Basics of AI for Physicians

A plain-English introduction to AI for clinicians — what it is, where it touches medicine, and how to use it safely and legally with patient data.

Who it's for: Practicing physicians and advanced practitioners across all specialties — no technical background required

A plain-English introduction to artificial intelligence for clinicians — what it is, where it already touches medicine, how to use it safely and legally with patient data, and how to build a personal AI workflow that saves time without creating liability.

10 lessons · $299 · lifetime access · certificate of completion

Syllabus

  1. Lesson 1: What AI Actually Is (and Isn't): A Physician's Mental Model (25 min)

    • Distinguish the three kinds of AI you'll actually encounter in medicine.
    • Explain, in plain language, how a large language model produces an answer.
    • Tell the difference between AI that's genuinely useful today and AI that's hype.

    Takeaway: AI is a powerful pattern engine, not a colleague. Once you understand that it predicts rather than reasons, you can use it with confidence — and catch its mistakes.

  2. Lesson 2: How AI Works: Neurons, Tokens & Weights — A Look Under the Hood (25 min)

    • Describe, in plain language, how a model learns from examples instead of rules.
    • Picture what a neuron and a neural network actually do.
    • Explain what tokens and weights are, and how they let a model produce language.

    Takeaway: Underneath the jargon, it's a simple loop: examples tune weights, layers of neurons form a network, and tokens let it work with language. Understand that, and the rest of AI stops being mysterious.

  3. Lesson 3: The AI Landscape in Medicine Today: What's Real, What's Hype, and Where It Lives (25 min)

    • Map the major categories of AI already deployed in clinical and administrative medicine.
    • Tell FDA-cleared clinical AI apart from general-purpose and administrative tools.
    • Evaluate any vendor's “AI” claim and judge what's real versus marketing.

    Takeaway: AI is already in medicine — mostly in documentation and the back office, not diagnosis. Knowing the map tells you where the real time savings are for your practice.

  4. Lesson 4: Using AI Safely with PHI: HIPAA, BAAs & the Hard Rules (25 min)

    • State exactly what counts as PHI and why it can't enter a consumer AI tool.
    • Explain why the Business Associate Agreement is the dividing line for compliant use.
    • Run a de-identification habit and pick the right AI tier for any task.

    Takeaway: No PHI in a tool without a BAA. Master de-identification and the enterprise tiers, and you can use AI every day without legal exposure.

  5. Lesson 5: How Rules-Based AI Works: Expert Knowledge, Encoded as Logic (25 min)

    • Name the three components of a rule-based system and how they interact.
    • Explain the recognize–act cycle, and forward vs. backward chaining.
    • Explain why rule systems can't scale — and where they hand off to machine learning.

    Takeaway: A rule-based system reasons but never learns: a knowledge base of human-authored rules, an inference engine that chains them. It's transparent and dependable inside its rules — and capped by everything an expert couldn't write down.

  6. Lesson 6: How Machine Learning Works: Finding the Logic in the Data (25 min)

    • Explain what 'learning from data' means — features, labels, and the model as a function.
    • Describe how a model improves: loss, gradient descent, and back-propagation.
    • Spot overfitting, and judge a model with sensitivity, specificity, and ROC-AUC.

    Takeaway: Machine learning doesn't follow rules — it finds them. A model is a function fit to data by minimizing a loss, one downhill step at a time, with back-propagation sharing the error across every weight. The danger is always overfitting — memorizing instead of learning — which is why you judge a model only on data it never saw, using sensitivity, specificity, and AUC rather than accuracy alone.

  7. Lesson 7: How Large Language Models Work: From Tokens to Attention to a Grounded Answer (26 min)

    • Explain how text becomes tokens and embeddings a model can compute on.
    • Describe attention and the transformer — how a model reads context.
    • Explain next-token generation, why models hallucinate, and how RAG grounds them.

    Takeaway: A large language model turns text into tokens and embeddings, uses attention to read each word in context, and generates an answer one token at a time — always the most plausible next word, which is not the same as the true one. That gap is why they hallucinate. Retrieval-augmented generation closes it by feeding the model real sources to answer from — which is exactly how Doximity and OpenEvidence keep their answers tethered to the evidence.

  8. Lesson 8: AI for the Business of Medicine: Where the Technology Earns Its Keep (25 min)

    • Map where AI fits in the revenue cycle and the practice workflow.
    • Put real numbers on prior-auth time saved and denials prevented.
    • Judge a tool with ROI and break-even — and know where autonomy must stop.

    Takeaway: In the business of medicine, AI pays for itself first in the back office — prior auth, denials, coding, documentation — where the work is repetitive and the safety bar is about time, not diagnosis. Size the prize with volume × time × rate, judge tools by ROI and break-even, and never cross the autonomy line: AI drafts, a human reviews and signs. Speed is the gift; accountability stays with you.

  9. Lesson 9: Build a CPT Code Checker: Putting It Together with AWS Bedrock and Claude (25 min)

    • Describe what the tool does and the structured output it returns.
    • Lay out the architecture — and why the note stays inside your AWS account.
    • Explain the Claude-on-Bedrock call, and the guardrails that keep it safe.

    Takeaway: A CPT checker is the whole course in one small app: a React front end and a FastAPI backend, a single Bedrock call that asks Claude to read the note and return structured JSON, and — the part that makes it responsible — hard validation against the real code set plus a human in the loop. The model proposes; deterministic rules and a clinician dispose. That skeleton — read, propose, validate, confirm — is how to build, or judge, almost any clinical AI.

  10. Lesson 10: Course Summary: From What AI Is to Putting It to Work (20 min)

    Takeaway: You came in with one intimidating word and you leave able to take it apart: three engines — rules, machine learning, language models — each with a predictable strength and failure; four habits that outlast any tool; and five questions that size up anything a vendor shows you. You don't need to fear AI, and you don't need to be dazzled by it. You can reason about it — which is exactly where a physician should stand.

What physicians say

Finally, an easier way to understand Artificial Intelligence and how it can affect my practice. Great courses and easy to follow.

Yvonne L. Figarella, OB/Gyn

Related guides