Why Clinical Data Standards Matter In An AI-Driven World

By Bill Illis, Novartis

Hardly a day passes without a new headline claiming AI will transform every step of drug development. Many of these highly publicized claims come from pilot studies, small-scale implementations, or isolated point solutions. Most rely on self-reported data from technology providers and sponsors with significant financial stakes in the outcomes.

Rigorous evaluations, independent reports and solutions at scale are less common. They require more time and effort, but they offer a more balanced perspective and one that exposes gaps in how AI models are developed and trained and highlight the limits of real-world usage and impact. A recent example is "Dozens of AI disease-prediction models were trained on dubious data" (Nature, April 15, 2026).

Still, improvements are accumulating, and evidence suggests the drug development industry is on the verge of a historic transformation. While daily reports offer glimpses of what may be possible, reports of failed AI projects due to poor data quality and lack of AI-ready data are also appearing. This may be a good moment to pause and cut through some of the hype. What changes are required across the drug research and development data landscape to achieve sustained, meaningful progress with AI to bring new therapies to patients faster?

Failed Predictions

One of the most cited failed predictions in healthcare AI came from Geoffrey Hinton, a renowned AI researcher known as the Godfather of AI. In 2016, Hinton claimed that medical schools should stop training radiologists because AI would replace them within five years.

Ten years later, radiologists remain not only employed but indispensable. They leverage AI as a tool to enhance their diagnostic capabilities rather than being replaced by it. The number of radiology residency positions offered in the U.S. continues to grow each year. Hinton's prediction overlooked the nuances of clinical practice and human expertise, intuition, and interpretation of complex medical scenarios that no current model can replicate. The training data used to develop radiology models was too limited and too narrow in scope. As a result, these models outperform humans on standardized benchmarks but fail to match that performance in real-world conditions.

IBM Watson's oncology platform offers another cautionary tale. Promoted as a revolution in cancer care, it instead produced unsafe treatment recommendations and struggled with complex cases. The root causes? Poor data quality and a lack of standardized clinical data. IBM sold off Watson Health in 2022, a stark reminder of how weak data standards can derail even the most ambitious AI projects.

AI has also stumbled in clinical research. Patient recruitment algorithms designed to match participants with study protocols have failed due to inconsistent or incomplete EHR data. A lack of standardized data produced skewed recommendations and undermined recruitment efforts. A review of 50 reports on AI models for optimizing clinical trial recruitment and retention found that unreliable results and limited generalization stemmed largely from errors and incompleteness in the data sources used for training (JAMIA, September 11, 2024).

Ambiguous study designs pose similar challenges. When protocols are not clearly defined or structured, AI struggles to interpret them, leading to deviations and costly errors. Progress is being made through customized clinical trial retrieval-augmented generation (RAG) solutions, which improve on results from LLMs alone. A recently published study illustrates both the advances and the challenges, concluding that deeper alignment with emerging protocol data standards is the pathway to reducing errors while keeping "embedded human approval at key checkpoints to mitigate hallucinations and propagation of AI-generated errors" (Journal of Biomedical Informatics, July 2026).

The Need For Standardized Data

The recognition that accurate, reliable AI outputs require high-quality, representative, well-structured, standardized data inputs is not new. It is as simple as "garbage in, garbage out."

One recurring challenge, however, is that AI applications trained on lower-quality, localized data can still produce impressive results, but results that are prone to hallucinations and errors once the models are scaled for enterprise use. These remarkable-looking outcomes can lead us to believe the hard work of developing the AI model is done, when in fact it is only just starting. The deficiencies often become apparent only years later, when predicted outcomes fail to materialize at scale in real-world usage.

AI technology alone cannot overcome systemic flaws in data quality and standardization. Clinical data standards play a crucial role in addressing these issues. They establish methodologies and frameworks that make data uniform, consistent, and interpretable across institutions, studies, and systems. Without standardized data, AI systems face significant barriers in interpreting and generalizing information.

Individual organizations and AI developers can demonstrate compelling solutions today. But these efforts are often limited by the high cost of acquiring, cleaning, and transforming data into high-quality training data sets. This creates a low ceiling on what the industry can achieve at scale. Today's reported gains suggest some headroom still remains but without broad adoption of industry data standards, we will see only incremental progress and miss the order-of-magnitude improvements that are otherwise within reach.

Clinical Data Standards Today

The good news: Over the past 10 years, while AI has been emerging, clinical data standards have advanced in parallel. Standards development organizations are collaborating more closely, and adoption is maturing. Progress has been slow, but there are good reasons to be optimistic.

Since 2016, several milestones have reshaped the landscape:

• CDISC Standards became mandatory for drug submissions at the FDA and PMDA.
• HL7/FHIR (Fast Healthcare Interoperability Resources) emerged as the dominant exchange standard for healthcare data and EHR systems.
• SNOMED CT, LOINC, and ICD-11 gained broader adoption as clinical terminologies.
• The OMOP Common Data Model gained traction for observational research.

More recently, new and emerging advances point to a data foundation that can support meaningful improvements in clinical development:

• Digital, machine-readable clinical study protocol standards, enabled by the TransCelerate Digital Data Flow (DDF) initiative working in partnership with CDISC  (Full disclosure: I have served as the initiative lead since its inception.)
• The ICH M11 common trial protocol technical specification approved in November 2025 and the continued alignment with the CDISC Unified Study Definitions Model (USDM)
• The HL7 Vulcan Accelerator UDP (Using the Digital Protocol) initiative, which aims to connect digital protocols to EHR systems and enable FHIR API-based exchange
• Integration of HL7 standards with and CDISC standards which has demonstrated how FHIR-based APIs can streamline real-world data collection for AI-driven clinical research
• CDISC's 360i initiative, which is working to logically connect formerly disparate CDISC standards horizontally across end-to-end, design to submission, clinical study processes using new metadata terminology components known as biomedical concepts, analysis concepts, and derivation concepts

Four trends emerge from these advancements:

1. Documents to Data. The shift from paper and PDF-based exchange to API-driven, machine-readable standards is underway.
2. Silos to Ecosystems. Greater harmonization is emerging across clinical care, research, and regulatory domains.
3. Static to Dynamic. Standards are becoming metadata-driven, automated, and continuously updated.
4. Local to Global. International regulatory alignment via ICH across FDA, EMA, PMDA, and others is reducing submission duplication.

