Artificial Intelligence Terms

Artificial Intelligence: Machine-based systems that can perform tasks that typically require reasoning, learning, perception, and/or the ability to make decisions.

Machine Learning: A subset of AI in which the system is trained to learn from patterns in data to assist in decision making; it can also improve its performance over time.

Deep Learning: A type of machine learning in which the AI is trained through multiple layers of neural networks to learn hierarchical feature representations from large datasets.

Algorithm vs Artificial Intelligence: An algorithm is a set of predefined, fixed instructions that tells a computer a set of steps to follow in order to complete a task, whereas artificial intelligence refers to systems that can perform tasks that require learning and adaption. As such, all AI systems use algorithms, but not all algorithms are AI.

Generative vs Predictive AI: Generative AI creates new content, such as text, code, or images, in line with the data it was trained on, while predictive AI is designed for estimating the likelihood of a specific outcome based on existing data. 

Hallucination: A response generated by AI that produces confident-sounding but false, nonsensical, or fabricated information, such as fake facts, nonexistent sources, or incorrect data. The problem occurs because AI is predicting likely word patterns rather than understanding the truth of the data, leading to outputs that seem plausible but are not grounded in reality. 

Agentic AI: AI that has the ability to independently choose the next step needed to complete a given task, including the capacity to search for information and determine which tool is best for accomplishing its goal. 

Neural Network: Computational model structured to mimic the human brain with interconnected nodes arranged in layers for processing information and identifying data patterns.

Convolutional Neural Network: A deep learning architecture that analyzes data by detecting patterns—such as edges, shapes, and textures—in one small portion of the data at a time and combining the patterns to support classification or prediction tasks. Convolutional neural networks are primarily used with images and other visual data.

Large Language Model (LLM): An AI model trained on large textual datasets to understand natural language processing and to be able to generate responses (example: ChatGPT is a commonly used LLM). 

Natural Language Processing: A branch of AI for training computers to understand and analyze human language. Natural language generation uses fundamental principles of natural language processing to generate text.

Transfer Learning: A strategy for adapting a model trained on one task to be reused for a different, related task to reduce the training time needed.

Foundational Model: A large-scale, general-purpose AI model designed to learn general knowledge from extensive, heterogeneous data for multiple purposes. These models serve as a starting point that can be adapted or customized for more specific needs without retraining the model from scratch.

Multimodal AI: An approach to AI models for processing and integrating multiple different data types—such as text, images, audio, video, genomics, etc.—to learn the relationships between them for improved prediction performance.

Human in the Loop (HITL): Refers to AI systems that require human oversight or participation as part of its operation to ensure greater accuracy or end-user trust.

Digital Twin: A virtual representation of a physical, real-world entity—a patient, organ, device, or healthcare system—that mimics its structure, context, behavior, and evolution and is dynamically updated with real-world data. Digital twins are used to simulate and predict how the entity may respond to different conditions and interventions.

Synthetic Data: Artificially generated datasets designed to resemble real-world data without replicating original records, which results in reduced privacy and data-sharing risks.

Federated Learning: A collaborative approach to training machine learning models in which the data come from multiple decentralized sources with only model updates rather than raw patient data shared to one centralized server for greater privacy protection. 

Ground Truth: The "gold standard" or reality/reference that the AI is being compared against or trained on.

Overfitting: A modeling error that occurs when an AI learns the "noise" or specific quirks of its training dataset too well, rather than the underlying medical patterns. An overfitted model may perform perfectly on the data it was built on but fail completely when applied to a different dataset.

Inference: The process of a trained AI model making a prediction on new, "unseen" patient data. While "training" happens in the lab, "inference" is what happens when a clinician runs a patient's scan through the software to get a result.

Retrieval-Augmented Generation (RAG): A technique used to "anchor" a large language model to a specific, trusted set of documents. This significantly reduces hallucinations by forcing the AI to cite specific sources before generating an answer.

Fine-Tuning: The process of taking a pre-trained foundational model and providing it with a smaller, specialized dataset to make it an expert in a specific niche without having to build a model from scratch.

Transformer: The specific type of neural network architecture (the "T" in GPT) that uses a mechanism called "attention" to weigh the significance of different parts of input data. It is the breakthrough that allowed AI to understand context in long medical notes rather than just individual words.

Drift: The phenomenon where an AI model’s performance degrades over time because the real-world data an AI model encounters over time starts to look different from the data it was originally trained on.

Autonomous AI: AI-enabled products that have the ability to perform tasks, operate independently, and make decisions without human intervention. The level of autonomy can vary based on the product. In clinical practice, most AI tools today are assistive, meaning that they flag, suggest, or highlight findings for a clinician to review. An autonomous AI, on the other hand, operates without that checkpoint.

For additional terms, visit the FDA’s Digital Health and Artificial Intelligence Glossary