Knowledge Base
Life Sciences & Healthcare AI Glossary
Not textbook definitions. These are terms I use every day - explained with the context you actually need to work with them, not just recognize them.
A
ADMET Prediction
Machine learning models that predict a drug compound's Absorption, Distribution, Metabolism, Excretion, and Toxicity properties before wet-lab synthesis. The real value isn't in replacing chemists - it's in ruthlessly filtering the candidate space so you're only running expensive experiments on compounds with a realistic shot. Most pharma AI companies live or die on how well their ADMET models hold up against real clinical outcomes.
AI Governance
The policies, processes, and oversight structures that control how AI systems are developed, deployed, and monitored in an organization. In healthcare, governance isn't optional - it's how you stay out of regulatory trouble and maintain clinical trust. The gap between companies that say they have AI governance and companies that actually do is enormous; real governance means someone accountable can answer "why did the model make that decision?" at 2am during an audit.
Ambient Scribe
AI-powered clinical documentation that listens to patient-provider conversations and automatically generates structured clinical notes. The holy grail of reducing physician burnout - but getting the accuracy right for billing codes and CPT compliance is where most products quietly fail in production. The best systems I've seen still require a physician review loop; the ones that skip it are engineering toward liability.
B
Bias (Algorithmic)
Systematic errors in AI model outputs that disadvantage certain patient populations - typically arising from training data that underrepresents minorities, women, elderly patients, or rare disease cohorts. In healthcare, biased models don't just give wrong answers, they give wrong answers to the people who are already most underserved by the system. Every model card worth reading quantifies performance across demographic subgroups, not just aggregate accuracy.
Biomarker Discovery
Using machine learning to identify biological signals - genes, proteins, imaging features, wearable readings - that correlate with disease onset, progression, or treatment response. The challenge isn't finding correlations (there are millions); it's finding ones that are causally meaningful and clinically actionable. Most "discovered" biomarkers don't survive external validation, which is why replication cohorts are non-negotiable before anyone calls it a breakthrough.
C
Clinical Decision Support (CDS)
Software that presents patient-specific information to clinicians at the point of care to assist clinical decisions - drug interaction alerts, dosing recommendations, diagnostic suggestions. Passive CDS (alerts that appear) has a 90%+ override rate in busy hospitals because alert fatigue is real. The products that actually change clinical behavior are tightly scoped, embedded in natural workflow, and show evidence not just flags.
Clinical Trial Matching
AI that parses patient records against eligibility criteria for open clinical trials and surfaces relevant matches to providers or patients. The matching problem is hard because eligibility criteria are written in ambiguous natural language and patient data is fragmented across systems. Companies that crack real-time matching on structured + unstructured EHR data have a genuine moat - fewer than a handful do it well today.
Computer-Aided Detection (CADe)
AI systems that flag regions of interest in medical images - a lesion, a nodule, an anomaly - for a radiologist or pathologist to review. CADe is the "second reader" model: it doesn't make the diagnosis, it makes sure you don't miss the finding. FDA clearance for CADe tools is generally more achievable than full diagnostic AI, which is why most imaging AI companies start here before building toward autonomous read.
Continuous Learning AI
AI systems that update their models using real-world deployment data, rather than being frozen after initial training. In healthcare, this is both the dream and the regulatory nightmare - a model that improves from patient outcomes is powerful, but the FDA's predetermined change control plan framework exists precisely because continuous learning without guardrails is how you get silent model drift that harms patients. Most clinical AI today is still static; true continuous learning under regulation is an open problem.
D
Data Drift
The gradual shift in the statistical properties of data flowing into a deployed AI model, causing degraded performance over time without any code changes. In clinical settings, drift happens constantly: coding practices change, EHR systems upgrade, patient populations shift seasonally. A model that hit 94% accuracy at launch can quietly degrade to 81% six months later - and nobody notices until an adverse event surfaces it. Monitoring for drift is not optional in production healthcare AI.
Digital Biomarker
Objective, quantifiable physiological or behavioral data collected by digital devices - wearables, smartphones, continuous glucose monitors - that correlates with health status or disease. The signal-to-noise problem is acute: your Apple Watch generates millions of data points per day, most of which are irrelevant. The companies winning in this space are those who've done the clinical validation work to prove a specific digital signal predicts a specific clinical outcome in a specific population.
