Reality bytes

Getting real about artificial data in drug development

Here’s the thing about drug development – we’re drowning in data but starving for the right information.

Regulatory constraints, privacy requirements and rare disease populations mean we are constantly working with incomplete pictures.

This is where synthetic data can help by providing artificial data sets that behave like the real thing. In order to generate and implement synthetic data effectively, you must start with understanding the source data and challenges in using real-world data directly for most applications. The most common implementation failure involves treating synthetic data generation as purely algorithmic.

Deploying sophisticated techniques, like generative adversarial networks and SMOTE algorithms, while overlooking critical upstream and downstream considerations such as data pre-processing, post-processing, governance and distribution, can result in technically proficient output, but, at the same time, surface data quality issues and data leakage risks.

The result is that patient studies get compromised even if they provide useful information.

Developing useful synthetic data starts with understanding the source data, identifying gaps and addressing them before deploying generation techniques. For example, if you take a clinical trial data set comprised largely of white males aged 35–65 and generate synthetic data from it, you might exacerbate bias inherent to the source data.

This is why it’s critical to begin by profiling the data to understand underlying distributions better, and find out what populations are underrepresented by the source data.

Data preparation demands equal attention to privacy. Traditional anonymisation approaches such as simple masking or identifier removal provide insufficient privacy protection. Differential privacy techniques offer more robust safeguards, introducing controlled statistical noise that preserves analytical utility while reducing the risk of individual-level traceability.

In the age of AI, human expertise remains non-negotiable. Clinical knowledge determines whether a synthetic data set serves its purpose. Population health studies need different data from precision medicine research, for example, even from identical sources.

Whether synthetic data should supplement existing data sets or serve as primary analytical inputs depends on specific research objectives, regulatory requirements and acceptable risk thresholds.

Success isn’t measured by technical sophistication: it’s measured by impact on drug development timelines and patient outcomes. KPIs should take this factor (called fit for purpose) into account. Does the synthetic data allow you to run proof-of-concept studies earlier? Identify safety signals faster? Stratify patients more effectively?

Synthetic augmentation shows particular promise in rare disease research where baseline event rates are low and patient populations number in hundreds.

Digital twin technologies push boundaries even further. Virtual patient populations now support drug interaction modelling, dosing strategies and adverse event prediction at scales impossible with real-world data alone.

Herein lies the paradox: the more powerful these capabilities become, the more critical proper process management becomes.

The organisations getting synthetic data right aren’t necessarily those with the fanciest algorithms. They’re the ones with the most disciplined approaches to data quality, bias detection and clinical validation.

As synthetic data moves from experimental to operational necessity, strong governance is a competitive advantage. It’s important to treat generation models like any analytical asset by documenting everything and tracking performance while maintaining audit trails.

The future isn’t about choosing between real and synthetic data. It’s about combining both intelligently to accelerate therapeutic innovation while protecting patient privacy and safety.

Built and governed properly, synthetic data transforms how quickly organisations move from hypothesis to evidence.

Anette Dalsgaard Jakobsen is Senior Industry Advisor, Global Health and Life Sciences Advisory at SAS and Sundaresh Sankaran, Senior Solutions Architect at SAS