Wenxin Zhang and Mark van der Laan

Please see the attached comment from Wenxin Zhang and Mark van der Laan.

Comment on “Considerations for the Regulation of
Generative AI-Enabled Medical Devices: Discussion
Paper and Request for Feedback”
Wenxin Zhang ∗1,3 and Mark van der Laan †2,3
Department of Epidemiology and Biostatistics, University of California,
San Francisco
Division of Biostatistics, University of California, Berkeley
Center for Targeted Machine Learning and Causal Inference, University of
California, Berkeley

October 19, 2026

We appreciate the FDA’s timely discussion paper [1] on the regulation of generative
artificial intelligence (GenAI)-enabled medical devices, particularly its consideration of riskbased assessment, clinical confirmation, and appropriate performance standards. As with other
medical products, evaluating the safety and effectiveness of GenAI-enabled devices requires
clearly defining the scientific questions relevant to regulatory decisions and determining what
evidence can credibly address them. The diverse functions and clinical interactions of GenAIenabled devices introduce additional challenges to this process. In this comment, we highlight
how systematic, question-driven causal and statistical thinking can help translate these
regulatory considerations into well-defined evaluation targets, guide appropriate evidence
generation, and support reliable statistical estimation and inference.
One established framework for organizing this question-driven statistical thinking is the
Causal Roadmap [2], which provides an explicit process from formulating scientific questions
and defining targets of estimation to planning studies and analyses aimed at generating
high-quality evidence. It includes key steps of defining the scientific question and statistical
target of estimation, characterizing the processes that generate data to answer that question,
articulating the assumptions required to give the results a causal interpretation, selecting
appropriate statistical analyses, and prespecifying sensitivity analyses. The roadmap brings
together concepts developed over decades of research in causal inference, experimental and
observational study design, and statistical estimation and inference; it also aligns with existing
regulatory frameworks (see [2] for an overview and further discussion).
∗
wenxin.zhang@ucsf.edu
†
laan@berkeley.edu
 A recent approach to evaluating diagnostic tests and AI-enabled medical devices [3]
illustrates how the question-driven statistical thinking underlying the Causal Roadmap can be
applied to evaluation questions relevant to the FDA’s discussion paper. For medical tests and
devices that provide diagnostic information, two important questions arise beyond diagnostic
accuracy: whether the additional information reveals treatment-effect heterogeneity beyond
existing information, and whether using this information to guide treatment decisions improves
patient outcomes. Following the roadmap, these questions can be formulated as two distinct
causal evaluation targets: explanatory effectiveness, which quantifies the additional treatmenteffect heterogeneity explained by diagnostic information beyond a reference information set,
and pragmatic effectiveness, which quantifies the difference in expected outcomes under an
optimal personalized treatment strategy informed by both existing and additional diagnostic
information versus a reference strategy based on existing information alone. This type of
approach can also accommodate comparisons of a range of treatment strategies that reflect
how the additional information is used to guide downstream decisions in different settings.
More broadly, this causal inference perspective highlights two interconnected roles of
AI-enabled medical devices, including GenAI-enabled devices, in clinical decision-making:
information augmentation, which enriches the information available for clinical decisions, and
decision-rule optimization, which uses artificial intelligence and machine learning algorithms
to learn or improve decision rules for selecting clinical actions based on available patient
information. An AI-enabled device may contribute through either role or both.
The causal and statistical thinking underlying the roadmap, as illustrated by this example,
is also relevant to the two-axis risk assessment framework discussed in Section IV of the
discussion paper, which considers device functions ranging from non-directive informational
functions to “action-directing” and “action-taking” functions, alongside the potential consequences of relying on incorrect outputs. By explicitly considering the roles of different device
functions in the causal pathways from information to clinical decisions and patient outcomes,
this statistical thinking helps clarify how these functions may affect clinical care and lead
to benefits or harms. Applying the causal roadmap can also help formulate well-defined
evaluation questions and estimands to quantify these potential benefits and harms relative
to appropriate comparators, as discussed in Section V.D.1 and Questions 14 and 15 of the
discussion paper.
Importantly, the statistical reasoning underlying the roadmap can help assess whether a
well-defined causal or statistical estimand can be identified from available or proposed data
under plausible assumptions, whether the corresponding analysis plan can support reliable
estimation and valid statistical inference, and what additional evidence may be needed [2].
This is directly relevant to the clinical confirmation approaches discussed in Section V.C.1
of the discussion paper, including retrospective evaluations and prospective clinical studies
potentially incorporating real-world data.
Overall, systematic statistical thinking provides a foundation for formulating scientific
questions that reflect regulatory objectives, defining meaningful evaluation targets, and
guiding appropriate evidence generation with rigorous statistical estimation and inference.
As new evaluation questions arise in regulatory decision-making, a structured causal roadmap
can help clarify what needs to be learned, what can be identified from the available data, how
reliably and efficiently the target estimands can be estimated, and where additional evidence
may be needed to evaluate the safety and effectiveness of emerging medical technologies.
References
[1] U.S. Food and Drug Administration. Considerations for the Regulation of Generative
AI-Enabled Medical Devices: Discussion Paper and Request for Feedback. Docket No.
FDA-2026-N-7874, August 2026. URL https://www.fda.gov/media/194242/download.

[2] Lauren E Dang, Susan Gruber, Hana Lee, Issa J Dahabreh, Elizabeth A Stuart, Brian D
Williamson, Richard Wyss, Iván Dı́az, Debashis Ghosh, Emre Kıcıman, et al. A Causal
Roadmap for Generating High-Quality Real-World Evidence. Journal of Clinical and
Translational Science, 7(1):e212, 2023.

[3] Wenxin Zhang, Rachael Phillips, and Mark van der Laan. A Causal Inference Approach for
Evaluating Diagnostic Tests and AI-Enabled Medical Devices: From Effect Modification
to Information-Augmented Decision-Making. arXiv preprint arXiv:2608.19501, 2026.