Drug Repurposing
Using AI to identify existing approved drugs that may be effective against different diseases than originally intended - dramatically compressing development timelines since safety profiles are already established. COVID-19 accelerated this field: Baricitinib was identified via AI drug repurposing analysis and ultimately received FDA authorization for severe COVID treatment. The bottleneck isn't finding candidates; it's funding the clinical validation for indications that may not be commercially attractive to large pharma.
E
Electronic Health Record (EHR)
The digital system of record for patient clinical data - diagnoses, medications, labs, notes, imaging orders. Epic and Cerner dominate the US market to a degree that shapes the entire healthcare AI industry: if your product doesn't integrate with their APIs, you simply don't exist in most large health systems. The EHR is simultaneously the richest source of clinical training data and the hardest system to extract data from - interoperability is still genuinely broken.
Embedding Model
A model that converts text, images, or other data into dense numerical vectors that capture semantic meaning - the foundation of modern RAG systems and semantic search. In healthcare, medical embedding models trained on clinical corpora (MIMIC, PubMed, clinical notes) significantly outperform general-purpose embeddings on clinical NLP tasks. The embedding model choice is often more consequential for RAG quality than the LLM choice.
Enterprise RAG
Retrieval-Augmented Generation deployed at enterprise scale - connecting LLMs to internal knowledge bases, documents, and data systems rather than relying on training data alone. The gap between a RAG demo and production enterprise RAG is enormous: chunking strategy, retrieval quality, re-ranking, citation integrity, access control, and latency all matter at scale in ways they don't in a weekend prototype. Most enterprise RAG failures are retrieval failures, not LLM failures.
F
FDA SaMD
The FDA's regulatory framework for Software as a Medical Device - AI/ML software intended to diagnose, treat, mitigate, or prevent disease. SaMD clearance (510k, De Novo, PMA) is the threshold that separates clinical decision support "tools" from regulated medical devices. Most healthcare AI companies underestimate the timeline (18-36 months is typical for novel AI diagnostics) and the post-market surveillance requirements that kick in after clearance.
FHIR
Fast Healthcare Interoperability Resources - the HL7 standard for exchanging healthcare information electronically. FHIR R4 is now federally mandated for patient data access via the 21st Century Cures Act, which means every major EHR must expose FHIR APIs. In theory, this makes healthcare data more accessible for AI; in practice, FHIR implementation quality varies wildly, and "FHIR compliant" can mean anything from a fully searchable clinical API to a barely-functional read-only endpoint.
Fine-Tuning
The process of further training a pre-trained foundation model on a domain-specific dataset to improve performance on target tasks. In healthcare, fine-tuning on clinical notes or medical literature can dramatically improve model accuracy on specialized tasks - but it also creates a new version of the model that needs its own validation, documentation, and potentially FDA review. The regulatory implications of fine-tuning a cleared medical AI model are still being worked out by the FDA.
Foundation Model
Large-scale AI models trained on broad datasets (GPT-4, Gemini, Claude, Med-PaLM) that can be adapted to many downstream tasks through prompting or fine-tuning. The shift from task-specific models to foundation models fundamentally changes the economics of healthcare AI - you're no longer building a sepsis model from scratch, you're adapting a general-purpose model with domain data. Med-specific foundation models (BioMedLM, Med-PaLM 2, ClinicalBERT variants) are still maturing but closing the gap on general models for clinical tasks.
G
GenAI
Generative AI - models that produce new content (text, images, code, audio, video) rather than classifying or predicting from existing data. In life sciences, GenAI is being applied to molecule generation, synthetic data creation, clinical document drafting, and patient communication. The hype cycle peaked in 2023-2024; the current challenge is distinguishing GenAI deployments that create genuine clinical or operational value from those that create plausible-sounding output with no measurable outcome improvement.
GxP Compliance
Good Practice regulations (GMP, GLP, GCP, GDP) governing pharmaceutical manufacturing, laboratory processes, clinical research, and distribution. When AI is used in any GxP-regulated process - automated batch record review, clinical trial data analysis, quality control - it inherits the full validation burden of that process. Computer System Validation (CSV) for GxP AI is not a checkbox exercise; regulators expect documented evidence that the system does what it claims to do, consistently, across its intended use environment.
