summit Therapeutics, Merck, Takeda, Bristol-Myers Squibb, Mirati Therapeutics, and Gateway for Cancer Research (all unrelated to this manuscript’s content). No other authors report any conflicts of interest.

Data availability

Data will be shared upon reasonable request to the corresponding author after completion of a signed data transfer agreement.