Mihaela van der Schaar
Marika NiihoriProf Mihaela van der Schaar’s latest white paper sets out a new framework for transforming how clinical trials are derisked. The full paper and abstract are available below.
Clinical development is entering a regime in which the dominant source of risk is no longer only the absence of data, but also the inability to reason coherently across heterogeneous, evolving, and only partially observed evidence streams, where uncertainty necessarily depends on priors derived from external knowledge. Biological uncertainty, operational variability, regulatory interpretation, and competitive dynamics interact non-linearly, yet current derisking approaches treat them as separable problems addressed at different moments in the development lifecycle. As a result, organisations continue to make large, irreversible commitments based on fragmented and imperfect reasoning, discovering critical failure modes only after substantial capital, time, and patient participation have already been invested.
The opportunity created by advances in machine learning is therefore not merely to improve individual trial components, but to fundamentally redesign how clinical development decisions are reasoned about and made. The central thesis of this white paper is that derisking should instead be supported by a continuous multi-agent learning system operating across the entire development system. Such a system must integrate heterogeneous evidence streams from discovery, prior trials, real-world data, operational signals, regulatory constraints, and the evolving therapeutic landscape, reason about how these jointly shape the conditions under which trials fail, succeed, or make a true impact on patients, and based on this knowledge, inform how clinical trials are conducted and derisked.
To enable this shift, I propose a coordinated multi-agent machine learning framework operating over shared causal digital twins of patients, trial populations, and development environments.
Specialised agents representing biological, statistical, operational, regulatory, and competitive perspectives will interact through structured and auditable communication and negotiation, enabling structured disagreement, counterfactual exploration, and early identification of fragile assumptions across the trial lifecycle. By transforming derisking from a sequence of isolated predictive tasks into a continuous learning and reasoning process, the framework enables proactive decision-making to shape trial success conditions, earlier identification of avoidable failures, and sustained alignment between evidence generation and real-world clinical relevance. It is important to note that both the specialised agents and the digital twins over which they make decisions are designed to use interpretable, auditable machine-learning models that make assumptions, uncertainties, and decision logic explicit, ensuring that the resulting reasoning remains scientifically transparent and suitable for regulatory evaluation.
This white paper aims to move clinical development beyond protocol-centric optimisation toward continuously learning clinical trial systems capable of adapting not only trial execution, but the broader development strategy itself as scientific, operational, and competitive environments evolve. This is why the framework proposed here cannot be designed, evaluated, or deployed within a single disciplinary perspective. Therefore, this paper is written for a broad interdisciplinary audience including clinical trialists, statisticians, clinicians, regulators, operational leaders, and machine learning researchers; it aims not to present a field-specific methodological proposal, but to articulate a shared system-level vision and technical foundation that can enable these communities to come together to collectively develop continuously learning clinical development infrastructures.
Join us on February 16th, 4-6pm UK time at the Revolutionary Healthcare session, where we’ll explore how agentic ML, digital twins, and multi-agent reasoning can be translated into real clinical development infrastructure and de-risking. Sign up here.
The full white paper is available below:
