Healthcare AI models are computer systems trained on medical data to perform a specific task, and they come in distinct types: large language models that generate and summarize clinical text, computer-vision models that read medical images, and predictive models that estimate risk from structured data. Each type solves a different problem, takes different inputs, and produces different outputs. Understanding the taxonomy matters because the strengths, failure modes, and oversight needs of a radiology model are nothing like those of a documentation assistant or a risk-adjustment tool.
The confusion in most "AI in medicine" conversations comes from treating "AI" as one thing. It is not. Below is a practical map of the model families you will actually encounter, what each does, and where the guardrails belong.
Generative, Predictive, and Discriminative: The Three Big Families
Almost every healthcare model falls into one of three broad categories based on what it produces.
Discriminative models draw boundaries between categories. Given an input, they answer "which class is this?" A model that labels a chest image as "pneumonia present" or "pneumonia absent" is discriminative. Most medical imaging and classification tools live here.
Predictive models estimate a numeric likelihood or future value from structured inputs. "This patient has a 22% probability of readmission within 30 days" is a prediction. Risk-stratification and early-warning tools are predictive.
Generative models create new content — text, and sometimes images — rather than choosing a label. The large language models behind ambient scribes are generative: they produce a paragraph of clinical narrative that did not exist before.
The practical takeaway is that generative models are the hardest to validate, because their output space is open-ended. A discriminative model is either right or wrong about a label; a generative note can be fluent, plausible, and still contain a fabricated detail. That difference drives everything about how you deploy and monitor each type.
Large Language Models for Documentation and Summarization
Large language models (LLMs) are generative systems trained on vast amounts of text. In healthcare their headline use is ambient clinical documentation — "AI scribes" that listen to a patient encounter and draft the visit note automatically.
The problem they solve is administrative burden. Documentation is a leading contributor to clinician burnout, and studies of ambient AI scribes have reported reductions in documentation time, cognitive load, and after-hours charting, with improvements in clinician-reported burnout. A randomized clinical trial of two ambient AI scribes found modest reductions in documentation time and improvements in work exhaustion relative to a control group.
- Inputs: the audio or transcript of a clinical conversation, plus context such as the patient chart.
- Outputs: a structured draft note, a summary, or an answer to a natural-language query.
- The catch: LLM-generated notes have been shown to be more thorough and better organized than physician notes in some evaluations, but also less succinct and more prone to hallucination — inventing plausible-sounding details. That is why every ambient note requires clinician review and sign-off before it enters the record.
The same technology underpins medical scribe workflows and broader summarization tasks. For a wider view of where this is heading, see how AI is transforming healthcare.
Computer Vision and Deep Learning for Medical Imaging
Computer-vision models — usually deep neural networks — interpret pixels. They are the most mature and most regulated category of medical AI. As of late 2025, the U.S. FDA had authorized roughly 1,400 to 1,500 AI-enabled medical devices, and radiology has consistently accounted for around three-quarters of them, according to FDA data reported by industry trackers.
These models detect, classify, or measure findings in images: a suspicious nodule on a CT, diabetic retinopathy in a retinal photo, a large-vessel occlusion on a brain scan, or the ejection fraction on an echocardiogram.
- Inputs: images or image series — X-ray, CT, MRI, ultrasound, pathology slides, retinal photos.
- Outputs: a detection box, a probability score, a segmentation mask, or a triage flag that reprioritizes a worklist.
- The catch: performance depends heavily on the scanners, protocols, and populations the model was trained on. A model validated at one health system can degrade at another — a portability problem well documented across medical AI.
Imaging AI is largely upstream of revenue cycle work, but the accuracy of what gets documented from these reads flows directly into coding and claims.
Predictive and Risk-Stratification Models
Predictive models estimate the probability of a clinical event. They power sepsis early-warning systems, 30-day readmission scores, deterioration alerts, and — importantly for home health — risk adjustment.
Sepsis prediction is the classic example. Published models report reasonable discrimination, but external validation commonly shows performance drops when a model moves to a new site, and high false-alarm rates can drive alarm fatigue. This is why subgroup fairness analysis and independent validation are non-negotiable before clinical use.
