Real-world data (RWD) is transforming clinical research, augmenting existing randomized controlled trial (RCT) data to de-risk studies and improve generalizability. With regulators setting clearer standards for their use, innovative designs are increasingly gaining traction.

However, drug developers must answer key questions to ensure feasibility and successful implementation. These include which use cases and under which conditions it is most appropriate to utilize RWD, which are the most effective RWD sources and statistical methods to utilize and how to create a more sustainable approach to data augmentation.

Innovation in the Use of Synthetic Patients

The development of stronger algorithms, faster computers and increased access to standardized RWD has resulted in significant innovation in the creation of synthetic patients in recent years. AI algorithms learn from both RCT and RWD to generate plausible synthetic patients in three main ways – emulation, simulation and synthesis.

Emulation applies AI to match study participants to historical or external data. It only uses existing and historical data, combining and reweighting results to generate robust comparators. Methods include matching and weighting approaches such as inverse probability of treatment weighting when using real-world data.

Simulation uses existing data and evidence to build disease and drug effect models with algorithms to interpolate or extrapolate. Such models are built with patient-level and aggregated data from multiple data sources. Inference and data integration methods include Bayesian disease progression and predictive modeling, trial simulations and causal inference and generative adversarial networks (GANs).

Synthesis uses generative AI algorithms to learn from large data sources, including publications, aggregated and patient-level data sources. Methods include large language models (LLMs), deep learning (DL) and machine learning (ML).

A systematic review of the ecology of synthetic data generator (SDG) methods published by Kaabachi et al in 2025¹ found around 40% of synthetic patients are generated by GAI and GAN methods. The larger proportion are generated by non-GAN methods including sequential trees, Bayesian network and statistical modeling.

The Value of in silico Trials or Cohorts

The impact of in silico trials or cohorts for drug developers can be classified into four buckets: accelerating time to evidence, predicting long-term benefit/risk and drug value, securing regulatory and HTA discussions and optimizing medical practice and value.

Accelerated time to evidence is the most understood value driver because of the capacity for parsimonious data collection and smarter trial designs. However, within this bucket, we should also consider how we can enrich trial design, ensure results will be generalizable to the real-world population and minimize the risk of a follow-up trial.

Long-term prediction of benefit/risk and a drug is about enabling the projection of short-term endpoints to long-term outcomes. For example, projecting public health impacts that may drive a decision of authorization or reimbursement in a given country from phase 3 trial results.

If we consider regulatory and HTA discussions, in silico disease cohorts or trial simulationscan optimize label defense. It can also help identify high value subgroups and allow the design of a pricing agreement where the price is directly derived from the performance of a drug on the market.

Finally, if an approved drug is not prescribed or misused, hence bringing no benefit to patients, then the medical and economic value of that drug is very low for health systems. Trials augmented with RWD can capture such pragmatic effects and consequently support decisions on the right endpoints, population and so on.

Each of these buckets have use cases which can be mapped versus acceptability and familiarity from health agencies. So, for example, if we look at accelerated time to evidence, specifically external control arms, emulation is already recognized in guidance, simulation has existing pilots leading to decisions and synthesis has exploratory use. However, the use of the younger AI-based synthetic approaches is currently widely exploratory apart from some use cases around the privacy of data and anonymization which can be granted by specific AI algorithms.

Case Studies to Demonstrate Impact

Impactful use of RWD modeling and patient synthesis in clinical trials comes in augmentation of existing trial data. The case studies below demonstrate how innovative approaches can deliver trial acceleration, increase the chance of regulatory success and increase medical value.

Case study 1: Historical control arm with adaptive design
There is still high unmet need in the prevention of perinatal HIV transmission, especially in low and middle-income countries. Yet it is hard to recruit pregnant women at risk in prevention trials. In addition, the standard of care (SoC) in HIV prevention can change very fast during the course of trial implementation. To overcome these challenges, there was a need for an efficient design which leveraged large historical data on SoC and treatment effects on prophylaxis. The chosen solution was to use an adaptive single arm trial design with a Bayesian meta-analysis and modeling of historical data to simulate controls, with early stopping rules based on the probability of reaching sufficient effect on transmission rates. This innovative approach divided the total sample size by five compared to a standard randomized controlled trial. The approach resulted in a successful trial and public health impact, changing WHO guidelines on prevention of HIV transmission. Similar approaches based on a platform trial design and integration of RWD on the standard of care are now considered for Africa.

