Presentient Proposes Pilot in Response to FDAs AI-Enabled Optimization of Early-Phase Clinical Trials Pilot Program RFI

Presentient appreciates the opportunity to comment on FDA’s request for information regarding the proposed AI-Enabled Optimization of Early-Phase Clinical Trials Pilot Program.  

Presentient Technologies Ltd is a London-based biostatistics technology company. Our principal system, BRAKES, is the subject of a Letter of Intent submitted in May 2026 to FDA's ISTAND qualification program for which we are now grateful to have received a response. Presentient is separately engaged with the EMA via their Innovation Task Force. BRAKES detects failing trials substantially earlier than current practice and identifies likely successes, both without unblinding analysts, sponsors or regulators until the point of detection. BRAKES offers a scalable trigger and selection mechanism to enable the Agency to potentially review data in real-time for those trials at the point where success becomes highly likely while avoiding allocating resources to trials highly likely to fail.

BRAKES uses a deterministic algorithm to evaluate unblinded comparative data near-continuously within a secure computing environment where no human has access. Where conventional interim analysis offers a single snapshot at a pre-specified midpoint, often too early to be informative, sometimes too late to be useful, BRAKES surfaces evidence at the point in time where a decision can be made. Mass testing on simulated trial datasets demonstrates a false-negative rate approximately five times lower than conventional methods. 

We respond as a methodological commenter on the sub-questions below; on all others we defer to commenters with more direct operational standing.

II.A.1.a - The issue of ethical use of participants and trial resources

The pilot should weight earlier futility detection heavily, because its ethical benefit is one of the clearest and most measurable of the potential uses of AI. Once accumulating data indicate a trial is unlikely to meet its primary endpoint, continuing it:

• keeps participants exposed to an ineffective intervention, bearing its risks and burdens for no realistic prospect of benefit;
• holds a finite eligible population in the trial; in scarce-patient indications such as rare disease and much of oncology, this is a high opportunity cost 
• strains the premise of informed consent, under which participants contribute to an answerable scientific question rather than bearing risk for one already resolved;
• is poor stewardship of finite collective resources; investigator time, sites, funding and patient goodwill - spending them on a question already settled.

Earlier, well-founded detection addresses each of these and BRAKES does so without trade-off against integrity: no human accesses unblinded data, and the signal is non-binding, the decision remaining with the DMC and sponsor.

II.A.1.c - Use-case priority: a proposed futility detection-led pilot

Objective. We propose a two-stage pilot (in silico evaluation of BRAKES followed by BRAKES’ realtime monitoring of a live trial) built around BRAKES, prioritising earlier futility detection as its lead use case. Success (on-track) signal, safety monitoring and a Zero Knowledge Regulatory Primitive are additional contributions enabling real-time clinical trials.

The pilot's aim is to establish whether BRAKES can identify a failing trial reliably and earlier than single midpoint conventional interim analysis, while no human accesses the unblinded comparative data, and to quantify the benefit across 4 areas: patient exposure to ineffective treatment avoided, trial capacity freed for studies that may succeed, decision latency reduced against conventional interim timing, reduced Type II error rate vs conventional statistical approaches. As a monitoring layer overlaid on conventional frequentist trials, requiring minimal change to design, powering, or analysis, and incurring no alpha penalty, BRAKES scales across the existing trial base rather than a niche of bespoke designs.

Futility is the lead use case. An early futility signal from BRAKES gives the DMC real-time, well-founded evidence that a trial is unlikely to meet its primary endpoint, enabling an earlier no-go that shortens patient exposure to ineffective treatment and returns scarce trial capacity to studies that may succeed. In scarce-patient indications, where each enrolled participant carries the highest opportunity cost, this is both the clearest ethical benefit and the most measurable. 

