BRIDGE Trials: Reusing Data to Validate AI Models Faster and Cheaper
Validating a new AI model for clinical use typically requires a prospective clinical trial—a costly and time-consuming endeavor. A novel trial design called "BRIDGE" offers a compelling alternative: reusing data from previously completed randomized controlled trials (RCTs) to validate AI risk prediction models. This approach could dramatically reduce the cost and time required to bring new AI diagnostic and prognostic tools to market.
The Challenge of AI Validation
The AI and machine learning community has developed a vast number of risk prediction models for clinical applications—from predicting cancer recurrence to forecasting sepsis onset in the ICU. However, the path from a model with promising performance on a retrospective dataset to a model that is trusted and used in clinical practice is long and difficult.
Regulators and clinicians rightly demand rigorous validation before an AI tool can influence patient care. This typically means a prospective clinical trial where the model's predictions are tested against real-world outcomes. Such trials can cost millions of dollars and take years to complete, creating a significant barrier to the adoption of beneficial AI technologies.
The Validation Bottleneck:
- Prospective validation trials for AI models can cost $2-10 million or more.
- Enrollment and follow-up can take 3-5 years.
- Many promising AI models never reach clinical practice because validation is not feasible.
Introducing the BRIDGE Design
The BRIDGE (Biomarker Reuse for Independent Development and Generalization Evaluation) trial design, introduced in a seminal paper published in November 2025, offers a solution. The core idea is simple but powerful: instead of enrolling new patients, leverage the rich data from previously completed RCTs that already have the outcomes needed to evaluate the AI model.
How It Works
The BRIDGE design involves several key steps:
- Identify a Suitable "Donor" Trial: The first step is to find a completed RCT that collected the input data required by the AI model (e.g., imaging, genomic data, clinical variables) and followed patients long enough to capture the outcome the model is designed to predict (e.g., disease recurrence, mortality).
- Define the Validation Hypothesis: Before accessing the data, the AI model developer pre-specifies the validation hypothesis. This includes the performance metric (e.g., AUC, sensitivity, specificity), the threshold for success, and the statistical analysis plan.
- Apply the AI Model to the Donor Trial Data: The AI model, which must have been developed and locked before accessing the donor trial data, is applied to the patient data from the donor trial. The model generates predictions for each patient.
- Evaluate Model Performance: The model's predictions are compared against the actual patient outcomes recorded in the donor trial to assess performance.
Statistical Rigor
A critical feature of the BRIDGE design is its attention to statistical rigor. The methodology includes provisions for:
- Type I Error Control: Ensuring that the probability of falsely claiming that a bad model is good is kept at an acceptable level (e.g., 5%).
- Sample Size Calculations: Determining how many patients from the donor trial are needed to adequately power the validation study.
- Handling of Multiple Comparisons: If multiple models or endpoints are being evaluated, appropriate adjustments are made.
Case Studies
The original BRIDGE paper demonstrated the design's applicability across several therapeutic areas.
Oncology: Breast Cancer Recurrence Prediction
A genomic-based AI model was developed to predict the risk of distant recurrence in early-stage breast cancer patients. Using data from a large, completed adjuvant therapy RCT, the BRIDGE design allowed for independent validation of the model's prognostic accuracy. The study confirmed the model's ability to stratify patients into high- and low-risk groups, providing evidence that could support clinical adoption.
Cardiovascular Disease: Heart Failure Risk
An AI model trained on EHR data to predict 1-year heart failure hospitalization was validated using data from a completed RCT of a heart failure medication. The BRIDGE validation showed strong discrimination (AUC > 0.80) and good calibration, providing evidence for the model's potential use in guiding therapy intensity.
| Therapeutic Area | Model Type | Outcome | Validation Result |
|---|---|---|---|
| Oncology (Breast) | Genomic Risk Score | Distant Recurrence | AUC 0.78 |
| Cardiovascular | EHR-based ML | HF Hospitalization | AUC 0.82 |
| Sepsis | ICU Vital Signs ML | Sepsis Onset | Sens 0.85, Spec 0.72 |
Pros and Cons of the BRIDGE Design
Advantages
- Dramatically Reduced Cost: Leveraging existing data eliminates the need for new patient enrollment, site fees, and lengthy follow-up periods.
