Artificial Intelligence, Machine Learning, NLP, and Generative AI

Rules-based systems execute explicit logic written by humans. “If potassium is below X, display alert Y” is automation and may be called AI in broad marketing language, but the decision logic is directly specified rather than learned from data.

Machine learning learns statistical patterns from examples. Instead of defining every rule, developers fit a model so inputs are associated with an output such as risk, category, probability, or numerical estimate.

Deep learning uses multi-layer neural networks to learn complex representations. It has enabled major advances in imaging, speech, language, and other high-dimensional data, but usually requires substantial data and computation.

Natural language processing covers computational methods for human language. NLP includes older rule-based and statistical methods as well as modern transformer-based language models.

Generative AI produces new content rather than only assigning a score or category. It can generate text, code, images, audio, summaries, structured data, or combinations. Generative capability creates new workflow possibilities and new risks because the output can be plausible without being true.

Supervised learning uses labeled examples. Inputs such as age, vital signs, diagnoses, and laboratory values are paired with an outcome label, such as readmission or deterioration. The model learns patterns that help estimate the label for new cases.

Features are the input variables available to the model. Feature engineering can encode clinical knowledge, but features can also become proxies for race, socioeconomic conditions, healthcare access, or institutional practice.

Training fits the model; validation estimates whether it generalizes. Evaluation data should be separated from training so performance is not measured only on examples the model already optimized against.

Overfitting occurs when a model learns idiosyncrasies of its training data rather than stable patterns. A model may appear excellent internally and perform poorly at another hospital or after workflows change.

Inference is the use of a trained model on new input. Deployment adds new variables: missing data, latency, workflow timing, user response, and differences between the development population and actual patients.

Sensitivity measures how often true positive cases are detected. High sensitivity can be important when missing a dangerous condition carries substantial cost, but increasing sensitivity may increase false positives.

Specificity measures how often true negative cases are correctly excluded. High specificity reduces false positives but may miss more true cases depending on threshold.

Precision asks how many positive predictions are actually positive. Precision depends partly on prevalence. Even a strong model can generate many false alarms when the event is rare.

Discrimination describes how well a model separates higher-risk from lower-risk cases. Measures such as area under the ROC curve summarize ranking ability but do not tell whether predicted probabilities are accurate.

Calibration asks whether predicted probabilities correspond to observed frequency. If patients assigned 20% risk experience the outcome about 20% of the time, the model is well calibrated in that range.

Subgroup performance matters. Overall metrics can conceal poor performance for specific ages, races, sexes, languages, conditions, care settings, or other populations. The relevant groups depend on intended use and plausible harm.

LLMs process language as tokens rather than as human-like thoughts. Tokens are chunks of text used by the model. The model predicts probable next tokens based on learned patterns and the context provided.

Transformers use attention mechanisms to represent relationships among tokens across context. This architecture allows modern models to handle long-range language relationships and scale to very large training corpora.

Context includes instructions, conversation, documents, tool outputs, and other supplied information. Larger context windows allow more material to be considered, but more context does not guarantee that the model will use every detail correctly.

Generation is probabilistic. A fluent answer reflects a plausible continuation conditioned on the prompt and model—not a database lookup of truth. This explains why confidence in tone should never be treated as evidence.

Reasoning-oriented models may allocate additional computation before producing an answer. This can improve performance on complex tasks but does not remove hallucination, bias, or the need for external verification.

Embeddings represent content as vectors that capture patterns of similarity. They allow systems to retrieve text that is semantically related even when exact keywords differ.

Retrieval-augmented generation combines search/retrieval with generation. A system retrieves relevant passages from an approved knowledge source and supplies them to the LLM so the answer can be grounded in current organizational information.

RAG improves grounding but does not guarantee correctness. Retrieval can miss the right document, return stale content, rank irrelevant passages, or expose conflicting policies. The model can also misinterpret correctly retrieved text.

Chunking and metadata affect retrieval quality. Breaking documents into segments that are too small can remove context; segments that are too large can dilute relevance. Metadata such as department, effective date, and document type can improve filtering.

Citations should point to the actual retrieved source. Generated-looking citations are not provenance. A useful enterprise system lets the user open the supporting policy, paragraph, or source record.

Multimodal models process more than text. Depending on the model, they may analyze images, audio, video, documents, tables, or other modalities. This can support interface review, document extraction, transcription, and multimodal clinical workflows.

Tool use lets a model call external functions rather than answer from memory alone. Tools can search databases, retrieve files, run code, or interact with applications. Tool access increases usefulness and risk because an output can become an action.

Agents sequence multiple steps toward a goal. An agent may retrieve information, compare sources, create a draft, update a system, and report results. Reliability becomes a systems problem involving permissions, checkpoints, logging, and recovery.

Autonomy should scale with reversibility and consequence. Drafting a meeting summary can tolerate more autonomy than changing an EHR configuration or communicating a clinical decision to a patient.

Human review should be designed, not assumed. “A human is in the loop” is weak if the reviewer receives hundreds of outputs, lacks time to verify them, or believes the AI is almost always correct.

Hallucination is generated content that is unsupported or false. It is especially dangerous when the answer is plausible, specific, and delivered with confident language.

Bias can enter through data, labels, sampling, objectives, deployment context, and human interpretation. Eliminating one sensitive variable does not necessarily remove bias because other features may act as proxies.

Distribution shift occurs when deployment data differ from development data. Populations, clinical practice, coding, devices, disease prevalence, or workflows can change model performance.

Model drift refers to performance change over time. Monitoring should detect when relationships or data patterns shift enough to affect safety or value.

Automation bias occurs when people over-rely on computerized recommendations. AI can reduce vigilance if users assume the model has already considered information they themselves have not checked.

Prompt injection and data leakage affect tool-connected systems. External content can manipulate behavior, while poorly governed prompts or logs can expose sensitive information.