From robotic surgery to predictive analytics, artificial intelligence (AI) has permeated most facets of healthcare – and its reach is only expanding. However, with this rapid expansion, navigating the slew of technical jargon can feel overwhelming. Whether you’re new to the field or looking to deepen your understanding, it’s essential to know the key terms that are shaping AI in healthcare today.
Keep reading for a list of common AI terms, and make sure to check back periodically as we add new ones.
Artificial Intelligence (AI)
AI refers to the capability of machines and computer systems to simulate human intelligence processes, including reasoning, problem solving, decision-making, and learning. Today, many AI systems can perform tasks that would have previously required human assistance.
AI in healthcare
AI in healthcare is a general term that broadly refers to the use of natural language processing (NLP), machine learning (ML), and other AI functions to optimize healthcare processes in both clinical and non-clinical settings, enhance experiences for providers and patients alike, improve health outcomes, and more – the possibilities are vast. AI in healthcare demands extreme care given the stakes (not to mention federal regulations). As such, organizations should seek out vendors with domain-specific models trained on rich terminology that can navigate clinical nuances and deliver highly specific, dependable results.
Clinical AI
While this term relates closely to “AI in healthcare” and is sometimes used interchangeably, clinical AI is more specific to the use of AI within the clinical environment. It typically refers to AI tools that enhance patient care and clinical decision-making, such as NLP for interpreting provider notes and accurately mapping diagnoses.
Natural language processing (NLP)
NLP is a broad branch of AI that focuses on enabling computers to understand, produce, and respond to human communication, including speech and text. Many people interact with NLP-driven applications daily, such as autocorrect, language translation services, and voice-activated digital assistants. One of the core capabilities of NLP is tranforming chaotic data into clean, structured information, thus enabling more sound healthcare decisions.
Large language model (LLM)
LLMs, which represent a subset of NLP, can interpret and generate text based on copious amounts of data that have been used to train them. LLM-powered apps, such as OpenAI’s ChatGPT, excel at producing human-like text, including articles, social media captions, and even creative works. Plus, when trained on rich clinical terminology and infused with expert-backed clinical AI, LLMs can dramatically increase the accuracy of medical coding.
Algorithm
An algorithm is like a recipe – a set of specific rules that is given to an AI machine to help it complete a task or solve an issue, resulting in a model that is capable of self-learning. Healthcare has hundreds of languages that require hundreds of algorithms to understand them – there is no “one-size-fits-all” clinical NLP algorithm. The key is training all these algorithms on clinical terminology that captures the many ways providers describe clinical concepts.
Machine learning (ML)
ML is a branch of AI that uses math, computer science, and coding to mimic how humans learn, creating self-learning machines that can tweak their actions based on new information. While AI refers to the general creation of machines with human-like cognitive abilities, ML refers to the use of algorithms and data sets to do so. For example, in the healthcare space, ML drives detection algorithms used in many medical devices such as automated electronic defibrillators (AEDs) and smart monitors.
Neural networks
Similar to how a human brain functions, neural networks are used in ML to organize data and make decisions. One of the most iconic examples of a neural network is Google’s search algorithm. The research into the application of neural networks to healthcare is still ongoing, but so far, many potential use cases have been put forward. Some of these include; detecting and diagnosing anomalies in medical images, identifying cohorts for clinical studies from aggregated patient data, predicting patient no-shows, drafting responses to patient inquiries, and much more.
Deep learning
Deep learning is a subset of ML that also mimics how the human brain functions via neural networks; in this case how it organizes and processes information to make decisions. Unlike typical algorithm-based ML models, deep learning models can learn from unstructured data without explicit rules, gaining accuracy over time. While these models require more powerful (and costly) resources, they can execute extraordinarily complex tasks, like powering surgical robots, making them indispensable in healthcare.
Generative AI
This is a type of AI that relies on deep learning models to create original content in response to user prompts. Some health IT experts view generative AI as key in helping to solve healthcare’s significant unstructured data problem and increasing connectivity among multiple data sets; others envision generative AI revolutionizing application development by simplifying the interaction between healthcare professionals and data.