The secondary success signal. A success (on-track) signal from BRAKES gives the DMC and if agreed, the Agency, evidence of a trial’s trajectory toward its primary endpoint at the earliest moment this trajectory passes a prespecified threshold in Phase 2. This would allow Phase 3 protocol design, site identification, SAP development to begin in parallel rather than sequentially, compressing the Phase 2-to-3 whitespace while the trial runs to its planned end. Further details may be found in Presentient’s ISTAND LOI. For sponsors it may also surface expedited-programme candidates (e.g. Breakthrough Therapy) sooner. For the Agency, the pilot could test whether a verified on-track signal lets FDA identify which running trials might warrant early Phase 3-design engagement before database lock, allocating scarce review resources by evidence rather than by waiting for Phase 2 readout. The BRAKES signal isn’t an interim analysis or claim that the drug is effective; it is not suitable to initiate early stopping for efficacy.

The tertiary safety layer. BRAKES can automate the generation of standard safety tables, listings, and figures that currently require weeks of manual preparation. Beyond standard tabulations, hierarchical classification of accumulating safety data into pre-defined groups can produce a composite score that enables real-time safety monitoring against pre-specified boundaries. This equips the agency or the DMC with more timely and more complete evidence on which to exercise their judgement. 

Zero Knowledge Regulatory Primitive. We propose a joint exploration of the cryptographic regulator-query interface described in II.A.4.b, which gives the Agency (or other authorised regulator) a means to extract, in real-time, specific information about safety (or other) data from the trial without revealing further details about the data itself. For example, a regulatory query can return a YES/NO on whether there have been e.g. more than 3 severe adverse events up to the present, without revealing the safety data itself or any other information about the trial. Another use case might involve a query requesting information about interaction of safety with efficacy, e.g. >3 SAEs and a positive estimated effect size up to the current time. This could be useful to protect trial integrity and to support wider engagement with real-time clinical trials among sponsors concerned about potential over-reaction to partial data. 

Stage 1: in-silico vetting (FDA-controlled gate). Before any live use, the Agency verifies the BRAKES’ key performance characteristics itself, using demoBRAKES against synthetic and real historical trials with known outcomes, truncated at decision points of its own choosing. Because the Agency sets the conditions and holds the ground truth, the result is one it can act on rather than take on trust. Live deployment proceeds only if BRAKES clears this gate.

Stage 2: live overlay (one to three trials). The instrument is overlaid on blinded randomised Phase 2 or 1b two arm frequentist design trials; either with endpoints observed relatively early after enrollment or time-to-event studies. A single futility trigger is prespecified in protocol before any data accrue, with no other change to powering, sample size, primary analysis or stopping rules. Monitoring is near-continuous within BRAKES’ secure enclave; the Agency views no patient-level data. If BRAKES meets the pre-specified conditions for likely futility (e.g. 99% likelihood of futility), then a signal is released to the DMC, and to the Agency as observer; any stop decision remains with the DMC and sponsor. This stage tests what simulation cannot: real-world integration, operability, and behaviour under live conditions. A similar prespecified approach can be used for BRAKES detecting success early, safety monitoring.

II.A.2.a — Criteria for selecting sponsors, trials, and technologies

For selecting technologies, it might be reasonable for the Agency to select on the expected return of a scarce pilot slot and consider a framing such as p · V / t: the probability of a conclusive result (p), the value of the pilot (V), and the time and FDA resources required to reach a result (t). On this framing, V is highest where the pilot delivers better trial decisions, faster trial decisions, patient protection, and a trial-efficiency gain that translates into measurable domestic competitiveness — plausibly recovering 10%+ of national trial capacity at scale. The following criteria, which apply equally to sponsor-developed and proprietary technologies, follow from that framing; the first is a gate, the rest are scored.

• Integrity preservation (gating). The technology must deliver its benefit without compromising blinding of trial team/participants, inference quality, and without removing human judgement from the decision point. A trustworthy AI pilot should not itself introduce integrity risk.
• Independent verifiability. Key performance characteristics should be verifiable by FDA itself in silico, against data with established ground truth, at decision points of the Agency's own choosing, before any live deployment. Establishing performance under the Agency's own conditions, rather than relying on vendor-supplied evidence, yields a result FDA can more confidently act on rather than one it must take on trust. As such, it reduces risks to pilot delivery. 
• Auditability and stability. The technology's decision behaviour should be specifiable in advance and stable enough to validate, ie, with versioning and change-control such that a single validation remains durable. This favours technologies whose performance can be characterised and confirmed, over those whose behaviour drifts or cannot be pinned down for review.
• Magnitude and breadth of benefit. Priority to technologies whose benefit, where it applies, is large in effect and broad in reach; materially changing trial decisions, accelerating them or scalably improving efficiency across many trials.
• Structural novelty. Priority should be to technologies that enable what the current paradigm cannot, rather than automating what it already does. 
• Regulatory maturity. Prior engagement with FDA or another regulator (e.g. an ISTAND submission) lowers the risk that a slot is consumed by a science project rather than a deployable system.

