Marika Niihori
Mihaela van der SchaarAdvances in machine learning have opened new opportunities to rethink how clinical trials are designed, executed, and interpreted. The van der Schaar Lab has been at the forefront of this transformation for the past decade.
Our research has focused on four interconnected pillars.
Causal Effect Inference
Unlocking individualized patient insights by combining traditional trial data with real-world evidence to prove what truly works.
Adaptive Clinical Trials using AI
Replacing rigid protocols with intelligent, real-time designs that learn as they go to find the right treatments faster.
Digital Twins
Using generative AI to simulate “what-if” scenarios and virtual patients, reducing the need for costly and risky physical testing.
AI for Pharmacology
Moving beyond “one-size-fits-all” dosing by merging biological expertise with AI to predict exactly how a drug will behave in any individual
Causal Effect Inference
Causal Effect Inference expands the role of clinical evidence generation by both empowering randomized controlled trials (RCTs) and, in some settings, reducing the need for them altogether. While RCTs remain the gold standard for establishing efficacy, causal inference methods allow researchers to extract substantially more value from trial data by moving beyond average treatment effects toward individualized and subgroup-specific understanding, enabling robust estimation of effects across heterogeneous and vulnerable populations, informing optimal dosing, and guiding treatment choice when standards of care are unclear. At the same time, modern causal inference frameworks make it possible to integrate trial data with observational and real-world evidence, extending insights beyond the limited duration and scope of traditional trials to capture long-term outcomes, safety signals, and evolving patient trajectories. Importantly, when randomization is infeasible, unethical, or prohibitively expensive — for example in rare diseases, rapidly changing treatment landscapes, or post-approval settings — causal inference can provide a principled alternative by reconstructing counterfactual outcomes from high-quality observational data. In such cases, causal methods can partially or, in carefully validated scenarios, effectively replace the need for new RCTs, enabling faster and more scalable evidence generation while maintaining scientific rigor. Together, these capabilities transform clinical research from a rigid, trial-centric paradigm into a continuous learning framework that combines experimental and real-world data to deliver precise, actionable, and patient-centered insights.
Digital Twins
Digital Twins represent a new generation of generative, AI-driven computational models that can simulate the biological, clinical, and operational trajectories of individual patients, populations, or entire trials, enabling researchers to explore “what-if” scenarios before decisions are made in the real world. Unlike traditional mechanistic models, modern digital twins combine domain knowledge with machine learning to build adaptive, data-driven representations that evolve continuously as new information becomes available, integrating diverse data sources such as longitudinal clinical records, genomics, trial data, and real-world evidence. Their generative power lies in the ability to create realistic counterfactual scenarios — predicting safety and efficacy under different dosing strategies, treatment combinations, patient characteristics, or trial designs — and to simulate outcomes that would otherwise require costly or ethically challenging experimentation. In clinical trials, this fundamentally changes how evidence can be generated: digital twins enable virtual or partially synthetic control arms, improve safety by forecasting adverse trajectories before they occur, support adaptive trial designs through real-time prediction of patient responses, and allow trials to test multiple hypotheses or comparators with greater efficiency and diversity. They also extend the value of trials beyond completion by supporting personalized treatment decisions and continuous learning in real-world deployment. By turning trials from static experiments into dynamic, continuously simulated systems, digital twins have the potential to accelerate answers, improve probability of success, reduce cost and patient burden, and ultimately reshape clinical development into a more predictive, patient-centered, and scalable process.
Adaptive Clinical Trials using AI
Adaptive clinical trials represent a shift from fixed, one-shot experimental designs toward continuously learning studies that can modify decisions as evidence accumulates. Machine learning plays a central role in enabling this transition by providing principled methods to decide when to continue, refine, or stop trial strategies, how to identify promising subpopulations, and how to allocate patients more efficiently under uncertainty. Our work has shown that adaptive experimentation can be formulated as a sequential decision-making problem in which trial designers must balance the value of gaining additional information against the cost of remaining committed to potentially failing strategies, leading to new frameworks such as optimal commitment policies that determine when trials should be adapted or redesigned in real time. At the same time, ML-driven adaptive designs enable flexible discovery of patient populations with treatment benefit, moving beyond traditional enrichment approaches by allowing dynamic subgroup identification while maintaining statistical rigor and error control — a critical requirement for regulatory acceptance. Our research further demonstrates how adaptive designs can improve sample efficiency through techniques such as synthetic controls and adaptive recruitment, allowing trials to learn faster from limited data and to focus resources where treatment effects are most likely to emerge. Together, these advances transform clinical trials from static protocols into intelligent, data-driven systems that continuously update their hypotheses, improve probability of success, and accelerate the delivery of effective therapies to patients.
AI for Pharmacology
AI is transforming pharmacology by enabling a shift from population-level, static models toward adaptive, data-driven understanding of how drugs behave across individuals, biological contexts, and time. Building on more than a decade of work at the van der Schaar Lab, our research integrates machine learning with pharmacometrics and systems pharmacology to improve prediction of pharmacokinetics and pharmacodynamics, support precision dosing, and enable individualized treatment strategies that go beyond traditional one-size-fits-all approaches. A central theme of our work is combining the strengths of mechanistic pharmacological modeling with the flexibility of modern AI — for example by integrating expert-designed ODE models with neural dynamical systems to better capture disease progression and drug response even in small-sample settings. We also develop methods that leverage multiple models and heterogeneous datasets through intelligent model combination, allowing robust predictions under covariate shift and across diverse patient populations. More recently, our work extends toward data-driven discovery of dynamical systems and interpretable equation discovery for pharmacological processes, enabling AI not only to predict outcomes but also to generate new mechanistic insight into drug dynamics. Together, these advances position AI as a core engine for next-generation pharmacology — accelerating drug development, enabling safer and more effective dosing, and creating a continuous learning loop between pharmacological theory, clinical data, and real-world patient outcomes. Together, these approaches aim to make clinical trials more reliable, efficient, and patient-centric. Readers interested in learning more are invited to visit the dedicated pages for each topic on our website.
New White Paper: Clinical Trials as Continuously Learning Systems
This white paper proposes a continuous, multi-agent machine learning framework built on shared causal digital twins to proactively derisk clinical development across the entire trial lifecycle.
The ideas presented here build on a 2022 clinician discussion on next-generation trial design hosted by the van der Schaar Lab. See that conversation here:
