A new study shows AI-generated synthetic patients can replicate cancer trial results, potentially shrinking the need for human control groups.

Why recruit hundreds of critically ill patients just to give them a placebo? In oncology, finding enough matching patients for clinical trials is slow, expensive, and ethically challenging. A new study suggests we can skip the human control group entirely by using AI-generated clones.

This is not just a statistical trick. It challenges the sacred cow of clinical medicine: the randomized controlled trial. If synthetic cohorts can reliably mimic real-world responses, insisting on physical control groups for terminal diseases like leukemia may soon look both inefficient and unethical.

Building the digital cohort

Researchers built their synthetic cohort using data from 1,377 acute myeloid leukemia (AML) patients. These historical records came from previous clinical trials and a real-world registry. They used this data to fine-tune a tabular foundation model, which then generated synthetic patients with matching clinical, genetic, and survival features.

The researchers matched these synthetic patients to the original SORAML trial’s active treatment group using Cox risk scores. The results were nearly identical to the real-world trial.

• The original trial showed a treatment hazard ratio of 0.64 (95% CI 0.47-0.87, p=0.004).
• The synthetic control arm produced a hazard ratio of 0.66 (95% CI 0.48-0.90, p=0.009).
• Median event-free survival rates between the two analyses were virtually indistinguishable.

This level of precision suggests that historical data, when processed through the right architecture, contains all the signal needed to simulate standard-of-care outcomes.

The limits of simulation

The catch is the data pipeline. This method only works because the researchers had access to high-quality, deeply characterized data from over a thousand real AML patients. For rare diseases or entirely novel therapies with no historical precedent, synthetic controls will struggle. If the training data is biased, the synthetic patients will inherit those exact blind spots.

Regulators will also need convincing. While the FDA has shown openness to synthetic controls in rare diseases, adopting them for mainstream oncology trials requires a massive cultural shift.

Why this matters

For oncology, this shifts the bottleneck of drug development. Instead of spending years recruiting patients who might end up on standard therapies that fail them, researchers can run “in silico” trials. This speeds up the validation of life-saving drugs. It also means more patients who enroll in trials can actually receive the active, experimental treatment rather than being randomized to a placebo.

Ultimately, this study proves that the past can predict the present with startling accuracy. If we can trust the math, we can stop experimenting on people when we already know how they will respond.

Read the full preprint on medRxiv.