Case study 2: Long-term benefit /risk projections for regulatory submission
A successful early prostate cancer Phase 3 trial showed superior progression-free survival (PFS) and histology for a new therapy versus active surveillance. However, the EMA questioned longer term safety and benefit/risk ratio. Modeling long term disease and outcomes progression from RCT and RW data was used to build a flexible external control to project benefit/risk outcomes such as erectile disfunction and incontinence over time. Cohort simulations were used to compare care strategies in terms of population-time with adverse events over more than five years and stratified analyses optimized label definition. The patient simulations were instrumental to gain market authorization, characterizing long-term benefit risks and to inform PASS and PAES requirements.

Case study 3: Optimizing switching time in heart failure to maximize drug effectiveness
New heart failure drugs beat the SoC in trials, yet poor relative effectiveness was observed in the real-world, hence resulting in poor medical benefit to patients and economic return to drug sponsors, especially in countries where performance-based pricing was implemented. There was a need to demonstrate when, and for whom, to switch to maximize medical value. A model patient pathway and disease progression was modelled from RCT and RWD. Various sequences of treatments were simulated and compared in terms of hospitalization and mortality rates, including differential switching strategies in various subgroups. Researchers found an early switch maximized medical value and care efficiency. The results were used to better design outcomes-based pricing and guide medical practices so to enhance effectiveness, reducing hospitalization rates by up to 30 to 40%.

A More Sustainable Approach to Patient Simulators

Case studies like those above are useful to showcase how in silico trials can demonstrate impact to patients and health systems, and delivering the full value of drugs in development. However, to truly build momentum in such use of RWD we need to shift away from ad hoc projects and try to identify patterns and build standard frameworks in the way we develop, use and maintain such patient simulators. Ad hoc approaches offer lower acceptability by health agencies and are often only valid for a very short window of time. This results in a huge investment and uncertain return.

As we shift to more sustainable solutions, we can build simulation platforms that collect real-world evidence from the best fit-for-purpose sources within a given disease and enable the development of a disease and care model which could be developed, maintained and validated in a mutualized way across companies and stakeholders. This would enable the simulation of virtual cohorts at the pre-competitive stage – not for a specific drug in development but for the whole research and public health communities.

The three pillars needed for this approach are always the same – modelling natural disease history, modelling drug and exposure effect and simulating patients’ disease and care. This work is already happening, e.g. via the Critical Path Institute engaging with KOLs, regulators such as EMA and FDA and many other stakeholders, so I believe we are going to see a lot more development in the next few years.

Is generative and agentic AI the Future of Synthetic Patients?

I want to conclude this article by looking further into the future and sharing a more provocative talking point – the use of AI to generate synthetic cohorts of patients only from the public aggregated information, just like ChatGPT does. When patient-level data from historical studies are available, GANs and deep learning have been tested to learn from such existing patient level data so to generate similar synthetic patients. It is computationally intensive, requires hard-to-access patient data for training and faces data privacy issues.

We recently explored whether ChatGPT-like LLMs could learn enough from Pubmed and generate cohorts of synthetic patients with their individual trajectories and outcomes. A prompt was fed into an LLM to generate synthetic data. Evaluation of performance was carried out by confronting results to real cohort data in Alzheimer's and Parkinson’s disease. These conditions were chosen because there is access to large cohorts, well documented patient table data and lots of published trials. This offers high potential for learning. We compared cohorts around three dimensions typically used in patient synthesis – fidelity, privacy and utility.

The results revealed an LLM without any patient-level data compared well to state-of-the-art DL approaches trained on rich data. While fidelity was often slightly under performing, privacy and utility performed well or very well.

A much larger evaluation exercise would be needed to see if the results could be replicated more widely. In the absence of patient-level data, we also need a validation and regulatory framework. However, it gives a flavor of what the future could look like in five or ten years thanks to wider data access the smarter algorithms.

Source

1. A scoping review of privacy and utility metrics in medical synthetic data. National Library of Medicie January 27, 2025. https://pmc.ncbi.nlm.nih.gov/articles/PMC11772694/