AI & Health Glossary

In healthcare, algorithms power everything from simple clinical decision rules to complex AI systems. A traditional algorithm might follow explicit rules — for example, “if the patient’s heart rate exceeds 100 bpm and blood pressure drops below 90/60, alert the care team.” AI algorithms go further by learning patterns from data rather than relying solely on predefined rules, enabling them to identify subtle relationships that would be difficult for humans to specify manually.

When healthcare professionals encounter the term “algorithm” in the context of AI, it typically refers to the mathematical method used to train a model or make predictions. Different algorithms are suited to different tasks — some excel at classifying images, others at predicting numerical values, and others at finding hidden patterns in data. Understanding that algorithms are simply structured problem-solving methods, not mysterious black boxes, is the first step toward evaluating AI tools critically.

Artificial intelligence in healthcare encompasses a wide range of technologies, from rule-based expert systems that have existed for decades to modern machine learning and deep learning approaches. The common thread is that these systems process information and produce outputs — diagnoses, predictions, recommendations — that would otherwise require human cognitive effort.

It is important to understand that “AI” is an umbrella term, not a single technology. When evaluating a product marketed as “AI-powered,” the relevant question is what specific type of AI it uses, what data it was trained on, and what evidence supports its claims. A simple rule-based alert system and a sophisticated deep learning model are both forms of AI, but they differ enormously in their capabilities, limitations, and appropriate use cases.

Bias in healthcare AI is a critical concern because it can directly impact patient outcomes and exacerbate existing health disparities. Bias can enter AI systems through multiple pathways: training data that underrepresents certain populations, historical healthcare data that reflects past discrimination, feature selection that inadvertently proxies for race or socioeconomic status, or evaluation methods that fail to test performance across demographic subgroups.

A well-known example is a widely used commercial algorithm that was found to systematically underestimate the health needs of Black patients because it used healthcare spending as a proxy for illness severity — spending that was already shaped by unequal access to care. Identifying and mitigating bias requires diverse training data, rigorous subgroup testing, ongoing performance monitoring, and inclusive development teams. Healthcare organizations evaluating AI tools should always ask vendors how bias has been assessed and what demographic groups the tool has been validated on.

Clinical decision support systems range from simple rule-based alerts (such as drug interaction warnings) to sophisticated AI-powered tools that analyze complex patient data to suggest diagnoses or treatment options. Modern AI-enhanced CDS can process far more variables simultaneously than traditional rule-based systems, potentially identifying patterns and risks that simpler approaches would miss.

Effective CDS integrates seamlessly into clinical workflows, presenting relevant information at the right time without creating alert fatigue. One of the longstanding challenges in CDS design is balancing sensitivity (catching real issues) with specificity (avoiding false alarms that train clinicians to ignore alerts). AI-driven approaches have the potential to improve this balance by tailoring alerts to individual patient contexts rather than applying blanket rules.

Computer vision is one of the most mature and impactful applications of AI in healthcare. It powers tools that analyze medical images — X-rays, CT scans, MRIs, pathology slides, retinal photographs, and dermatological images — to detect abnormalities, measure structures, and classify findings. Many FDA-cleared AI medical devices rely on computer vision, including systems for detecting diabetic retinopathy, identifying lung nodules, and flagging potential fractures.

The technology works by training deep learning models on large datasets of labeled medical images. The model learns to recognize visual patterns associated with specific conditions, much as a radiologist develops pattern recognition through years of training — but with the ability to process images at much greater speed and scale. Computer vision does not replace the specialist’s interpretation but adds an additional layer of detection that can improve sensitivity and help prioritize urgent cases.

Deep learning has driven many of the most impressive AI breakthroughs in healthcare. Its strength lies in its ability to automatically discover relevant features in raw data — rather than requiring humans to manually specify what the model should look for. This makes it particularly powerful for complex data types like medical images, genomic sequences, and unstructured clinical text, where the important patterns are difficult to define explicitly.

In healthcare applications, deep learning models can match or exceed specialist performance on specific tasks, such as classifying skin lesions, detecting arrhythmias from ECG data, or identifying cancerous cells in pathology slides. However, deep learning models typically require very large training datasets, significant computational resources, and can be difficult to interpret — meaning it may not be clear why the model made a particular prediction. This “black box” characteristic is an active area of research, as clinical adoption often depends on the ability to explain AI reasoning.

Electronic health records are both a critical data source for healthcare AI and a primary integration point for AI tools. EHR systems like Epic, Cerner (now Oracle Health), and MEDITECH contain structured data (lab results, vital signs, medications, billing codes) as well as unstructured data (clinical notes, radiology reports, discharge summaries) that AI models can analyze to generate predictions, identify risks, and support clinical decisions.

Many AI applications are designed to work within the EHR environment, appearing as embedded alerts, dashboards, or workflow enhancements so that clinicians can benefit from AI insights without leaving their primary work system. The quality and completeness of EHR data directly impacts the performance of AI models trained on it, making data governance and standardization important prerequisites for successful AI implementation in healthcare organizations.

