Molecular classification now divides patients with the same hematologic diagnosis into subgroups that may each need a different treatment. In rare diseases and narrowly defined populations, recruiting both investigational and control arms can take years, and standards of care may change before a study finishes.
Generative artificial intelligence (AI) can create synthetic patient cohorts. Trained on existing clinical trials and real-world datasets, these models generate patient-level data that reproduce the statistical characteristics of real populations, including laboratory findings, molecular profiles, and survival outcomes. Such cohorts could supplement conventional control groups, reduce recruitment requirements, and allow more patients to receive investigational treatments. Their ability to support reliable comparisons of treatment efficacy remains unresolved, scientifically and for regulators.
OncoDaily Blood spoke with Yan Leyfman, Medical Correspondent at OncLive, Co-Founder and Executive Director of MedNews Week, and freelance medical journalist, about synthetic control arms in hematology, their methodological limits, and the path from AI-generated evidence to clinical practice.
Can synthetic control arms reproduce the effects of randomization?
“Randomization remains the gold standard for minimizing both measured and unmeasured confounding, and synthetic control arms cannot fully replicate that protection. However, I believe the more important question is not whether artificial intelligence can replace randomization, but whether it can help generate reliable evidence in settings where traditional randomized trials are impractical, ethically challenging, or limited by small patient populations.
In hematologic oncology, particularly in relapsed or refractory disease, we frequently encounter substantial molecular heterogeneity, rapidly evolving therapeutic landscapes, and patients who have exhausted conventional treatment options. Generative AI may facilitate the integration of longitudinal clinical data, identification of comparable external control populations, and modeling of disease trajectories.
Nevertheless, these approaches must be grounded in rigorous causal inference, transparent methodology, and independent validation. A statistically convincing synthetic patient is not necessarily a clinically valid counterfactual. Our goal should be to complement randomization where appropriate, not to replace scientific rigor with algorithmic sophistication.”
How can synthetic trials remain relevant when standards of care change?
“One of the greatest challenges in hematologic oncology is that therapeutic innovation often outpaces conventional evidence generation. Consider multiple myeloma, where CAR-T therapies, bispecific antibodies, and novel combination regimens are rapidly transforming treatment sequencing and patient outcomes. A historical control cohort from even a few years ago may no longer accurately reflect contemporary standards of care. I envision a transition from static historical controls toward dynamic, continuously updated evidence frameworks capable of incorporating evolving treatment patterns, molecular risk stratification, supportive care advances, and real-world outcomes.
Generative AI could facilitate data harmonization and identify changes in patient populations or treatment practices, while prespecified statistical methods would be needed to address temporal confounding. Importantly, updating a model does not automatically eliminate historical bias. Any modifications must be transparent, version-controlled, prospectively defined where possible, and independently validated. The future of synthetic trials lies not simply in generating historical comparisons, but in developing adaptive evidence infrastructures that remain clinically relevant without sacrificing reproducibility or methodological integrity.”
Rare subgroups: greatest need, hardest data
Synthetic control arms may help most in hematologic malignancies where recruitment is hardest: molecularly defined AML subgroups, rare lymphomas, and multiply relapsed disease. These populations are small, heterogeneous, and spread across institutions.
They are also the populations for which reliable synthetic data is hardest to generate. Models trained on limited datasets can’t necessarily reproduce uncommon molecular characteristics, treatment histories, or resistance patterns. Pooling multiple trials and registries may improve representativeness, but comparison quality still depends on the completeness and clinical relevance of the data.
Could synthetic controls add value in rare molecular subgroups?
“Absolutely. I believe rare molecular subgroups represent one of the most compelling opportunities for AI-assisted external control arms. As hematologic malignancies become increasingly classified by genomic and molecular characteristics, traditional randomized trials may become prohibitively difficult to conduct in narrowly defined populations. This is particularly relevant in rare AML molecular subsets, uncommon T-cell lymphomas, and patients with specific resistance mechanisms following cellular or targeted therapies.
