FDA Guidance on Using AI/ML in Clinical Trial Data Analysis

By Rhizome Team

Artificial intelligence and machine learning technologies are increasingly being considered for use in clinical trial data analysis, from patient monitoring to endpoint detection and statistical analysis. As these technologies evolve, sponsors and researchers need clear guidance on how FDA views their application in the clinical trial context.

The FDA has provided commentary on AI/ML tools through various guidance documents, statements, and regulatory frameworks. Understanding the agency's current position helps sponsors determine when and how these technologies can be appropriately integrated into clinical trial protocols and data analysis plans.

Here we examine FDA's public statements and guidance regarding the use of AI/ML in clinical trial data analysis, including considerations for validation, transparency, and regulatory acceptability of these emerging analytical approaches.

Want to read the citations or ask Rhizome a follow-up question? Schedule a call

What has FDA said about using AI/ML in clinical trial data analysis?

Answer

FDA has not (in the provided excerpts) issued a single, clinical-trial-specific rule that “endorses” AI/ML for trial analysis. Instead, FDA’s public guidance statements relevant to using AI/ML to analyze clinical study data focus on (1) where AI/ML may be used, (2) documentation and transparency expectations, and (3) data quality/bias controls.

  1. FDA recognizes AI/ML can be used to process clinical data (including unstructured records) for studies, but does not endorse specific AI tools
  • In guidance on using EHR/claims real-world data for drug/biologic regulatory decision-making, FDA notes that technological advances in AI may enable more rapid processing of unstructured electronic health care data, including using NLP/ML/deep learning to extract data elements from unstructured EHR text, develop algorithms to identify outcomes, or evaluate images/lab results 23. FDA explicitly states it “does not endorse any specific AI technology” 23.
  1. FDA expects rigorous documentation/QA/QC when automated (including AI-enabled) methods are used to retrieve/transform data for analysis
  • FDA recommends that all manual and automated data retrieval and transformation processes be thoroughly assessed from data collection through final report writing to ensure data integrity, and that sponsors ensure curation/transformation do not alter meaning or lose important context 27.
  • FDA also says documentation of processes used to mine/evaluate unstructured data should describe the techniques employed (e.g., NLP) to abstract unstructured data and supplement structured data 27.
  • FDA further recommends that processes for managing/preparing the final study-specific analytic dataset be described in the protocol or analysis plan, and that submitted analysis programs be thoroughly annotated to explain the intent of each data management/analysis step 27.
  1. FDA highlights representativeness and bias as key issues when AI/ML is involved (especially for AI models and data pipelines)
  • In FDA’s AI-enabled device lifecycle/marketing submission draft guidance, FDA explains that AI system performance depends heavily on the quality, diversity, and quantity of training/tuning/testing data, and that data management is important for identifying and mitigating bias (including risks like overfitting to site- or subgroup-specific artifacts and underrepresentation of certain populations) 6.
  • In FDA’s final guidance on Predetermined Change Control Plans (PCCPs) for AI-enabled device software functions, FDA recommends data practices that include sequestering training/tuning/test data to prevent overfitting and “misquotes” of test performance, and methods aimed at identifying/mitigating unwanted bias; it also lists bias-mitigation approaches such as cross-validation, bootstrapping, bagging, ensembling, and synthetic/augmented data 11. While this is written for AI-enabled devices, the principles FDA emphasizes (separation of datasets, bias mitigation, representativeness) are directly relevant to AI/ML-based analytical pipelines used to generate clinical evidence 11.
  1. FDA encourages clear descriptions of AI/ML methodology, data sources, and bias controls in submissions when AI/ML is used
  • In FDA’s guidance on the content of premarket submissions for device software functions, FDA recommends describing the analysis methodology when software performs signal/pattern/image analysis, including whether it uses AI/ML, neural networks, and whether algorithms are fixed (“locked”) or adaptive 22. FDA also recommends describing what data informed the model(s), how/when/where data were collected, steps taken to identify/address potential biases/limitations, and approaches used to provide transparency about development/performance/limitations 22.
  • Although this is framed for device software submissions, it reflects FDA’s broader expectation that when AI/ML is used to generate evidence, sponsors should be able to explain the method, data provenance, limitations, and bias controls 22.

Bottom line from FDA statements in these sources: FDA acknowledges AI/ML can be used to extract and analyze clinically relevant data (notably unstructured EHR text) for evidence generation, but emphasizes that sponsors should not treat AI/ML as a “black box.” FDA expects strong documentation, QA/QC, and transparency about methods and data handling, and it highlights representativeness and bias as central risks to manage; FDA also states it does not endorse specific AI technologies 232722611.