Supervised vs. Unsupervised vs. Reinforcement

Three ways machines learn -- and why it matters for your project

The Setup

In Lesson 1, you learned where a model sits in the AI family tree -- traditional ML, deep learning, LLM. That tells you what kind of tool it is.

But there is a second question that is just as important: how did the model learn? Two papers can both use random forests, but one is trying to predict outcomes and the other is trying to discover patient subgroups. Those are fundamentally different tasks, requiring different data, different validation, and different interpretation.

The learning paradigm tells you what data you need before you start. Get this wrong and your entire project is dead on arrival -- not because the algorithm is bad, but because you collected the wrong information.

The Three Paradigms

Every ML model learns in one of three ways. Tap each card to see the details.

Supervised Learning

You give it the questions and the answers. It learns to predict answers for new questions.

Most Common

What you need

A dataset where every row has both the predictor variables (features) and a known outcome (label). The model learns the mapping from features to label.

The surgical analogy

Teaching a resident by showing them 1,000 completed cases: "Here is the patient, here is what happened -- now predict the next one." The resident learns because they can see both the setup and the outcome.

Two flavors

Classification -- the outcome is a category (amputated vs. not, dead vs. alive). This is the ML version of logistic regression. Regression -- the outcome is a number (length of stay, blood loss). This is the ML version of linear regression.

Vascular example: Predicting 30-day mortality after EVAR using VQI data. Features = age, creatinine, aneurysm size, etc. Label = died within 30 days (yes/no). The model sees 10,000 labeled cases and learns to predict the outcome for new patients.

Unsupervised Learning

You give it the questions but no answers. It finds patterns you did not know to look for.

Pattern Discovery

What you need

A dataset with predictor variables but no outcome column. You are not trying to predict anything -- you are asking the data to organize itself into natural groups or reduce complexity.

The surgical analogy

Handing a resident 1,000 PAD patient charts with no diagnoses, no outcomes, and asking: "Do you see any natural groupings?" The resident might notice clusters -- patients who look metabolically similar, or those with similar wound patterns -- without being told what to look for.

Common techniques

Clustering -- grouping similar patients together (k-means, hierarchical clustering). Dimensionality reduction -- collapsing 50 variables into 3 or 4 that capture most of the variation (PCA, t-SNE, UMAP). Often used as a first step before supervised learning.

Vascular example: You have 8,000 CLTI patients with 40 comorbidity variables. Clustering reveals 4 distinct patient phenotypes -- one dominated by renal disease, another by cardiac comorbidity, a third by poorly controlled diabetes, and a fourth with minimal comorbidities. You did not define these groups. The algorithm found them.

Reinforcement Learning

No labeled data at all. The model learns by trial and error, optimizing a reward signal over time.

Rare (for now)

What you need

An environment where the model can take actions, observe consequences, and receive a reward or penalty. It learns a policy -- a strategy for which action to take in each situation -- through repeated interaction.

The surgical analogy

A resident learning to titrate a heparin drip -- not from a textbook, but by adjusting the dose, checking the aPTT, and adjusting again over hundreds of patients. The "reward" is a therapeutic aPTT; the "penalty" is a supratherapeutic or subtherapeutic result.

Where it shows up

Treatment optimization (dynamic dosing, ventilator management), robotic surgery training, and clinical trial adaptive designs. Rare in surgical research today, but growing in critical care and chronic disease management.

Vascular example: An RL agent learning to manage postoperative blood pressure after carotid endarterectomy -- choosing between IV medications, dose adjustments, and timing based on continuous hemodynamic monitoring. The reward is stable SBP in the target range; the penalty is hypertensive or hypotensive episodes. This is largely experimental.

!

Why This Matters for Your Project

The paradigm determines what data you need to collect. If you want to predict outcomes (supervised), you need labeled outcome data for every patient. If you want to discover patient subtypes (unsupervised), you need lots of variables but no outcome is required. If you want to optimize sequential decisions (reinforcement), you need time-series data with actions and results.

This is the first question for any new ML project: Am I predicting, discovering, or optimizing?

How to Tell Them Apart

When you are reading a paper or planning a project, one question separates all three paradigms.

Do you have labeled outcome data?

i.e., do you know the "answer" for each patient in your training set?

YES

Supervised

You know the outcome. The model learns to replicate your labels on new data.