The FDA has cleared hundreds of AI-enabled medical devices across specialties including radiology, cardiology, ophthalmology, and pathology. The most common regulatory pathways are the 510(k) process (demonstrating substantial equivalence to an existing device) and the De Novo pathway (for novel, lower-risk devices without a direct predicate). The FDA has also introduced a framework for AI/ML-based software that can adapt and improve over time, recognizing that these tools differ from traditional static medical devices.

Understanding FDA clearance is important for healthcare professionals evaluating AI tools. Clearance indicates that the FDA has reviewed evidence of the device’s safety and effectiveness, but it does not guarantee that the tool will perform equally well in every clinical setting or patient population. Healthcare organizations should look beyond the regulatory status to examine the specific clinical evidence, the populations studied, and how the tool’s performance has been validated in real-world settings similar to their own.

Generative AI has rapidly become one of the most visible and accessible forms of AI in healthcare, largely through tools like ChatGPT, Claude, and Gemini. In clinical settings, generative AI is being used for drafting clinical documentation, creating patient education materials, summarizing medical literature, generating referral letters, and assisting with administrative tasks. Ambient AI scribes that listen to patient encounters and produce structured notes are a prominent example of generative AI in healthcare.

While generative AI is powerful, it carries unique risks in healthcare contexts. These models can produce fluent, confident-sounding text that is factually incorrect — a phenomenon called “hallucination.” They may not reflect the most current medical evidence, and their outputs should always be reviewed by a qualified professional before being applied to patient care. Privacy is another concern: entering patient data into consumer-facing generative AI tools may violate HIPAA unless appropriate enterprise agreements are in place.

HIPAA is a foundational consideration for any AI deployment in U.S. healthcare. The law’s Privacy Rule and Security Rule govern how protected health information (PHI) can be collected, stored, processed, and shared. When healthcare organizations use AI tools that access, process, or store PHI, they must ensure compliance with these rules — including executing Business Associate Agreements (BAAs) with AI vendors, implementing appropriate technical safeguards, and maintaining audit trails.

The rise of cloud-based and generative AI has created new HIPAA compliance challenges. Consumer-facing AI tools like ChatGPT are not HIPAA-compliant by default, meaning clinicians should never enter identifiable patient information into them unless their organization has a specific enterprise agreement with the provider that includes HIPAA provisions. Healthcare organizations should develop clear, practical guidelines for staff about which AI tools are approved for use with patient data and provide training on how to use AI responsibly within regulatory boundaries.

Large language models like GPT-4, Claude, Gemini, and LLaMA are the technology behind the generative AI tools that have rapidly entered healthcare workflows. These models learn statistical patterns in language from billions of text documents and can then generate coherent, contextually relevant text in response to prompts. In healthcare, LLMs are being used for clinical documentation, literature synthesis, patient communication, medical education, and administrative tasks.

LLMs have notable limitations that are important for healthcare professionals to understand. They can generate plausible-sounding but incorrect information (hallucinations), they may reflect biases present in their training data, and they do not truly “understand” medicine the way a clinician does — they predict likely text sequences based on patterns. For these reasons, LLM outputs in clinical contexts should always be verified by a qualified professional. Despite these limitations, when used as assistive tools with appropriate oversight, LLMs can meaningfully improve clinical efficiency and access to information.

Machine learning is the driving force behind most modern healthcare AI applications. Rather than following hand-written rules, ML models are trained on datasets of examples — such as thousands of labeled medical images or millions of patient records — and learn to identify the patterns that distinguish one outcome from another. This data-driven approach allows ML systems to capture complex relationships that would be impractical to specify manually.

There are several main types of machine learning. Supervised learning uses labeled examples (e.g., images already marked as “normal” or “abnormal”) to train models to classify new cases. Unsupervised learning finds hidden patterns in data without predefined labels, useful for discovering patient subgroups or anomalies. Reinforcement learning trains models through trial and error to optimize sequences of decisions. Each approach has different strengths and is suited to different healthcare applications, from diagnostic imaging to treatment optimization to operational forecasting.

Natural language processing is particularly valuable in healthcare because so much clinical information exists as unstructured text — physician notes, radiology reports, pathology findings, discharge summaries, and research articles. NLP tools can extract structured data from these narratives, enabling analysis at scale that would be impossible through manual review. For example, NLP can identify all patients in a health system whose clinical notes mention a specific symptom pattern, even when different clinicians describe it in different ways.

Healthcare applications of NLP include ambient clinical documentation (converting spoken patient encounters into structured notes), automated medical coding, clinical trial matching, adverse event detection from medical records, and literature surveillance. The advent of large language models has dramatically expanded NLP capabilities, enabling more nuanced understanding of medical text and more natural interaction between clinicians and AI systems.

Neural networks are the architectural foundation of deep learning and many modern AI systems used in healthcare. They consist of an input layer (which receives data), one or more hidden layers (which process the data through mathematical transformations), and an output layer (which produces the result — such as a classification or prediction). During training, the network adjusts the strength of connections between nodes to minimize errors, gradually learning to map inputs to correct outputs.