II.A.3.a  Effective partnerships

Analytical tools and data-pipeline providers occupy distinct operational layers; treating them as separable would allow multiple qualified providers at each. We would welcome discussion with the Agency on integrating into a suitable pilot our BRAKES real-time futility and success signals released to the Agency on pre-specified trigger conditions to support earlier Agency review of a trial’s data. This BRAKES module set is already built and validated, ready to support Agency activities immediately. It provides a scalable selection mechanism to trigger real-time review of trial data on pre-specified conditions. 

II.A.4.b - Infrastructure

Enabling Agency (and DMC) access to trial data in real-time without compromising blinding, patient privacy, or sponsor data custody is the core infrastructure question. Procedural safeguards such as training, role-based access and contractual restrictions are necessary but not sufficient to assure trial integrity. They depend on human self-restraint at every access point, and the operational bias they risk introducing across the data lifecycle cannot be adjusted for analytically. 

The properties worth specifying are: elimination of human access to unblinded comparative data across the data lifecycle; full auditability with logs available to sponsor and Agency; attestation of computational integrity as well as operational security capability to maintain these at scale. Evaluating against properties rather than architectural specifics leaves room for multiple valid implementations.

Presentient's institutional commitment to cybersecurity capability, with more than one-third of current staff having cybersecurity backgrounds and a commitment to maintain that proportion above one-quarter, is one expression of the operational investment such properties require.

An upcoming BRAKES module implements a cryptographic regulator-query interface; a structured framework of conditional information access rights between sponsor and Agency, which offers a joint project to explore together. This module enables sponsor data to remain in its custody while the Agency receives answers to a pre-specified set of potential queries in real-time. Our approach is based on a Zero Knowledge Proof system. Functions under development include continuous queries against the safety database with conditional queries against efficacy and futility probabilities released only when pre-specified circumstances arise. For example, when safety concern crosses a threshold within a trial where probability of success is predicted to be low. The latter gives the Agency decision-relevant information outside the sponsor's standard trigger schedule, but only under circumstances to which both parties have pre-agreed and that are cryptographically enforced. This could be useful to protect trial integrity and to support wider engagement with real-time clinical trials among sponsors concerned about potential over-reaction to partial data.

II.B.4.a - AI system performance

Validation of deterministic analytical tools such as BRAKES falls between pure analytical validation and the higher scrutiny appropriate for AI/ML models on dimensions such as potential model drift. A low-resource independent validation path is available in this space, consistent with the risk-based credibility framework in the Agency's January 2025 draft guidance.

The Agency's evaluation of performance need not depend solely on review of proprietary implementations or on validation evidence supplied by the sponsor. A direct approach is Agency-conducted performance verification: the Agency generates its own simulated trials, truncates them at decision points of its own choosing, runs the truncated data through a fixed configuration of BRAKES matching the sponsor’s deployment, and compares observed outputs against the ground truth the Agency holds. Presentient operates a benchmarking instance of BRAKES, demoBRAKES, that supports this evaluation process and is available on request.

Closing

Presentient welcomes the Agency's initiative. We see four opportunities for engagement: first, integration into the pilot's decision infrastructure of BRAKES futility and success signals, released on pre-specified trigger conditions, building on our ISTAND submission; second, automated safety reporting as a continuously available layer for DMC and Agency review; third, demoBRAKES as a reference deployment for Agency-conducted performance verification; fourth, joint exploration of the cryptographic regulator-query interface described in II.A.4.b. The first three are available now and require no new framework. The fourth is a forward-looking patent-pending architecture worth the Agency's consideration.

If you would like more information, please contact us on via our webform or email us at info@presentient.com.