Findings
The document introduces a risk-based credibility assessment framework for establishing and evaluating the credibility of an Artificial Intelligence (AI) model's output when used to support regulatory decisions regarding drug safety, effectiveness, or quality. The framework outlines a 7-step process beginning with defining the question of interest and the Context of Use (COU). Credibility is defined as trust, established through evidence, in the AI model's performance for a particular COU. The credibility assessment is tailored to the AI model risk, which is a combination of model influence (the AI model's evidence contribution relative to other evidence) and decision consequence (the significance of an adverse outcome from an incorrect decision). The document highlights challenges with AI use, including variability in development datasets (training/tuning), the need for methodological transparency due to model complexity, difficulty in quantifying and interpreting uncertainty in model output, and the potential for performance change over time (data drift), which necessitates life cycle maintenance.

Recommendations
Sponsors and interested parties should define the question of interest and clearly define the COU, detailing the AI model's specific role and scope and whether other information will be used. They should assess the AI model risk (low, medium, or high) to ensure that subsequent credibility assessment activities (Step 4) are commensurate with that risk and tailored to the COU. For Step 4, the credibility assessment plan should include a description of the model, model development process (including inputs, architecture, feature selection, and rationale), and data used (training and tuning data). Development data must be deemed fit for use (relevant and reliable) to mitigate issues like algorithmic bias. The plan should also detail the model evaluation process using independent test data and include performance metrics with confidence intervals, an estimate of uncertainty, and a description of model limitations. Early engagement with the FDA is strongly encouraged to discuss model risk and the adequacy of the credibility assessment plan.

Regulatory Considerations
The risk-based credibility assessment framework is intended to help organize and document information for regulatory submissions. The required stringency of assessment activities and the level of documentation should be commensurate with the AI model risk. For AI models whose performance can change over time (e.g., in pharmaceutical manufacturing or postmarketing), sponsors must implement life cycle maintenance plans to monitor performance and manage changes in a risk-based manner. Changes to AI models should be evaluated through the manufacturer's change management system and may require re-execution of parts of the credibility assessment plan. Early engagement can be facilitated through formal meetings (e.g., Pre-IND) or other specialized programs listed in the guidance, such as the Center for Clinical Trial Innovation (C3TI), the Model-Informed Drug Development (MIDD) Paired Meeting Program, and the Emerging Technology Program (ETP) or Advanced Technologies Team (CATT).

Findings
The guidance applies to four types of Clinical Outcome Assessments (COAs): Patient-Reported Outcomes (PROs), Observer-Reported Outcomes (ObsROs), Clinician-Reported Outcomes (ClinROs), and Performance Outcomes (PerfOs). A COA is considered fit-for-purpose when the validation evidence is sufficient to support its context of use (COU). To determine if a COA is fit-for-purpose, sponsors must clearly describe the Concept of Interest (COI) and the COU, and present sufficient evidence to support a clear rationale for the COA's proposed interpretation and use. The rationale for using a COA should include up to eight components, such as justification for the COA type, capturing the important parts of the COI, appropriate administration and scoring, minimal influence from irrelevant factors or measurement error, and correspondence with the Meaningful Aspect of Health (MAH). The most direct assessment of how a patient feels or functions (MAH) should be used as the COI whenever possible.

Recommendations
Sponsors should use the Roadmap to Patient-Focused Outcome Measurement to guide the selection, modification, or development of a COA. The process begins with understanding the disease/condition (including patient perspectives) and conceptualizing clinical benefits and risks (defining the MAH, COI, and COU). When feasible, existing COAs are generally preferred, especially for well-established COIs, as this approach is often the least burdensome. If an existing COA is modified or used in a different context, additional evidence (e.g., cognitive interviews, psychometric studies) must be collected to justify its fitness for the new context of use. For new COA development, sponsors should involve patients, document all steps, and generally avoid using the new COA for the first time in a registration (pivotal) trial; a standalone observational study or early phase trial is recommended for evaluation.

Regulatory Considerations
Sponsors are encouraged to interact early and throughout medical product development with the relevant FDA review division to ensure COAs are appropriate for the intended COU. Sponsors should communicate their proposed COA-based endpoint approach, including the MAH, COI, COA type/name/score, and the final COA-based endpoint, ideally using the suggested format. The type and amount of evidence required to support the rationale for a COA's use is weighed against the degree of uncertainty regarding that part of the rationale. For ClinROs, it is recommended to use an assessor masked to treatment assignment and study visit for primary endpoints, if feasible. FDA strongly discourages proxy-reported measures for concepts known only to the patient (e.g., pain) and recommends using an ObsRO to measure observable behaviors instead when the patient cannot self-report.

Recommendations
Clearly define the concept of interest and its context of use to ensure COAs align with trial objectives.
Use conceptual and measurement frameworks to communicate how COAs measure patient experiences and generate interpretable scores.
Leverage existing COAs where possible, modifying them only when justified, and document all modifications rigorously.
Ensure COAs are accessible and inclusive, incorporating features like large fonts, touch interfaces, or audio assistance for diverse populations.
Conduct early engagement with FDA to discuss COA selection, development, and validation plans.

Regulatory Considerations
Fit-for-purpose validation requires evidence of conceptual alignment, scoring reliability, and sensitivity to clinically meaningful changes.
Digital health technologies used for COAs must comply with FDA’s guidance on data integrity, usability, and technical performance.
COAs intended for regulatory submissions must be developed and validated before pivotal trials to avoid jeopardizing trial outcomes.
Modifications to COAs or scoring methods during trials necessitate justification and revalidation.
Sponsors should submit comprehensive documentation on COA development, including scoring algorithms and item tracking matrices.

Findings
The use of Artificial Intelligence (AI) and Machine Learning (ML) is being applied to a broad range of drug development activities with the potential to accelerate the process and make clinical trials safer and more efficient. The inclusion of AI/ML is most common in the clinical development/research phase of regulatory submissions. Concerns exist that AI/ML algorithms could amplify errors and preexisting biases in underlying data sources, which raises issues related to generalizability and ethical considerations. Other challenges include limited explainability due to model complexity and proprietary reasons, as well as managing risks related to data quality, reliability, and representativeness. The FDA recognizes that a careful, risk-based assessment of the specific context of use (COU) is needed when evaluating AI/ML.

Recommendations
Stakeholders should adhere to practices in three key areas: human-led governance, accountability, and transparency; quality, reliability, and representativeness of data; and model development, performance, monitoring, and validation. A risk management plan should be applied to identify and mitigate risks based on the COU, guiding the level of documentation and transparency. Practices are needed to ensure the integrity of AI/ML and address issues like bias and missing data. For models, developers should use pre-specification steps and clear documentation for development and assessment criteria. Models must be monitored over time for reliability and consistency, and Real-World Data (RWD) performance can provide valuable feedback, including for potential re-training.

Regulatory Considerations
The FDA encourages early engagement through mechanisms like the Critical Path Innovation Meetings (CPIM), ISTAND Pilot Program, and Emerging Technology Program to discuss relevant AI/ML methodologies or technologies. The Verification and Validation (V&V 40) risk-informed credibility assessment framework and the principles for Good Machine Learning Practices (GMLP), while not specific to drug development, are helpful guides for evaluating models. The industry is exploring the use of a Predetermined Change Control Plan (PCCP) mechanism for AI/ML-based devices to proactively specify and manage modifications, enhancing adaptability. In general, a risk-based approach should guide the level of evidence and record keeping needed for the verification and validation of AI/ML models for a specific COU.