- Inputs: structured data — vitals, labs, diagnoses, demographics, claims history.
- Outputs: a risk score or probability, often with a threshold that triggers an alert or a workflow.
- The catch: these models drift as populations and practices change, and they can encode bias if training data underrepresents certain groups.
In home health revenue cycle management, risk-adjustment models help surface the diagnoses that should be captured to reflect a patient's true acuity — the HCC (Hierarchical Condition Category) logic behind accurate reimbursement. Model output is a starting point; certified coders confirm what the documentation actually supports. That is the heart of risk adjustment coding and disciplined healthcare data insights.
NLP for Coding and Information Extraction
Natural-language processing (NLP) models read unstructured clinical text and pull out structured facts — diagnoses, procedures, medications, findings. Unlike a generative LLM writing a note, extraction-focused NLP is about turning narrative into codeable, queryable data.
The problem it solves is that most clinical value is locked in free text. Computer-assisted coding tools use NLP to suggest ICD-10 and other codes from a chart, flag documentation gaps, and route encounters for review.
- Inputs: clinical notes, discharge summaries, OASIS assessments, and other narrative documents.
- Outputs: suggested codes, extracted entities, and links back to the supporting text.
- The catch: NLP suggestions are only as good as the documentation they read. Ambiguous, contradictory, or incomplete notes produce weak suggestions — which is exactly where human coding & OASIS review and clinical documentation review close the gap.
For home health specifically, coding accuracy drives the payment model. If you want the mechanics, the PDGM guide explains how diagnoses and clinical grouping map to reimbursement, and AI in medical coding and RCM covers how these tools fit into the revenue cycle management workflow.
Recommendation and Clinical Decision Support
Decision-support models sit on top of the other families and translate output into an action suggestion: a drug-interaction alert, a suggested order set, a next-best-step prompt, or a care-management flag. They combine predictions, extracted facts, and rules.
The value is reducing missed steps and cognitive load. The risk is the opposite — too many low-value alerts cause clinicians to tune them out. Good decision support is specific, timely, and traceable back to the evidence that triggered it. For a longer view of where this category is going, see the future of AI in healthcare and AI in home health agencies.
The Shared Limits: Validation, Drift, Bias, and Oversight
Regardless of type, four constraints apply to every healthcare AI model.
- Validation: performance on the developer's data does not guarantee performance on yours. External, independent validation is the standard.
- Drift: models decay as populations, coding rules, and clinical practice change. They need monitoring, not set-and-forget deployment.
- Bias and fairness: subgroup analysis across age, sex, and other factors is essential to catch uneven performance.
- Human oversight: for anything touching diagnosis, documentation, or payment, the model produces a draft and a qualified human owns the decision.
Frequently Asked Questions
What is the difference between generative and predictive AI in healthcare?
Generative AI creates new content, such as a large language model drafting a clinical note from a recorded visit. Predictive AI estimates the likelihood of an event from structured data, such as a readmission or sepsis risk score. Generative models are harder to validate because their output is open-ended, while predictive models produce a bounded score that is easier to benchmark but still prone to drift.
Which type of AI model is most common in healthcare today?
Computer-vision models for medical imaging are the most established and most heavily regulated. FDA data through late 2025 shows radiology accounting for roughly three-quarters of all authorized AI-enabled medical devices. Documentation-focused large language models are growing quickly but sit in a different regulatory and workflow category.
Do AI models replace medical coders and clinicians?
No. Current models assist rather than replace. NLP tools suggest codes and surface documentation gaps, imaging models flag findings, and LLMs draft notes — but each output requires review by a qualified professional because of hallucination, drift, and bias risks. The reliable pattern is model-as-draft, human-as-decision-maker.
Model accuracy is only half the equation — the other half is disciplined human review of what those models produce. Medeoan's certified coders validate AI-suggested codes against the documentation that actually supports them, so risk capture and claims reflect the true clinical picture. Learn how that works in our coding & OASIS review service.