In healthcare, neural networks power applications ranging from image classification (detecting tumors in radiology scans) to sequence prediction (forecasting patient deterioration from time-series vital signs) to language generation (producing clinical documentation). The term “deep” in deep learning refers to neural networks with many hidden layers, which can capture increasingly abstract and complex patterns. While the internal workings of neural networks can be difficult to interpret, research into explainable AI is making these systems more transparent and trustworthy for clinical use.

AI is a key enabler of precision medicine because it can analyze the vast, multi-dimensional datasets required to identify which treatments work best for which patients. Genomic data alone can contain millions of data points per patient, and when combined with clinical history, lab results, imaging, and lifestyle factors, the complexity far exceeds what traditional statistical methods can handle effectively. Machine learning models can discover patterns across these variables to predict treatment response, identify disease subtypes, and guide therapeutic selection.

In oncology, AI-powered precision medicine is already in clinical use, helping match patients with targeted therapies based on their tumor’s genetic profile. Pharmacogenomics applications use AI to predict how individual patients will metabolize specific drugs, reducing adverse reactions and optimizing dosing. As datasets grow and AI models become more sophisticated, precision medicine is expected to expand into more specialties and conditions, moving healthcare toward a model where treatment plans are truly individualized rather than based on population averages.

Prompt engineering has become an essential skill for healthcare professionals using generative AI tools. The quality of an AI’s output is directly influenced by how the input prompt is structured. In healthcare contexts, effective prompts specify the clinical scenario, desired format, evidence standards, target audience, and any constraints — such as avoiding information beyond a certain date or focusing on specific patient populations.

Key principles for healthcare prompt engineering include providing sufficient clinical context, specifying the desired level of detail and format, requesting citations or evidence levels, asking the model to flag uncertainty rather than guess, and never including identifiable patient information in prompts to non-compliant tools. Building a library of tested prompt templates for recurring tasks — such as patient education, literature review, or differential diagnosis brainstorming — can help clinicians work efficiently while maintaining quality and safety standards.

Reinforcement learning differs from other machine learning approaches in that the model learns through interaction with an environment rather than from a static dataset of labeled examples. The agent takes actions, observes outcomes, and adjusts its strategy to maximize cumulative rewards over time. This trial-and-error approach is particularly well-suited to sequential decision-making problems where the best action depends on the current state and future consequences.

In healthcare, reinforcement learning is being explored for treatment optimization — for example, determining the best sequence and dosing of medications for sepsis management, chemotherapy regimens, or insulin dosing for diabetes. It is also applied to adaptive clinical trial design, robotic surgery planning, and resource allocation in hospitals. Because reinforcement learning involves exploration (trying different actions to learn), its application in healthcare requires careful safety considerations, and much of the current work uses simulated environments or retrospective data rather than real-time patient interaction.

Supervised learning is the most common form of machine learning in healthcare AI. The process involves providing the model with a dataset where each example has both input features (such as pixel values of a medical image, or patient vital signs) and a corresponding label (such as “malignant” or “benign,” or “will be readmitted within 30 days”). The model learns the relationship between inputs and labels, then applies that learned mapping to new, unlabeled data.

The quality of supervised learning depends heavily on the quality and quantity of labeled training data. In healthcare, creating these labels often requires expert annotation — for example, radiologists marking regions of interest on thousands of images, or clinicians reviewing charts to confirm diagnoses. This annotation process is time-consuming and expensive but essential for building accurate models. Common supervised learning tasks in healthcare include classification (is this lesion benign or malignant?), regression (what is this patient’s predicted length of stay?), and risk scoring (what is the probability of a specific adverse event?).

Training data is the foundation upon which all machine learning models are built. In healthcare, training data might include medical images with expert annotations, electronic health records with outcome labels, genomic sequences paired with disease diagnoses, or clinical notes tagged with relevant codes. The principle is straightforward: a model can only learn patterns that are present in its training data, so the quality, size, diversity, and representativeness of this data directly determine the model’s capabilities and limitations.

Several challenges surround training data in healthcare. Patient privacy regulations restrict how health data can be collected and shared, which can limit the size and diversity of available datasets. Imbalanced data — where certain conditions, demographics, or outcomes are underrepresented — can lead to biased models that perform poorly for minority populations. Data quality issues such as missing values, inconsistent coding, and documentation variability add further complexity. Techniques like federated learning (training models across multiple institutions without sharing raw data) and synthetic data generation are emerging as solutions to some of these challenges.

Validation is a critical step in determining whether a healthcare AI tool is ready for clinical deployment. Internal validation tests the model on a held-out portion of the original dataset, while external validation tests it on entirely independent data — ideally from different institutions, patient populations, and time periods. External validation is far more informative because it reveals whether the model’s performance holds up outside the specific conditions in which it was developed.

For healthcare organizations evaluating AI tools, understanding the validation process is essential. Key questions include: Was the model validated on a population similar to yours? Were results broken down by demographic subgroups to check for bias? Was the validation performed prospectively (on new data as it arrived) or only retrospectively? Were the performance metrics clinically meaningful — not just statistically significant? Robust validation does not guarantee perfect performance, but it provides the evidence needed to make informed decisions about clinical deployment and to set realistic expectations for how the tool will perform in practice.