A Future for MedTech: Navigating AI in Pharmaceuticals

As we stand on the brink of a technological revolution, artificial intelligence (AI) is reshaping various sectors, and pharmaceuticals are no exception. From drug discovery to clinical trials, AI is set to transform the landscape by analyzing vast datasets, speeding up processes, lowering costs, and enabling personalized medicine.

The Role of AI in Pharmaceuticals

AI’s transformative capabilities are being employed in several crucial areas, including identifying promising drug candidates, predicting protein structures, optimizing supply chains, and automating regulatory tasks. The efficiency and precision of drug development are notable benefits, as AI aids in finding new targets, designing molecules more swiftly, and improving patient recruitment for trials. However, this progress isn’t without challenges, such as ensuring data quality and transparency.

Regulatory Oversight: A Balancing Act

As AI technologies gain traction, the question arises: how do they align with the regulatory frameworks governing pharmaceuticals? The European Medicines Agency (EMA) has taken a groundbreaking step by producing a draft guidance focusing on the application of AI in drug development and manufacturing. This initiative, known as Annex 22, aims to provide a clear framework for AI governance in Good Manufacturing Practices (GMP).

Scope of Annex 22: Clear Boundaries

Annex 22 outlines strict boundaries, applying only to static, deterministic AI/ML models used in crucial GMP processes. In this context, only static machine learning models (which yield consistent outputs for the same inputs) and critical applications with stringent controls are permitted. Notably, dynamic models, generative AI, and large language models (LLMs) are excluded from this regulatory framework, deemed acceptable solely for non-critical GMP tasks under Human-in-the-Loop (HITL) oversight.

Cross-Functional Collaboration: A Necessity

One of the cornerstone mandates of Annex 22 is the emphasis on cross-functional accountability. It requires collaboration among subject matter experts, data scientists, Quality Assurance teams, IT departments, and vendors. All stages—from algorithm selection to operation—demand clear documentation and robust governance frameworks within pharmaceutical organizations utilizing AI.

Defined Intentions: The Importance of Clarity

In the pharmaceutical landscape, acceptance testing is vital for equipment validation. Annex 22 insists upon a comprehensive characterization of the input sample space prior to acceptance testing, ensuring that even rare variations are identified. Each pharmaceutical entity must meticulously define intended uses and establish clear HITL responsibilities to uphold accountability.

Acceptance Criteria: Statistical Rigor

When measuring the success of AI models, Annex 22 stipulates that companies must establish transparent acceptance criteria before testing begins. Specific metrics, like accuracy and sensitivity, need to be clearly defined by experts. Moreover, the performance of the AI model must exceed that of the traditional processes it intends to replace, ensuring a measurable improvement in operational efficiency.

Data Integrity: Statistical and Procedural Rigor

To accurately assess an AI model, the accompanying test data must represent the entire input space, including rare edge cases. The Annex emphasizes that the dataset should be expansive enough to achieve statistical significance while maintaining high accuracy in labeling. Intriguingly, it also prohibits the use of generative AI-created test data to ensure integrity.

Independence in Testing: Avoiding Bias

To mitigate bias in AI development, Annex 22 enforces strong separation of duties. Controls are in place, such as no shared use of training and test data and the need for access-controlled, audited repositories. Developers must not have access to test data unless stringent oversight is in place, reinforcing strict data segregation.

Rigorous Test Execution

The execution phase of AI performance testing is critical. The Annex demands a well-defined test plan, complete with metrics, test scripts, and data references. Any deviations encountered during testing must follow established GMP protocols, while all test artifacts, including audit trails, must be meticulously documented.

Ensuring Explainability

In the realm of pharmaceuticals, transparency is paramount. Each AI model must utilize explainable AI techniques, which assign importance scores to input features. Tools like SHAP (SHapley Additive ExPlanations) and LIME (Local Interpretable Model-agnostic Explanations) are employed to elucidate model behavior, ensuring compliance with transparency standards.

Confidence Measures Against Uncertainty

To cultivate confidence in AI predictions, models must incorporate mechanisms to log confidence scores and implement thresholds that avoid unreliable outputs. When confidence levels fall below acceptable thresholds, models should indicate “undecided” to prevent inappropriate automated decisions from being made.

Life-Cycle Governance: A Rigorous Approach

The operational phase of AI within pharmaceuticals mandates strict life-cycle governance as outlined in Annex 22. Documentation and assessments are required for any modifications to AI models, with stringent configuration controls established to detect unauthorized changes throughout the model’s lifespan.

As AI steadily integrates into the pharmaceutical sector, the guidelines set forth by Annex 22 aim to clarify expectations for regulatory compliance, emphasizing the balance between innovation and safety. With public comment now closed, the finalized version is anticipated later in 2026, setting a precedent that may shape the future of medical technology in profound ways.