Randomized controlled trials (RCTs) are the gold standard for demonstrating treatment efficacy, but they can be costly and time-consuming. The increasing availability of patient-level data from previous clinical trials and electronic health records (EHRs) offers the potential to improve the efficiency of new treatment evaluations. However, using external control (EC) data without appropriate statistical design can introduce bias. This paper addresses this challenge by proposing a hybrid trial (HT) design that combines the strengths of both RCTs and externally controlled trials (ECTs). The HT design aims to provide reliable treatment effect estimates even when external data have limitations, while maintaining efficiency comparable to ECTs under ideal conditions. The research is crucial as it directly tackles the limitations of existing trial designs, offering a potentially more efficient and reliable method for evaluating new treatments, particularly in areas with limited patient populations or high costs associated with traditional RCTs. The integration of external data, when properly handled, has the power to accelerate the drug development process while maintaining robust scientific rigor. The development and validation of this HT design is thus a significant step towards optimizing clinical trials in various settings.

The authors reviewed existing literature on RCTs, ECTs, and the use of external control data in clinical trials. They note the increasing interest in leveraging external data to reduce the cost and time of evaluating new treatments. However, they also highlight the potential for bias in ECTs if not properly accounted for. Previous work has focused on methodologies to integrate EC data into the analysis of single-arm trials (ECTs), using methods such as marginal structural models (MSMs), matching, and inverse-probability weighting (IPW). These methods rely on assumptions that may be difficult to validate in practice, such as the availability of all confounding variables and identical conditional outcome distributions between the trial and EC populations. The authors' previous work on ECTs and the use of external control data for predictions and futility interim analyses in clinical trials is also cited.

The study introduces a two-stage HT design. In the first stage, patients are randomized to experimental and internal control (IC) arms. An interim analysis (IA) determines if the study continues or stops for futility, and whether to update the randomization ratio for the second stage. This decision is based on an index of dissimilarity (W₁) between the IC and EC populations. At the end of the study, another dissimilarity index (W₂) is calculated, determining whether to use EC data for the final analysis. The HT design is compared to ECTs and RCTs using model-based simulations and in silico clinical trials generated using a resampling algorithm. The simulations include scenarios with varying levels of measured and unmeasured confounding and differences in outcome distributions between the IC and EC populations. For ECTs, marginal structural models (MSMs) were used to estimate treatment effects, while for RCTs and HTs (when EC data were not used), the difference in empirical response rates was used. A permutation test was also considered as an alternative for HTs using both trial and EC data. Two datasets, one from ES-SCLC studies and another from GBM studies, were used for the in silico trials. The leave-one-study-out resampling algorithm was used to generate in silico trials, treating the control arms of one study as the EC data and the remaining studies to generate in silico control and experimental arms. This process was repeated 2000 times to assess the operating characteristics of each trial design.

The simulations and in silico trials revealed several key findings. In the ideal scenario (no unmeasured confounding, similar outcome distributions), ECTs outperformed HTs and RCTs in terms of power. However, when key assumptions of ECTs were violated (unmeasured confounding, different outcome distributions), ECTs performed poorly, exhibiting inflated type I error rates and reduced power. In contrast, the HT design showed robustness across scenarios, maintaining control of type I error rates and providing good power. In ES-SCLC in silico trials, which reflected substantial heterogeneity, the ECT design substantially inflated the type I error rate (reaching up to 59%), whereas the HT design maintained type I error rates close to the nominal 5% level. In GBM in silico trials, where less heterogeneity was observed, both ECTs and HTs showed type I error rates near 5% and greater power compared to RCTs. Notably, both ECTs and HTs had higher probabilities of early termination for futility when treatment effect was null, leading to reductions in average sample size. The HT design also improved interim decisions by leveraging external data for futility analyses, leading to a higher likelihood of early study termination when the treatment was ineffective.

The results demonstrate the HT design's ability to balance the benefits of leveraging EC data with the need for robust inference in the presence of confounding or heterogeneity. The HT design safeguards against the risks associated with ECTs by incorporating a prospective assessment of the similarity between the EC and IC data. This adaptive approach allows the HT to leverage EC data when appropriate while maintaining the reliability of RCTs when external data are unreliable. The study highlights the importance of considering potential sources of bias when using EC data in clinical trials. The findings underscore the need for careful evaluation of EC data quality and the potential limitations of relying solely on ECTs in settings with unmeasured confounding or substantial heterogeneity across study populations. The proposed HT design provides a valuable alternative that offers a flexible and reliable approach for improving the efficiency and robustness of clinical trials. The efficiency gains observed in scenarios with high-quality EC data, such as the GBM datasets, highlight the design's adaptability and potential to optimize resource allocation in clinical research.

The HT design offers a robust and efficient approach to clinical trial design, combining the strengths of RCTs and ECTs. It provides reliable inference on treatment effects, even in settings with limitations in EC data, while achieving efficiency gains comparable to ECTs under ideal conditions. Future research could focus on exploring different dissimilarity indices, adaptive randomization schemes, and extensions of the HT design to other types of endpoints and study aims. The availability of high-quality, up-to-date patient-level data is crucial for the successful implementation of HT designs. Continued efforts in data sharing and standardization will play an important role in fostering wider adoption of this promising new approach.

The study's limitations include the relatively small number of GBM and ES-SCLC datasets used, limiting the generalizability of the findings. A larger number of datasets could provide a more robust evaluation of the HT and ECT designs across various outcome distributions and patient populations. The availability of prognostic variables in the ES-SCLC datasets was limited, potentially impacting the performance of ECTs and HTs. The heterogeneity within the ES-SCLC datasets, resulting from variations in treatment regimens and eligibility criteria, underscores the challenges in ensuring comparability across studies. Further, the inclusion of open-label and partially randomized studies in the ES-SCLC analysis might also contribute to heterogeneity.