H
Hallucination (LLM)
When a language model generates confident, fluent, but factually incorrect output - invented citations, fabricated drug interactions, nonexistent clinical guidelines. Hallucination is the single biggest barrier to deploying LLMs in high-stakes clinical settings, and "we've reduced hallucination rate to X%" is not a sufficient answer when the cost of a hallucinated drug interaction is patient harm. Grounding (RAG, tool use, structured outputs) reduces hallucination but doesn't eliminate it - which is why human-in-the-loop remains mandatory for clinical LLM deployments.
HIPAA
The Health Insurance Portability and Accountability Act - US federal law establishing privacy and security standards for protected health information (PHI). Every AI product touching patient data must be HIPAA-compliant, and Business Associate Agreements (BAAs) must be in place before PHI can flow to any third-party AI service. The practical implication for AI teams: you cannot just send patient notes to a public LLM API without a BAA, regardless of what the privacy policy says.
Human-in-the-Loop
System design where a human reviews, approves, or overrides AI decisions before they have downstream effects - particularly in high-stakes contexts like clinical recommendations, treatment plans, or regulatory submissions. Human-in-the-loop is not a concession to AI limitations; it's an architectural choice that allocates AI to pattern recognition and humans to judgment, accountability, and edge cases. The design question is where in the workflow the human checkpoint happens, not whether one should exist.
L
Large Language Model (LLM)
Neural networks trained on massive text corpora that develop emergent language understanding and generation capabilities - GPT-4, Claude, Gemini, Llama. In life sciences, LLMs are being deployed for clinical note summarization, medical literature synthesis, patient communication, protocol drafting, and regulatory document generation. The enterprise reality: most healthcare organizations are still in the evaluation/piloting phase because the combination of HIPAA, clinical accuracy requirements, and IT integration complexity makes production deployment harder than demos suggest.
LLM Evaluation
The systematic measurement of LLM performance on task-specific criteria - accuracy, hallucination rate, citation fidelity, toxicity, latency, cost per token. General benchmarks (MMLU, MedQA) are a starting point, but production evaluation requires domain-specific golden datasets that reflect your actual use case and patient population. "It performs well on GPT-4 Evals" is marketing; "it achieves 96% accuracy on our internal oncology note summarization benchmark with 0.3% hallucination rate" is an engineering claim.
M
Medical Coding AI
AI that automates the translation of clinical documentation into ICD-10, CPT, and HCPCS billing codes. This is one of the highest-ROI AI applications in healthcare administration - a 1% improvement in coding accuracy translates to millions of dollars in recovered revenue for large health systems. The accuracy bar is unforgiving: wrong codes mean rejected claims, compliance risk, and potential fraud exposure. Production systems almost universally combine AI suggestion with coder review rather than fully autonomous coding.
Model Card
A structured documentation artifact for a machine learning model - intended use, training data provenance, performance metrics across demographic subgroups, known limitations, and recommended monitoring thresholds. Model cards are increasingly required by hospital procurement committees and are expected by FDA for AI/ML-based SaMD. A good model card tells you not just what the model does well, but exactly where it breaks down and for whom.
Multimodal AI
AI systems that process and integrate multiple data types simultaneously - text, images, audio, structured data, genomic sequences. Healthcare is inherently multimodal: a complete patient picture combines imaging, labs, clinical notes, vitals, and social determinants. Multimodal models that can reason across radiology images and clinical notes simultaneously, or correlate genomic sequences with EHR data, represent the next frontier in diagnostic and prognostic AI. Most current deployments still silo data types; true clinical multimodality is where the compounding value lives.
O
OMOP CDM
The Observational Medical Outcomes Partnership Common Data Model - a standardized vocabulary and schema for transforming heterogeneous healthcare data from different EHR systems into a consistent format for observational research. OMOP is how research networks like PCORnet enable federated analyses across millions of patients without centralizing PHI. If you're building real-world evidence studies or training models across multiple health systems, OMOP fluency is foundational.
P
Pharmacovigilance AI
AI applications that monitor, detect, and assess adverse drug reactions from post-market data - EHR records, social media, regulatory reports, clinical literature. Post-market safety surveillance is a regulatory requirement for all approved drugs, and the volume of adverse event reports has grown beyond human processing capacity. The value proposition is clear: faster signal detection saves lives. The challenge is precision - alert fatigue from too many false positives makes pharmacovigilance AI systems as useless as ones that miss real signals.
