Our FDA Comment: AI-Enabled Optimization of Early-Phase Clinical Trials Pilot Program

Dear Sir or Madam:

Cureledger, Inc. (Cureledger) appreciates the opportunity to submit the following comments to the U.S. Food and Drug Administration (FDA or Agency) in response to the request for information titled “AI-Enabled Optimization of Early-Phase Clinical Trials Pilot Program” (the RFI). Cureledger is a life data trust dedicated to advancing the development of safe and effective therapies for people living with rare diseases.

I. GENERAL COMMENTS

Cureledger commends the Agency on the issuance of the RFI and believes the proposed pilot has significant potential to improve the efficiency, safety, and decision quality of early-phase clinical trials. The Agency’s recognition that rare-disease trials are among the use cases best positioned to benefit from AI is consistent with the Agency’s prior guidance on natural history studies and on the development of drugs and biological products for rare diseases, both of which name small populations and sparse longitudinal data as central challenges in this setting. Cureledger believes the pilot can produce durable, generalizable evidence about the use of AI in early-phase development if it is scoped, evaluated, and supported in a manner that reflects the practical conditions of small-population research.

Cureledger believes the evidence the pilot generates must satisfy three conditions simultaneously: it must rest on data that is contextualized, longitudinal, and contributor-verified; it must be evaluated against decision-quality metrics that account for biological and measurement variability at the individual level; and it must be governed by privacy and data-infrastructure protections appropriate to small rare-disease populations. The specific comments that follow address these conditions and offer recommendations for the Agency’s consideration.

II. SPECIFIC COMMENTS

1. Pilot Scope and Priority Use Cases (A.1.a, A.1.b, A.1.c, B.1.a)

Cureledger notes that the RFI identifies rare-disease trials as use cases where AI is well positioned to improve trial design and decision-making. Cureledger believes rare disease is where the pilot’s most generalizable evidence will be produced. Small, geographically dispersed populations combine high decision uncertainty, limited enrollment, and sparse longitudinal data, and AI methods that perform well in these conditions will translate to less constrained therapeutic areas.

Cureledger recommends that the pilot include at least one small rare-disease population with well-characterized genetics and objective, repeatable biomarkers. Conditions identified at birth through newborn screening carry an additional advantage that the Agency may wish to consider: enrollment is measurable against a known denominator, which allows efficiency metrics to be reported in absolute terms. Among the AI use cases the RFI lists, recruitment, retention, dose-finding, safety monitoring, biomarker assessment, and endpoint validation are the highest-priority candidates in small-population early-phase work.

2. Decision-Quality Metrics for Small Populations (B.1.b, B.1.c, B.2.a, B.2.b, B.2.c, B.3.a, B.3.b, B.3.c, B.6.a, B.6.b)

Cureledger notes that the RFI seeks input on metrics for assessing the quality and timeliness of go/no-go decisions and on methods for evaluating concordance between AI-supported and traditional decisions. Data latency and lack of context is a persistent disadvantage in the ability to develop meaningful therapies for known rare disorders. Cureledger believes that in small-population early-phase studies, the most informative decision-quality metric is whether an observed change exceeds each participant’s own biological and measurement variability before it informs a go or no-go.

Cureledger further believes that real-time, longitudinal, contextualized capture is the data condition that allows such a metric to be applied. With frequent, contributor-verified measurement, including biomarker values together with diet, timing, adherence, and intercurrent illness, each participant’s variability range becomes observable, and a treatment effect must clear that range before it is credited. Within-person comparison also controls for the between-population variability that often confounds rare-disease comparisons, and the Agency has recognized natural history data as an appropriate comparator where concurrent controls are limited.

Cureledger recommends that the pilot pair the Phase 1 to Phase 2 interval with the basis of the decision and track decisions later reversed alongside time saved, consistent with the Agency’s emphasis under Project Optimus. Cureledger further recommends that adverse-event detection be evaluated against the latency of episodic monitoring, recognizing that richer real-time capture will surface events earlier and more completely. The Agency has acknowledged this dynamic in finalizing guidance on clinical trials with decentralized elements. Retention should be measured directly through time in study, completeness of scheduled capture, and longitudinal participation rate, benchmarked against historical attrition in comparable populations. Cureledger recommends that AI-supported and traditional decision rules be pre-specified and run in parallel on the same data, with concordance and divergence reported and explained.

