Folia Health's Comment to FDA on AI-Enabled Early-Phase Clinical Trials
On June 25, 2026, Nell Meosky Luo, Founder and CEO of Folia Health, submitted a formal comment to the U.S. Food and Drug Administration in response to its Request for Information on the AI-Enabled Optimization of Early-Phase Clinical Trials Pilot Program (Docket FDA-2026-N-4390-0001). In the comment, Folia argues that AI models built to optimize early-phase trial design and decision-making are only as reliable as the data they're trained on, and that traditional visit-based data collection leaves gaps between clinic visits that limit what these models can see. The full comment, submitted ahead of the docket's June 29, 2026 deadline, is below.
June 25, 2026
Dockets Management Staff (HFA-305)
Food and Drug Administration 5630 Fishers Lane, Rm. 1061
Rockville, MD 20852
Re: Docket No. FDA-2026-N-4390 AI-Enabled Optimization of Early-Phase Clinical Trials Pilot Program; Request for Information
To Whom It May Concern:
Folia Health appreciates the opportunity to comment on the proposed AI-Enabled Optimization of Early-Phase Clinical Trials Pilot Program. The success of “AI-enabled optimization” is bounded by whether the underlying data represents the full impact of a drug candidate on the health status and disease burden of the patient. In current trial designs, the most consequential evidence gap is the patient experience between visits (e.g., symptoms, exacerbations, and flares that are experienced and treated at home; the side effects that never reach the next case report form, the functional setbacks that resolve before the next visit, and the tolerability signals that drive silent discontinuation). Very few people would agree that their clinical records, or the information collected about them in the course of a clinic visit, directly illustrates their disease burden and treatment experience. We write to argue that closing this gap is critical, as the AI-enabled pilot program is fundamentally limited by the quality of its inputs.
Additionally, the FDA has spent more than a decade establishing patient experience data as research-grade evidence. The Patient-Focused Drug Development (PFDD) guidance series formalizes these methods, particularly Guidance 3 (Select, Develop or Modify Fit-for-Purpose Clinical Outcome Assessments) and Guidance 4 (Incorporating Clinical Outcome Assessments into Endpoints for Regulatory Decision-Making). The Commissioner’s vision of “continuous trials,” in which the boundaries between traditional phases are streamlined or eliminated, depends on this additional evidence layer: the adaptive go/no-go decisions, futility analyses, and dose-finding heuristics that make continuous designs viable hinge on early tolerability and functional signals that visit-based data cannot capture in time to be useful. The AI-enabled pilots should utilize the insights made possible by following these PFDD guidance.
Finally, the broader modernization pressure now driving real-time data capture at the trial site level applies equally to patient experience data. The strongest platforms will be those that support real-time, contextualized, auditable, and interoperable research operations in place of batch-oriented, downstream reporting. Patient experience data (PED) has historically been the most batch-oriented signal stream in a clinical trial - administered as standalone surveys at visits, often retrospectively, sometimes on paper, and reconstructed from patient memory that may be exposed to recall bias. It is therefore the area where the shift to home-reported, point-of-experience capture produces the largest gain in signal quality. Capturing evidence on the patient experience in real-time - and sharing the resulting signals in real-time - should become the new standard in drug development.
Folia is a 10-year-old patient-driven health informatics organization founded by family caregivers that collects structured, actionable home-reported outcomes to support data-driven care for the individual and drug development anchored in what matters most to patients. . We work with biopharmaceutical sponsors to capture quantitative, longitudinal patient-reported outcomes and patient-generated health data between visits, across rare disease, immunology, cardiometabolic, neurology, dermatology, oncology, and other therapeutic areas. Our platform produces statistically rigorous, research-grade evidence at the patient level: symptom trajectories, functional status, tolerability signals, adherence, and the descriptive texture of daily life on therapy. These data generate signals alongside and in conjunction with clinical measures that, when measured longitudinally, inform a drug’s efficacy, safety, and tolerability. Our comments below respond to five RFI questions where longitudinal patient data is most necessary.
Question A.1.c: Priority AI Use Cases
The RFI asks whether priority should be given to specific AI use cases. We recommend that continuous, real-time capture and integration of structured patient experience data be designated as an explicit priority use case for the pilot.
AI applications downstream of clinical data (adaptive randomization, safety signal detection, retention prediction, sub-population identification) are bounded by the quality and completeness of the inputs they ingest. Clinical and biomarker signals at scheduled visits are necessary but not sufficient: they cannot tell the model what happened between visits, when the majority of a therapy’s impact is experienced; they cannot reveal that an endpoint stable visit-over-visit may mask substantial day-to-day variability; and they cannot detect that a coming dropout was already visible in patient-reported engagement metrics three weeks earlier.