Predetermined Change Control Plan (PCCP)
An FDA framework that allows AI/ML-based medical device manufacturers to pre-specify how their model will be updated post-approval without requiring a new regulatory submission for each change. The PCCP is how the FDA is trying to reconcile regulatory oversight with the practical reality that AI models need to evolve. Getting PCCP right in your initial submission is one of the highest-leverage regulatory moves a healthcare AI company can make - it's the difference between being able to ship improvements in weeks vs. years.
Prior Authorization AI
AI that automates the prior authorization process - the insurance approval required before many procedures, medications, and specialist referrals. Prior auth is one of the most hated administrative processes in US healthcare, consuming an estimated 13% of physician time and delaying patient care by days to weeks. AI-assisted PA automation has enormous ROI potential (Humana and Cigna are both investing heavily), but requires careful validation to avoid systematically denying care to specific patient populations.
Prompt Engineering
The practice of designing inputs to language models to reliably elicit desired outputs - through instruction framing, few-shot examples, chain-of-thought reasoning, or structured output formatting. In healthcare, prompt engineering is not a soft skill - a poorly engineered clinical summarization prompt can systematically omit critical information or generate unsafe outputs. The best clinical AI teams treat prompt development with the same rigor as software development: version control, A/B testing, and regression testing against curated evaluation sets.
R
RAG (Retrieval-Augmented Generation)
Architecture that augments LLM generation by first retrieving relevant context from an external knowledge base, then using that retrieved content to ground the model's response. RAG is the dominant architecture for enterprise clinical AI because it reduces hallucination, enables source citation, and allows knowledge bases to be updated without retraining the underlying model. In healthcare, RAG is not just a pattern - it's often the difference between a clinically defensible answer and a plausible-sounding fabrication.
Real-World Evidence (RWE)
Clinical evidence derived from real-world data sources - EHRs, claims, registries, wearables, patient-reported outcomes - rather than randomized controlled trials. The FDA increasingly accepts RWE to support regulatory decisions, particularly for expanded indications and post-market surveillance. The AI opportunity is in turning messy, unstructured real-world data into analysis-ready datasets at scale. The challenge is confounding: real-world data captures who got which treatment, but not always why - and that "why" is where the bias hides.
Regulatory Sandbox
A controlled environment where innovators can test new products and services under regulatory supervision with relaxed rules, enabling real-world learning before full regulatory frameworks are finalized. The FDA's Digital Health Center of Excellence and various EU regulatory sandboxes for AI medical devices are attempts to prevent regulation from becoming an innovation brake. Practically, sandboxes are most valuable for companies doing genuinely novel things where no established regulatory pathway exists yet.
S
SaMD (Software as a Medical Device)
Software intended to be used for one or more medical purposes without being part of a hardware medical device - the regulatory category that covers most clinical AI products. SaMD classification determines your regulatory pathway and burden: a low-risk clinical decision support tool vs. a high-risk autonomous diagnostic AI are the same "software" from an engineering perspective but completely different animals from a regulatory one. Understanding your SaMD risk tier before you start building saves enormous rework later.
Synthetic Data
Artificially generated data that mimics the statistical properties of real patient data, used to augment training datasets, test AI systems, or share data without privacy risk. Synthetic patient data is increasingly used to get around HIPAA barriers in AI development - but the devil is in the fidelity: synthetic data that doesn't preserve rare disease distributions, demographic correlations, or clinically important edge cases will produce models that fail exactly where they're most needed.
V
Validation (Clinical)
The process of demonstrating that an AI model performs as intended in the clinical environment for which it was designed - distinct from technical validation (does the code work?) and analytical validation (does the model predict accurately?). Clinical validation requires prospective studies or rigorous retrospective analyses on representative patient populations, independent from training data. Skipping clinical validation is how you get a model that works beautifully in development and fails in the ICU at 3am on a patient population you never tested on.
Vector Database
Specialized databases optimized for storing and querying high-dimensional embedding vectors - the infrastructure layer that makes semantic search and RAG architectures possible at scale. Pinecone, Weaviate, Qdrant, and pgvector are the most common choices in healthcare AI deployments. The vector DB choice matters less than people think at small scale, and matters enormously once you're indexing millions of clinical documents with strict latency requirements and HIPAA compliance obligations.