AI in Healthcare Glossary

A defined set of step-by-step rules a computer follows to solve a problem or complete a task.

In practice: A triage algorithm ranks emergency patients by severity based on vital signs entered at intake.

The broad field of building computer systems that perform tasks normally requiring human intelligence, such as reasoning, perception, or language.

In practice: An AI system flags chest X-rays that may show pneumonia for faster radiologist review.

A summary of how well a model separates classes across thresholds; AUC ranges from 0.5 (chance) to 1.0 (perfect).

In practice: An AUC of 0.94 indicates strong discrimination between diseased and healthy scans.

Systematic error that causes a model to perform unequally across groups or situations.

In practice: A model trained mostly on one population may underperform for underrepresented patients.

A model whose internal decision process is not readily interpretable by humans.

In practice: Clinicians may hesitate to trust a black-box risk score without knowing what drives it.

Software that provides clinicians with knowledge or patient-specific guidance at the point of care.

In practice: A CDS alert warns of a dangerous drug combination as an order is placed.

AI that interprets visual information such as images and video.

In practice: Computer vision measures tumor size automatically across serial CT scans.

Protecting patient information from unauthorized access or disclosure.

In practice: De-identifying records before model training helps preserve data privacy.

A branch of machine learning using multi-layered neural networks to learn complex patterns from large datasets.

In practice: Deep learning detects diabetic retinopathy directly from retinal photographs.

Digitizing tissue slides so images can be viewed, shared, and analyzed by AI.

In practice: Digital pathology lets an AI pre-screen slides and prioritize suspicious regions.

Methods that make a model's reasoning understandable to humans.

In practice: A heat map highlighting the lung region that drove a pneumonia prediction aids explainability.

A single metric balancing precision and recall, useful when classes are imbalanced.

In practice: The F1 score helps evaluate a rare-disease detector where positives are uncommon.

An individual measurable property or input variable the model uses to make predictions.

In practice: Age, blood pressure, and creatinine level are features in a kidney-risk model.

Training a shared model across institutions without moving raw data off-site.

In practice: Hospitals jointly train a model on their own patient data without exchanging records.

Further training a pre-existing model on a narrower dataset to specialize it for a task or domain.

In practice: Fine-tuning an LLM on oncology literature sharpens its cancer-related answers.

A large model trained on broad data that can be adapted to many downstream tasks.

In practice: A medical foundation model can be fine-tuned for summarizing notes, coding, or Q&A.

AI that creates new content — text, images, audio, or structures — rather than only classifying existing data.

In practice: Generative AI produces synthetic medical images to augment scarce training datasets.

When a generative model produces confident output that is factually wrong or unsupported.

In practice: An LLM inventing a non-existent drug interaction is a hallucination — a key clinical safety risk.

A design where a qualified person reviews, confirms, or overrides AI outputs before action.

In practice: AI suggests a diagnosis, but the physician makes the final call — human-in-the-loop.

The known correct output attached to a training example.

In practice: Each scan in the dataset carries a label indicating whether a tumor is present.

A foundation model trained on vast text to understand and generate human-like language.

In practice: An LLM drafts a plain-language discharge summary from a clinician's shorthand notes.

A subset of AI where systems learn patterns from data rather than being explicitly programmed with fixed rules.

In practice: An ML model learns from thousands of past cases to estimate a patient's risk of readmission.

Degradation of model performance over time as real-world data shifts away from training conditions.

In practice: A sepsis model may drift after a hospital changes its lab equipment or workflows.

AI that understands, interprets, and generates human language.

In practice: NLP extracts smoking status from thousands of free-text clinical notes.

A model loosely inspired by the brain, made of interconnected units (neurons) organized in layers that transform input into output.

In practice: A neural network converts ECG waveforms into a rhythm classification.

When a model memorizes training data too closely and fails to generalize to new cases.

In practice: An overfitted model scores perfectly in development but poorly on real patients.

Of all cases the model flagged positive, the fraction that were truly positive.

In practice: High precision means most patients the model flags as high-risk really are high-risk.

Tailoring prevention and treatment to individual differences in genes, environment, and lifestyle.

In practice: AI matches a tumor's genomic profile to the most promising targeted therapy.

Using data and models to forecast future events or risks.

In practice: Predictive analytics identifies patients likely to deteriorate in the next 24 hours.

The instruction or question given to a generative model to guide its output.

In practice: A well-designed prompt asks an LLM to summarize a paper for a patient audience.

Formal authorization by a health authority for a medical AI product to be marketed or used.

In practice: SFDA or FDA clearance signals an AI tool met safety and performance requirements.

Training an agent to make sequential decisions by rewarding desired outcomes and penalizing poor ones.

In practice: Reinforcement learning helps optimize insulin dosing strategies in simulation.

A technique where a model retrieves relevant documents and uses them to ground its generated answer.

In practice: A RAG assistant cites the specific guideline passage it used to answer a dosing question.

The proportion of true positive cases a model correctly identifies (true positive rate).

In practice: High sensitivity means a cancer-screening model misses very few actual cancers.

The proportion of true negative cases a model correctly identifies (true negative rate).

In practice: High specificity means few healthy patients are wrongly flagged as positive.

Information organized in a fixed format such as tables, fields, or codes.

In practice: Lab results and vital signs in an EHR are structured data.

Training a model on labeled examples where the correct answer is known, so it learns to predict labels for new data.

In practice: A model trained on labeled biopsy images learns to distinguish benign from malignant tissue.

The dataset used to teach a model the patterns it needs to make predictions.

In practice: A pathology model's training data might include 100,000 annotated slide images.

Information without a predefined format, such as free text, images, or audio.

In practice: Clinical notes, radiology images, and dictated reports are unstructured data.

Finding structure or groupings in data that has no predefined labels.

In practice: Clustering patients by lab-value patterns can reveal previously unrecognized disease subtypes.

Data held back from training and used to measure how well a model performs on unseen cases.

In practice: A diagnostic model is evaluated on a separate test set of patients it never learned from.