NO

Are you trying to optimize sequential decisions?

i.e., does the agent take actions over time and learn from feedback?

YES

Reinforcement

Learns by trial and error with a reward signal.

NO

Unsupervised

No outcomes, no feedback. Looking for hidden structure.

Quick comparison

	Supervised	Unsupervised	Reinforcement
Needs labeled outcomes?	Yes	No	No (uses rewards)
Goal	Predict	Discover	Optimize
Clinical translation	"What will happen to this patient?"	"Are there patient subtypes we have not recognized?"	"What is the best action to take right now?"
Evaluation	AUC, calibration, accuracy	Silhouette score, clinical validity of clusters	Cumulative reward, policy improvement
Frequency in vasc surg lit	Very common	Occasional	Rare

Common Confusion: Unsupervised learning does not mean the model is "unsupervised" in the sense of being unmonitored or unvalidated. It means the model does not have labeled outcomes to learn from. You still need to validate whether the clusters it finds are clinically meaningful -- that part requires your expertise as a clinician.

You Already Know This

Every paradigm maps to something you have done before. The vocabulary is new; the concepts are not.

Supervised = regression and classification

What you already do

Logistic regression predicting 30-day mortality? Supervised classification. Cox proportional hazards modeling time to graft failure? Supervised regression (with censoring). The only difference with ML is the algorithm is more flexible -- random forest instead of logistic regression, gradient boosting instead of Cox -- but the setup is identical: features in, labeled outcome out.

If you have built a REDCap database with predictor variables and an outcome column, you have set up a supervised learning problem.

Unsupervised = finding structure without a hypothesis

What you already do (less formally)

When you look at a group of CLTI patients and think "these fall into a few distinct patterns" -- that intuition is what clustering formalizes. When a paper uses principal component analysis (PCA) to collapse a long list of variables into a few summary scores, that is dimensionality reduction. You may not have called it "unsupervised learning," but the concept is not new.

Unsupervised learning is exploratory. It generates hypotheses. Supervised learning tests them.

Reinforcement = clinical decision-making over time

What you already do (every day)

You make a clinical decision (start heparin), observe the result (aPTT at 6 hours), and adjust (increase the rate). That iterative process is reinforcement learning in human form. The RL paradigm formalizes this: state (patient vitals), action (drug dose), reward (therapeutic range), repeat.

The difference: RL algorithms can process thousands of patients simultaneously and find optimal policies that account for interactions a single clinician cannot track. The limitation: they need extensive data and a well-defined reward function, which is hard to specify for complex surgical decisions.

Here is the practical implication: when you read an ML paper, the paradigm tells you how to evaluate it. Supervised? Check discrimination and calibration (Module 8). Unsupervised? Ask whether the discovered clusters are clinically actionable and reproducible. Reinforcement? Ask how the reward function was defined and whether it captures what actually matters for patients.

Exercise: Name That Paradigm

Real vascular surgery ML studies, simplified. Identify which learning paradigm each one used.

Study 1 of 8

Lesson Complete!

0

Total Points Earned

Studies (0/8 correct) +0 pts

Perfect Score Bonus +50 pts

Lesson Completed +100 pts

✓

Supervised vs. Unsupervised vs. Reinforcement

Module 1 - Lesson 2 complete

Key Takeaways

Supervised learning needs labeled outcomes: Features + known outcome = supervised. This is the vast majority of surgical ML papers. If you can describe the study as "predicting Y from X," it is supervised.
Unsupervised learning discovers hidden structure: No outcome column needed. Clustering finds patient phenotypes; dimensionality reduction simplifies complex data. It generates hypotheses, not predictions.
Reinforcement learning optimizes sequential decisions: Actions, consequences, rewards, repeat. Rare in surgery today, growing in critical care. Requires a well-defined reward function.
The paradigm dictates your data requirements: This is the first question for any project. Predicting? You need labeled outcomes. Discovering? You need lots of variables. Optimizing? You need time-series with actions and results.
You already understand the concepts: Supervised = regression/classification. Unsupervised = pattern recognition. Reinforcement = clinical titration. The algorithms are new; the logic is familiar.

The next time someone proposes an ML project, your first question is: "Are we predicting, discovering, or optimizing?" The answer determines everything that follows -- what data you collect, what algorithm you use, and how you validate the results.