AI and Propensity Matching in Rare-Disease Clinical Trials

Pharmaceutical biostatisticians are overcoming the challenges of rare-disease clinical trials by using generative artificial intelligence and propensity score matching to create external synthetic control arms. Because rare diseases affect small populations, recruiting enough patients for a traditional placebo group is often impossible or unethical. This article provides an overview of how biostatisticians collect historical patient data, utilize generative AI to simulate realistic control groups, and apply propensity matching to remove statistical bias, ultimately accelerating the approval of life-saving therapies.

The Challenge of Rare-Disease Trials

In standard clinical trials, patients are randomly assigned to receive either the experimental treatment or a placebo. In rare-disease trials, the patient pool is extremely limited. Furthermore, many rare diseases are severe or fatal, making it unethical to withhold potential treatments from a control group. To bypass this issue, researchers construct external synthetic control arms using data from patients who were previously treated under standard care.

Gathering Real-World Data

The foundation of a synthetic control arm is Real-World Data (RWD) combined with historical clinical trial data. Biostatisticians gather information from electronic health records, disease registries, and past studies. This data includes baseline patient demographics, genetic markers, disease progression timelines, and treatment histories. However, raw historical data is often incomplete, inconsistent, or subject to strict privacy regulations.

Generating Patients with Generative AI

To resolve data scarcity and privacy issues, biostatisticians use generative AI models, such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs).

Biostatisticians train these AI models on the existing historical datasets. The AI learns the complex, underlying mathematical relationships between different patient variables—such as how a specific biomarker correlates with symptom progression over time. Once trained, the generative AI creates entirely new, synthetic patient profiles. These digital profiles mirror the statistical properties of real rare-disease patients without duplicating any actual person's data, effectively solving both sample size limitations and patient privacy concerns.

Balancing Groups with Propensity Score Matching

Because a synthetic control arm is not created through traditional randomization, there is a high risk of selection bias. If the synthetic control group is inherently sicker or healthier than the real patients receiving the new drug, the trial results will be invalid. Biostatisticians correct this using Propensity Score Matching (PSM).

A propensity score is the statistical probability that a patient would be assigned to the treatment group based on their baseline characteristics (such as age, sex, and disease severity). Biostatisticians calculate this score for both the real patients in the active treatment arm and the AI-generated profiles in the synthetic pool.

Using matching algorithms, the biostatistician pairs each real patient in the trial with a synthetic patient who has an identical or nearly identical propensity score. Unmatched synthetic profiles are discarded. This process ensures the external control arm is statistically balanced with the treatment arm, successfully mimicking the conditions of a randomized controlled trial.

Validation and Regulatory Submission

Once the generative AI and propensity matching processes are complete, biostatisticians perform sensitivity analyses to test the robustness of the synthetic control arm. They verify that the baseline characteristics between the real and synthetic groups are indistinguishable. This validated dataset is then used to measure the true efficacy of the experimental drug, providing regulatory bodies like the FDA and EMA with the statistical evidence required for drug approval.