These are not purely hypothetical examples. They reflect the routine information losses that prevent visit-based trials from generating high-resolution evidence. The principle underlying continuous data capture at the trial site applies symmetrically at the patient level: site-level data is most reliable when captured at the point of care, integrated into the clinical workflow rather than reconstructed at the end of a visit; patient experience data is most reliable when captured at the point of experience - when a symptom occurs, a side effect emerges, or a functional setback happens - not when it is recalled at the next scheduled visit days or weeks later, after intervening events have altered the memory. Continuous, structured, point-of-experience patient data, consistent with the PFDD guidance series, closes this gap. AI systems that integrate continuous patient-experience data with clinical and biomarker measures may provide an important opportunity to improve early understanding of treatment response, treatment burden, and the real-world relevance of emerging efficacy signals. We recommend that the pilot designate this category as a priority use case and allocate at least one pilot slot to a sponsor whose proposed protocol embeds a continuous, real-time patient-experience data layer.
Question A.3.c: Role of Patient Groups in AI Governance
The RFI asks about the role of patient groups in AI governance - that is, how patients, patient advocacy organizations, and navigators participate in overseeing the AI systems used in clinical trials. In addition to that oversight role, we believe patients must participate in data capture itself. Both matter, but they require different infrastructure, different evaluation, and different commitments from participating sponsors.
Patient advocacy organizations and certified patient navigators play an essential governance role, particularly in eligibility design, communication of AI-generated assessments, and identification of where the AI system fails the patient experience. We endorse the comments submitted to this docket on that subject.
But the patient is not only a governance stakeholder. The patient is also the primary source of the richest data the study generates - and the only source of much of it. An AI-enabled trial pilot that incorporates patient governance without incorporating patient-generated data has solved the easier half of the problem. We recommend that the pilot’s governance framework explicitly require participating sponsors to articulate, for each AI use case in their proposed protocol, both (a) how patients participate in oversight of the system, and (b) how patient-generated data informs the system’s inputs. The two should be evaluated as paired commitments.
Question B.1.c: Measuring Improvements in Screening, Recruitment Efficiency, and Retention
The RFI asks how improvements in screening, recruitment efficiency, and retention can be quantified. Retention is the metric where between-visit patient experience data has the most direct, measurable impact. We recommend that the pilot evaluation framework include the following:
Between-visit data completeness. The proportion of scheduled between-visit patient-experience reports that are completed, stratified by indication, age, and visit cadence. This is the operational integrity metric of any real-time patient experience evidence layer and a leading indicator of whether the data infrastructure is fit for purpose.
Early tolerability detection lead time. The number of days between the first patient-reported signal of a tolerability event and the first appearance of that signal in clinic-visit-derived data. A pilot that demonstrates non-trivial lead times has demonstrated that real-time patient data adds regulatorily meaningful information.
We also recommend that participating sponsors be required to report subgroup-specific values for each of these metrics, consistent with the fairness expectations articulated in B.5.e of the RFI. Patient-experience data is particularly vulnerable to differential missingness across age, language, digital access, and caregiver-mediated reporting; surfacing these differentials should be a standing requirement of the pilot.
Question B.2: Decision Quality
The RFI asks how the contribution of AI-supported decisions to trial quality should be measured. We submit that continuous patient-experience data materially improves AI-supported decision quality, and that the pilot's evaluation framework should include metrics that capture this improvement directly.
Where AI models have access to real-time patient-reported signals in addition to clinical and biomarker data, they can identify meaningful response patterns, detect subgroup differences, and surface non-responders earlier in the trial - gains that are difficult to demonstrate when models are trained on visit-only data. We recommend the pilot evaluation framework include the following:
Time to identification of meaningful treatment-response patterns. The number of weeks from trial initiation to the point at which an AI-supported model reliably identifies a treatment-response pattern of regulatory interest. Pilots should report this metric for models trained with and without real-time patient-experience inputs, isolating the contribution of the patient-data layer.
Time to identification of patient subgroups with differential response. The number of weeks from trial initiation to the point at which an AI-supported model identifies clinically meaningful subgroups with differential response patterns. Earlier identification supports adaptive design decisions and dose-finding heuristics.