- Faster Time to Validation: Validation can be completed in months rather than years.
- High-Quality Data: Data from RCTs is typically of higher quality than real-world data, with rigorous outcome ascertainment and standardized data collection.
- Regulatory Acceptability: Regulators are increasingly open to leveraging existing trial data, especially when the validation is prospectively planned and rigorously executed.
Limitations
- Data Availability: Not all AI models will have a suitable donor trial available. The trial must have collected the right inputs and followed the right outcomes.
- Generalizability Concerns: Patients in RCTs are often not representative of the broader population that would use the AI model in practice. Selection criteria, comorbidities, and care settings may differ.
- Temporal Drift: Medical practice evolves. Data from a trial completed 10 years ago may not reflect current standards of care, potentially affecting the AI model's performance.
- Requires Locked Model: The AI model must be finalized before accessing the donor trial data. Any subsequent model modifications would require a new validation study.
When is BRIDGE Appropriate?
The BRIDGE design is best suited for validating AI models that:
- Use input data commonly collected in clinical trials (e.g., standard imaging, common lab values, genomic panels).
- Predict outcomes that are standard endpoints in RCTs (e.g., survival, disease recurrence, major adverse events).
- Are intended for patient populations that are well-represented in existing trial databases.
Implementation Roadmap for AI Model Developers
For organizations looking to leverage the BRIDGE design, here is a practical roadmap:
Step 1: Inventory Potential Donor Trials
Before finalizing your AI model, conduct a systematic search for completed RCTs that could serve as donor trials. Key databases include ClinicalTrials.gov, the Yale University Open Data Access (YODA) project, and pharmaceutical company data sharing platforms (e.g., Vivli, CSDR).
Step 2: Engage Early with Regulators
If you intend to use a BRIDGE validation to support a regulatory submission, engage with the FDA (via Pre-Submission meetings) or EMA early in the process. Discuss your proposed validation plan and get feedback on its acceptability.
Step 3: Pre-Specify Everything
The credibility of a BRIDGE validation depends on pre-specification. Before accessing donor trial data, finalize and document your AI model, your performance metrics, your success thresholds, and your statistical analysis plan. Consider registering your validation study publicly (e.g., on ClinicalTrials.gov or a preprint server).
Step 4: Execute and Report Transparently
Apply your locked model to the donor trial data and analyze the results according to your pre-specified plan. Report results transparently, including any sensitivity analyses and assessments of generalizability.
Need Help with AI Model Validation?
CTDSU's biostatistics team has extensive experience in designing and executing AI validation studies, including the BRIDGE methodology.
Conclusion
The BRIDGE trial design represents a significant methodological advance for the clinical AI community. By enabling rigorous, independent validation of AI models using existing RCT data, it addresses one of the key bottlenecks in translating AI research into clinical practice.
While not a panacea—data availability and generalizability remain important considerations—the BRIDGE design offers a powerful tool for AI model developers, particularly in therapeutic areas where large, well-annotated RCT datasets exist. As the field matures, we expect to see this and similar data-reuse methodologies become increasingly central to the AI validation toolkit.
Key Takeaways:
- BRIDGE trials reuse data from completed RCTs to validate AI risk prediction models.
- This approach can dramatically reduce validation costs (from millions to tens of thousands of dollars) and timelines (from years to months).
- Key requirements include a suitable donor trial, a locked AI model, and a rigorously pre-specified analysis plan.
- Limitations include data availability, generalizability concerns, and temporal drift.
Stay up-to-date on the latest in clinical trial methodology. Subscribe to the CTDSU newsletter for weekly insights.