Glossario

‍Annex (EU GMP): technical annexes to the European Good Manufacturing Practice (EU GMP) that address specific aspects of pharmaceutical manufacturing. Each Annex provides detailed guidelines on particular topics (e.g., sterile manufacturing, computerized systems, AI).

Information Asymmetry: a situation in which one party involved in a transaction possesses less information than the other, creating a gap that requires trust mechanisms to enable the exchange.

Audit trail: a chronological and immutable record of all actions performed on a system, documenting who did what, when, and why. Fundamental for traceability in a GMP environment.

Change control: a formal procedure that manages any modification to a validated system, assessing its impact, documenting the changes, and requiring approvals before implementation.

Computer System Validation (CSV): a systematic process demonstrating that a computerized system performs as intended, in a consistent and reproducible manner, through documented testing.

Controllability: a principle stating that any modification to a system must follow formal change control procedures, including an impact assessment.

Dataset: a set of data used to train, validate, or test a machine learning model. It is typically divided into a training dataset (for training), a validation dataset (for optimization), and a test dataset (for final evaluation).

EU GMP / Good Manufacturing Practice: a European regulatory framework establishing the requirements to ensure that medicinal products are manufactured and controlled according to quality standards appropriate for their intended use.

‍

FMEA (Failure Mode and Effects Analysis): a risk assessment methodology that identifies potential failure modes of a process or product, evaluating their effects and defining preventive actions.

Generative AI: artificial intelligence systems capable of generating new content (text, images, code) based on patterns learned from training data. This includes technologies like ChatGPT, DALL-E, and similar tools. They are typically probabilistic and non-deterministic.

Human-in-the-loop (HITL): an approach where a human operator retains an active role in the decision-making process, supervising and approving actions suggested by an AI system before they are executed.

Large Language Models (LLM): artificial intelligence models trained on massive amounts of text to understand and generate natural language. They are inherently probabilistic and non-deterministic: given the same inputs, they may produce different outputs due to their stochastic nature. Examples: GPT-4, Claude, Gemini.

Deterministic Models: models that, given the same inputs, always produce the same outputs in a predictable and reproducible manner.

Dynamic Models: AI models that continue to learn and update their parameters based on new data after deployment. Unlike static models, their behavior can change over time, which poses specific challenges for validation in regulated environments (as the system state is not fixed).

Probabilistic Models: models that, even when given the same inputs, can produce different outputs due to the stochastic component in their operation. They do not guarantee deterministic reproducibility.

Static Models: AI models with parameters that remain fixed and immutable during operational use. Once trained and validated, they do not modify their behavior without a formal update intervention.

Parameters (AI): internal numerical values of a machine learning model that determine how the model processes inputs and generates outputs. They are established during the training phase.

Predictability: a principle stating that a system must generate identical results given identical inputs.

Quality Management System (QMS): a structured system of processes, procedures, and responsibilities that ensures the maintenance of quality standards across all of an organization's activities.

R&D (Research & Development): pre-commercial research and development activities where new products or processes are designed, tested, and optimized before their introduction into production.

‍Retrain offline with revalidation: a strategy for updating an AI model by retraining it outside the production environment, followed by complete revalidation before operational redeployment.

Risk assessment: a systematic process of identifying, analyzing, and evaluating risks associated with an activity, process, or system, defining actions to mitigate them.

SOP (Standard Operating Procedure): a standard operating procedure that describes in detail how to perform a specific activity in a consistent manner and in compliance with quality requirements.

Traceability: a principle stating that every action performed by a system must be fully documented through an audit trail.

Validability: a principle stating that a system's performance must be verifiable and reproducible through documented testing.

‍Validation: a documented demonstration that a process, system, or method produces results compliant with predefined acceptance criteria in a consistent and reproducible manner.