A RAND Corporation study reported that more than half of AI projects across all industries stall because of a lack of data readiness (RAND, August 13, 2024). This number is likely even higher in clinical development. These data standards trends represent the heavy lifting necessary to improve the overall clinical data infrastructure. That infrastructure, in turn, increases the success rate of deploying robust, accurate, reliable, and explainable AI outputs and drives progress toward meaningful real-world usage.

Continued progress on these fronts is essential to raise the ceiling of what is possible with AI. More work is needed to qualify AI-ready data in common, standardized ways. Standards Developing Organizations (SDO’s) need to urgently work more closely together to update methods both to apply AI to accelerate the development of standards, and to harmonize their approaches to produce standards which better deliver AI-ready data.  This includes metadata specifications that capture machine-readable attributes like data lineage, data quality, and data set completeness, as well as the classification of and adherence to privacy and security measures. Investments in data standards infrastructure should not be considered secondary to investments in AI itself.

What Else Is Needed? Industrywide Governance

These advances in clinical data standardization are necessary to create and improve the foundational data infrastructure. But they are not sufficient on their own to realize the full potential of AI.

Across sponsors, sites, CROs, technology providers, and regulators, adoption and implementation of standards remain fragmented and inconsistent. Every standards implementation initiative reaches an inflection point between rigorous adherence to the published standard and the flexibility required to deliver practical benefits quickly at the local level. Each customization seems reasonable in isolation. In aggregate, however, these customizations break industrywide interoperability and undermine the goal of reducing friction in how quality data is exchanged and used.

Standards derive their primary value from the network effect: The more organizations that adopt them, the cheaper they become to implement and the more valuable they become overall. Each local variation erodes this effect, raising costs and lowering benefits across the clinical research ecosystem. This is not theoretical but a pattern that has been repeated with every clinical data standardization effort to date.

Despite the increasing maturity of standards development, the clinical development industry lacks cohesive, comprehensive governance to decide on, commit to, and enforce standards adoption. Models from the broader healthcare space may apply here. In the U.S., the Office of the National Coordinator for Health IT (ONC) regulates data exchange in the interest of public health, patient access, and secondary use of data. In the EU, the European Health Data Space (EHDS) initiative plays a similar role. To date, however, neither has been meaningfully extended to clinical trial processes.

At the same time, drug regulators have been hesitant to impose data standards requirements that extend beyond the data submitted to them. They have also been slow to transition from document-based to standardized, fully structured data formats. For clinical studies, this leaves us primarily following 10-year-old CDISC submission requirements in two regions while anticipating emerging ICH structured document formats such as eCTD v4.0, which will be required in most regions by 2029, and the more recently approved ICH M11 protocol format, for which submission requirements have yet to be announced.

These mandatory standards do not go far enough to support AI-based improvements in clinical study start-up, execution, reporting, or regulatory review. Regulators concerned with improving clinical trials for public benefit need to extend their reach and accelerate development and adoption of mandatory standards. Consider the possibilities: AI-based methods for reviewing protocols, optimizing study designs, and benchmarking patient and site burden; automated site selection and feasibility analysis; improved patient matching and recruitment; and source data acquisition directly from EHR systems. AI models and agents in all these areas are being developed and tested today, and progress can be accelerated through faster industrywide adoption of appropriate clinical data standards and clear policies for permitted use of standardized data.

Doing Better

The FDA recently announced a new platform called HALO (Harmonized AI & Lifecycle Operations for Data), which consolidates more than 40 disparate applications, submission data sources, systems, and portals across all FDA centers (FDA, May 6, 2026). This approach may deliver short-term benefits to internal FDA review operations. However, it leaves the data inputs from external sources unchanged. Advanced algorithms cannot fully repair a broken data infrastructure.

Instead, regulators can catalyze real change by working with industry to define targets that mandate the consistent end-to-end adoption of clinical data standards. The opportunity extends beyond faster reviews and approvals. It is about realizing the often stated vision of putting patients first and enabling better-designed studies that start faster, enroll patients more quickly, eliminate friction at sites, avoid many amendments and deviations, deliver data faster and ultimately lead to accelerated and higher-quality decisions and more therapies reaching patients sooner.

Past expert predictions about AI failed because they focused too heavily on the technology producing the outputs and too little on the data feeding the models. As we anticipate the next generation of AI, clinical data standards matter more than ever.

Since the first mainstream generative AI chatbot was released in 2022, AI release cycles have far outpaced standards development, adoption, and governance. That gap is widening without coherent industrywide governance and intentional actions. A growing gap will only create more barriers to realizing AI's full potential and more failed predictions.

It is time to rebalance attention and investment. The rigorous, disciplined adoption of clinical data standards across the drug development landscape is the foundation on which every meaningful AI advance will be built.

About The Author:

Bill Illis has over 30 years’ experience in drug development R&D, focused on sustainable and scalable innovation in the acquisition and use of data. In addition to his current role in technology and scientific computing at Novartis, he is the workstream lead for the TransCelerate Digital Data Flow Initiative (protocol digitalization) and serves on the CDSIC board.