AI could facilitate the identification and harmonization of clinically comparable patients across institutions, improve the utilization of fragmented real-world datasets, and support more efficient evidence generation. However, rare populations also introduce substantial risks of model overfitting, selection bias, and the generation of biologically implausible synthetic patients. For this reason, I would prioritize validated patient-level data, clinically informed matching, and rigorous sensitivity analyses rather than relying exclusively on fully AI-generated cohorts.
Beyond methodological efficiency, this approach has important implications for global health equity. Patients with rare malignancies, particularly those treated outside major academic centers or in resource-limited settings, are frequently underrepresented in clinical research. The true promise of AI is not simply to accelerate clinical trials, but to make meaningful evidence generation possible for patients who have historically been excluded from it.”
What safeguards are needed to prevent bias or manipulation of synthetic trial results?
“I believe responsible implementation requires a comprehensive framework centered on transparency, verification, reproducibility, and accountability. One of the central challenges I have emphasized in my work on responsible AI is the verification paradox: as AI-generated outputs become increasingly sophisticated and convincing, our ability to independently establish their validity may not advance at the same pace.
This becomes particularly consequential in clinical trials, where seemingly minor decisions regarding patient selection, endpoint definitions, missing data, or statistical adjustment can substantially influence observed treatment effects. Safeguards should include prospective protocol registration, prespecified eligibility and endpoint criteria, transparent data provenance, independent statistical oversight, and rigorous assessments of confounding and selection bias. Sensitivity analyses, negative controls where appropriate, external validation, and reproducibility across independent datasets are equally important.
We must also evaluate model performance across demographic, geographic, socioeconomic, and molecular subgroups to avoid perpetuating existing healthcare disparities. Importantly, AI systems should not independently determine favorable comparators or selectively report outcomes without predefined methodological constraints and human oversight. The more consequential the clinical decision, the greater the burden of verification. Scientific credibility must come from independently reproducible evidence, not from the apparent sophistication of the algorithm.”
Comparability: realistic records vs valid comparisons
A synthetic cohort may reproduce the survival distribution of a completed trial and still miss factors that affect outcomes in a new population. Treatment selection, supportive care, molecular testing, and eligibility criteria all affect how comparable patients from different datasets are. Regulatory acceptance will depend on evidence that synthetic controls support valid treatment comparisons.
What would it take for AI-generated control arms to influence clinical practice?
“For AI-generated control arms to meaningfully influence clinical practice, we must move beyond technological feasibility and demonstrate clinical and causal validity. A critical first step would be systematic benchmarking against completed randomized hematology trials. By reconstructing external control cohorts using prespecified methods and comparing their estimated treatment effects against independently established randomized results, we could begin identifying the clinical circumstances in which these approaches are reliable, as well as where they fail.
Importantly, retrospective agreement alone would not establish prospective validity. Multicenter validation, transparent reporting standards, independent replication, and prospective evaluation within appropriately designed clinical studies would be essential. I also envision opportunities for hybrid trial designs in which external controls augment randomized comparator groups, potentially reducing recruitment burdens while preserving important protections against bias.
Regulatory engagement with agencies such as the FDA, clear standards for data quality and model governance, and clinically meaningful endpoints would be necessary before broader adoption. Ultimately, the success of synthetic control arms should not be measured by their ability to produce statistically significant findings or accelerate regulatory submissions, but by whether they generate trustworthy evidence that improves therapeutic decision-making and patient outcomes. Our objective should never be to replace the randomized clinical trial simply because we can. It should be to leverage AI responsibly so that fewer patients, particularly those with rare or refractory hematologic malignancies, are left without meaningful evidence to guide their care.”
Perspective
Generative AI can reproduce characteristics of hematologic malignancies and simulate outcomes from completed trials. Its ability to support reliable comparisons of new treatments is unproven so far. Establishing which settings allow credible synthetic controls and which still require randomized comparison will take hematologists, clinical trialists, statisticians, AI researchers, and regulators working together.