Concordance between early patient-experience signals and later efficacy outcomes. The agreement between early patient-reported signals (collected in the first weeks of treatment) and later, formally adjudicated efficacy endpoints. High concordance demonstrates that patient-experience data is a valid early indicator of regulatorily meaningful outcomes.
Concordance between early patient-experience signals and later discontinuation patterns. The agreement between early patient-reported tolerability and adherence signals and eventual discontinuation events. High concordance supports the use of patient-experience data in early go/no-go decisions.
Earlier identification of non-responders. The lead time, in weeks, by which AI-supported models incorporating continuous patient-experience data identify non-responders relative to traditional, visit-based approaches. This metric captures the most direct operational value of the patient-data layer in adaptive designs.
We recommend that participating sponsors be required to report these metrics for AI systems trained on combined clinical-and-patient-experience data, with appropriate comparators where feasible, so that the pilot can demonstrate empirically whether the patient-data layer improves decision quality.
Question B.5.c: Transparency and Explainability
The RFI asks how transparency and explainability of AI systems should be evaluated. We submit that the answer depends in significant part on the nature of the inputs the AI system is built on. Patient-reported and patient-generated data provide a uniquely explainable layer of evidence because they are directly attributable to the patient's lived experience rather than derived from latent model features. A symptom score traceable to a specific patient's report of how they felt on a specific day carries a structurally different kind of interpretability than an output derived from high-dimensional, opaque transformations of clinical or imaging data.
AI systems that integrate real-time patient-experience data with clinical and biomarker measures provide an important opportunity to improve early understanding of treatment response, treatment burden, and the real-world relevance of emerging efficacy signals. Where the model's outputs can be decomposed into contributions from named, patient-attested experiences, the explainability burden becomes more tractable: reviewers can ask which patient-reported signals are driving a model's recommendation and verify against the patient's actual reports rather than against post-hoc model rationalizations.
We recommend that the pilot evaluate explainability not only in terms of model-internal interpretability but also in terms of input-side traceability - the extent to which an AI output can be traced back to specific, patient-attested inputs collected through fit-for-purpose instruments. Participating sponsors should be required to document, for each AI use case, how patient-reported and patient-generated data contribute to the system's outputs, and where in the output a patient could in principle recognize their own reported experience.
Conclusion
We support the Agency’s pursuit of an AI-enabled, continuous, real-time clinical trial paradigm. We submit that this paradigm cannot achieve its stated objectives - faster, more adaptive, more patient-relevant evidence generation - on clinical and biomarker signals alone, but that we must instead place the self-reported symptom and treatment experiences of patients on the same footing, if we are to establish the appropriate context to evaluate therapies on whether they are moving the needle on what matters most to patients. The patient-experience data layer is a co-equal real-time data stream, grounded in a decade of PFDD guidance and years of peer-reviewed research in home-reported outcomes across more than a dozen indications by organizations such as Folia Health, and the pilot’s design should treat it as such.
We add one further observation drawn from our own development experience. Continuous, auditable, interoperable patient-experience data infrastructure takes years to build, not months. Sponsors selected for the pilot will need to deploy this capability at protocol activation, not construct it during the pilot. We recommend that the selection criteria reflect this reality and prefer proposals that integrate with already-operational patient-experience data infrastructure rather than those that propose to build it concurrently with trial conduct.
Concretely, we recommend that the FDA:
Designate continuous, real-time capture of structured patient experience data as a priority AI use case in the pilot.
Allocate at least one pilot slot to a protocol that incorporates a fit-for-purpose, continuous, real-time patient-experience data layer.
Distinguish in the governance framework between patient roles in AI oversight and patient roles as data sources, and require sponsors to address both.
Adopt evaluation metrics that cover both operational integrity of the patient-experience data layer (between-visit data completeness, tolerability detection lead time, PFDD alignment) and the decision quality it enables (time to identification of response patterns and subgroups, signal–outcome concordance, earlier identification of non-responders).
Evaluate explainability of AI systems by input-side traceability - the extent to which outputs can be traced to specific, patient-attested inputs collected through fit-for-purpose instruments.
Reference the FDA Patient-Focused Drug Development guidance series, particularly Guidance 3 and Guidance 4, as evaluation references in the pilot's selection criteria.
Prefer pilot proposals that integrate with already-operational, continuous, auditable, interoperable patient-experience data infrastructure rather than those that propose to construct it during the pilot.
We welcome the opportunity to provide further input or discuss these comments with Agency staff.
Respectfully submitted,
Nell Meosky Luo
Founder and Chief Executive Officer
Folia Health
foliahealth.com