Synthetic data can reliably mirror real-world data (RWD) in chronic spontaneous urticaria (CSU), potentially enabling smaller clinical trial sample sizes without compromising statistical power, a recent study found.1 The findings, published in Clinical and Translational Allergy, highlight a significant challenge in CSU research—the ongoing difficulty of enrolling and retaining adequate patient numbers, especially among those with comorbidities, older age, or uncommon disease subtypes.
The results of the study show that synthetic data could maintain accuracy down to 25% of the original real-world data sample size. | Image credit: tippapatt – stock.adobe.com
The authors noted, “Robust data are essential for clinical and epidemiological research, yet in chronic spontaneous urticaria (CSU), certain patient groups, such as the elderly or comorbid patients, are often underrepresented. In clinical trials, strict inclusion and exclusion criteria frequently limit recruitment, making it difficult to achieve sufficient statistical power. Similarly, real-world observational studies may lack sufficient sample sizes for robust analysis.”
Using data from the Chronic Urticaria Registry (CURE), researchers extracted information on 4136 patients across 30 countries and 12 ethnicities, capturing a comprehensive set of demographic, laboratory, and patient-reported outcome variables. Synthetic datasets were generated using a Classification and Regression Trees (CART) algorithm, which allows the synthetic cohorts to “maintain the statistical properties and correlations of the original data without directly copying any individual records.”
This method preserves patient privacy while still capturing clinical and demographic diversity.
When systematically compared with real-world data, the synthetic datasets showed strong alignment across key measures. In terms of gender, RWD reported 72.4% female (n = 2994) vs 71.7% (n = 2965) in synthetic data, with no significant difference (P = .47). Age distributions were virtually identical: mean (SD) 44.2 (16.3) years in RWD compared with 44.3 (16.4) years in synthetic data (P = .85). Body mass index was similarly replicated (26.3 vs 26.1; P = .28).
Clinical characteristics were also successfully replicated. Daily wheals were reported by 28.6% of real patients compared with 28.8% of synthetic patients, while angioedema was absent in 24.3% of RWD patients, which was matched by 23.7% in synthetic data. Comorbidity burden was nearly identical, with mean comorbidities of 1.98 in RWD and 1.96 in synthetic datasets (P = .77). Atopic dermatitis prevalence was 4.8% in both groups, and allergic rhinitis occurred in 19.1% and 19.2% of patients, respectively (P = .98). Similarly, comorbidity burden, laboratory parameters such as IgE, and medication use showed no significant differences, reinforcing that synthetic datasets can reliably capture diverse clinical characteristics.
Further subgroup analyses, including patients aged ≥60 years, those with BMI ≥25 or ≥30, and both male and female cohorts, displayed no statistically significant differences in core characteristics and disease scores when comparing synthetic and real-world data. (all P > .10). Correlation analyses further validated synthetic fidelity. The strong negative correlation between UAS7 and UCT seen in RWD (P = -.73) was reproduced in synthetic data (P = -.72; P = .58).
The results of the study show that synthetic data could maintain accuracy down to 25% of the original RWD sample size. The authors explained, “enrolling just 38 patients in a clinical trial and applying GenDT allows us to generate a synthetic cohort of 150 patients. In other words, we can produce a synthetic patient population that is 4 times larger while maintaining high-quality data.” Previous technologies, such as Unlearn.AI, have only achieved a 33% reduction in control arm size, whereas this approach offers a 75% lower sample size for both control and treatment arms with equivalent statistical power.2
However, researchers caution that synthetic data performed best with continuous variables such as age, BMI, and patient-reported outcome scores, but categorical variables, including treatment type or symptom frequency, were more prone to errors when generated from smaller sample sizes.1 “Further research is necessary to establish and validate the standards of this method, allowing the scientific community to benefit from its advantages and safely use it in research settings,” the authors note.
These findings suggest that synthetic data generation, when vigorously validated, could ease barriers in CSU research, especially for understudied populations such as older adults, those with comorbidities, and patients with rare disease variants. By extending smaller cohorts into adequately powered synthetic populations, researchers may accelerate hypothesis testing, enhance subgroup analyses, and reduce the costs and burdens associated with recruitment.
References
1. Gutsche A, Salameh P, Jahandideh SS, et al. Can Synthetic Data Allow for Smaller Sample Sizes in Chronic Urticaria Research? Clin Transl Allergy. 2025;15(8):e70087. doi: 10.1002/clt2.70087
2. Yakubu A, Bogert J, Zhuang R, et al. Accelerating randomized clinical trials in Alzheimer’s Disease using generative machine learning model forecasts of progression. Alzheimers Dement. Published online January 9, 2025. doi:org/10.1002/alz.086486