GUIDE
Leveraging AI to meet client needs:
A guide for driving efficiency, maximizing cost effectiveness, and boosting user satisfaction
Ambient AI
Ambient AI refers to technology that operates in the background, analyzing encounters and delivering insights with minimal user involvement. In the realm of healthcare, Ambient Clinical Intelligence (ACI), also known as Ambient Clinical Documentation, is an advanced technology that leverages AI and NLP to capture and analyze physician encounters in real time without the use of a keyboard. Beyond personalizing the patient health journey and vastly improving clinician workflows, ambient AI models that are built on foundations of comprehensive terminology have the potential to transform sub-fields in healthcare, like telemedicine and remote care.
Knowledge graph
A knowledge graph, also known as a semantic network, is a data structure that represents real-life entities, such as objects or concepts, and the relationships between them. Knowledge graphs are often confused with or used in the same context as ontologies, as the two can resemble one another visually. In clinical AI, leveraging semantically structured knowledge graphs is key for the most accurate, detailed, and actionable results.
Ontology
Ontologies serve as structured frameworks that categorize and organize the entities within knowledge graphs, enabling AI systems to comprehend hierarchical relationships and form connections between different concepts. When it comes to clinical AI specifically, it’s crucial to build upon deep ontologies to address complex, clinical nuances and deliver accurate, actionable results.
Generative Pre-trained Transformer (GPT)
GPT refers to a group of LLMs that can analyze, summarize, and generate text in plain language. More specifically, as the acronym suggests, it’s a generative AI technology that has been previously trained to transform prompts, or inputs, into unique outputs. The most famous use case is arguably ChatGPT, which mimics natural conversation.
Retrieval-augmented generation (RAG)
RAG is a technique that permits LLMs to access external information, thus increasing the accuracy and reliability of results. In medical coding, RAG can enable LLMs to reference relevant information retrieved from comprehensive terminologies and normalization application programming interfaces (APIs). By minimizing the occurrence of fake or inaccurate codes, RAG empowers vendors to use smaller LLMs to build lower-cost and faster-running solutions.
Optical character recognition (OCR)
OCR, sometimes called text recognition, is a technology that converts images of text, such as PDF documents, into machine-readable formats through automated data extraction. The user can then edit the extracted text, reducing manual and redundant work. OCR methods are especially valuable in healthcare where free text narratives, like clinician notes, typically reside in PDFs or images.
Application Programming Interface (API)
An AI API allows developers to add AI capabilities to existing applications, enabling organizations to seamlessly integrate new features into their programs. This is particularly useful in the healthcare sector where the flexible deplyoment of both vendor-hosted and self-hosted technology is key.
Predictive analytics
Predictive analytics is one of the core types of data analytics, leveraging historical data, patterns, and trends to predict future events or behaviors. Predictive analytics, which relies partly on AI, poses many healthcare benefits, including more substantiated decision-making, optimized efficiency, and streamlined disease management.
Semantic interoperability
Different machines can exchange data, but this will not amount to much unless they fully understand what is being shared. Semantic interoperability, or the ability of computers and machines to communicate in a way that is understandable and standardized, is critical in healthcare. Developers must train LLMs on robust clinical terminology mapped accurately to standard code sets to maintain semantic interoperability and optimize AI for real insights.
Autonomous coding
As its name implies, autonomous coding is AI-powered coding that can function independently. In healthcare, it’s emerging as a formidable tool for alleviating the immense administrative burden of manual coding. However, complex clinical language and inconsistent data can pose challenges for autonomous coding in the healthcare space; the solution is relying on strong clinical terminology, combined with NLP and data normalization tools.
True positive
A true positive in ML is when a machine correctly predicts a positive outcome. For example, in a situation where the positive result represents a certain medical condition, a true positive would be when the machine correctly identifies a patient as having that condition. True positives are key in assessing a model’s performance and level of accuracy.
Precision
Precision refers to the measure of positive predictions that are correct. In AI-powered medical coding, it’s the number of terms a machine identified as ‘problems’ that are in fact, problems. A machine with high precision is one that correctly identifies problem terms with few falsely identified problem terms.
Recall
In the context of automated medical coding, recall refers to the measure of correctly identified problem terms from all the actual problem terms in the data set. A machine with high recall can comprehensively identify problem terms in the data set with few missed problem terms. In the healthcare space especially, machines should ideally have both high precision and high recall values.
F1 score
A machine’s F1 score indicates its balance of precision and recall, or the measure of both correct and incorrect identification of problem terms in the data set. High F1 indicates both high precision and high recall, meaning problem terms in the dataset are thoroughly identified, and that identification is correct; few problem terms are missed.