3. Privacy Framework for Small Populations (A.4.b, B.5.d)

Cureledger notes that the RFI seeks input on infrastructure needed to support the pilot and on the evaluation of privacy protections and data governance under the NIST AI Risk Management Framework. Cureledger is concerned that HIPAA Safe Harbor de-identification, as currently defined, may not provide meaningful protection for participants in small rare-disease populations, and the Department of Health and Human Services Office for Civil Rights has acknowledged the limits of Safe Harbor in such cohorts. Cureledger further notes that the Department of Justice final rule implementing Executive Order 14117 reflects a federal determination that bulk human genomic data is among the most sensitive categories of personal data, with thresholds set at 100 U.S. persons. Cureledger believes these federal positions are directly relevant to how the pilot evaluates privacy.

Cureledger recommends that the Agency consider recognizing privacy-preserving technical approaches as acceptable components of a pilot data infrastructure, including pseudonymous identifiers and tamper-evident consent and governance records. These approaches allow multiple sites and sponsors to learn from the same sensitive data while the underlying records remain in place, and they make the NIST AI RMF privacy function inspectable by the pilot. Cureledger believes architectural privacy controls are particularly well matched to the heightened sensitivity of small rare-disease populations.

4. Data Infrastructure for Compounding Evidence (A.3.a, A.3.b, A.3.c, A.4.a, A.4.b, A.6.a, A.6.b)

Cureledger notes that the RFI contemplates pre-competitive collaboration, knowledge sharing, and the role of patient groups and investigators in AI governance. Cureledger believes the pilot’s results will compound only if the data on which they rest remain persistent, verifiable, and reusable across programs. Natural history data generated under the pilot may serve as external control data for subsequent programs, and the value of that data depends on its persistence and verifiability beyond the program that generated it.

Cureledger recommends that the Agency consider data custody models in which the individual contributor’s participation is encouraged across the life of the data and operational custody rests with a fiduciary capable of delivering data to qualified researchers on demand, consistent with the recommendation Cureledger submitted in response to Docket No. FDA-2026-D-1256 on the Plausible Mechanism Framework. The Agency’s framing of fit-for-purpose real-world data sources is consistent with the principle that several developers may derive evidence from the same well-characterized population.

Cureledger further recommends that consent, governance, and data-standard methods be eligible for full disclosure as part of the pilot’s transparency expectations, while sponsor analyses and proprietary model internals remain protected. An interface-level disclosure standard aligns with the Agency’s draft AI guidance and the FDA-Health Canada-MHRA Good Machine Learning Practice principles, and it allows one transparency standard to work across sponsor-developed and proprietary systems.

5. Trustworthiness Assessed at the Interface (A.4.a, A.4.c, A.5.a-c, B.4.a-c, B.5.a-c, B.5.e, B.7.a-c)

Cureledger notes that the RFI aligns the pilot’s trustworthiness evaluation with the NIST AI RMF. Cureledger believes the most testable approach is to evaluate AI systems at the interface to the system: documented inputs, data provenance, a decision rationale a reviewer can inspect, and reproducibility of every result from a fixed dataset version and a logged analysis. Independent regeneration of an AI output from a versioned dataset is a strong form of validity evidence and maps directly onto the NIST AI RMF treatment of valid and reliable AI.

Cureledger recommends that workflow burden, participant trust, and portability across sites be treated as first-class outcomes in the pilot’s qualitative evaluation. Real-time longitudinal capture gives the pilot meaningful configurability across conditions: timelines, checkpoints, and read-out cadence can be tuned to each condition’s natural history, and the durability of the underlying data allows the pilot to adapt to what each condition requires. Cureledger further believes that early and repeated regulatory engagement and clear documentation expectations will be especially valuable to smaller sponsors and patient-community programs participating in the pilot.

III. CONCLUSION

Cureledger appreciates the opportunity to submit these comments and supports the Agency’s commitment to applying AI to improve the efficiency, safety, and decision quality of early-phase clinical trials. Cureledger believes the pilot represents an important step forward and that the final pilot design will be strengthened by addressing the pilot scope, decision-quality metrics, privacy framework, data infrastructure, and trustworthiness considerations outlined above, and looks forward to working with the Agency and other stakeholders to advance these objectives.

Respectfully submitted,

Nina Kilbride, JD
CEO and Founder
Cureledger, Inc.
cureledger.com