Skip to main content

CME Delivery — Continuing Medical Education That Doesn't Suck

A paid, clinician-facing product that shares the case-generation engine with the Diagnostic Game but is sold and used very differently. Two heads, one body.

The two-headed framing: the same engine — synthetic case generation + LLM patient role-play + structured medical artifact generation (labs, imaging, EKG) — feeds two products with different audiences. CME for licensed clinicians (paid) is one head. The consumer Diagnostic Game is the other. Build the engine once, ship two products.

The pain we're attacking​

Every licensed physician in the US is required to complete Continuing Medical Education (CME) to maintain licensure. Most jurisdictions require an annual or biennial credit total. The CME market is mostly box-checking.

Erik's read on Dr. P's behavior (which is widespread):

She grabs whatever free modules she can to satisfy the requirements. She's motivated to save money and get in and get out — even though it's defeating the educational purpose of CME.

The paid CME modules that exist aren't necessarily better — they're more polished theater. The whole industry has settled into "compliance product, not education product."

The opportunity​

Build CME that's actually engaging — and clinicians will pay for it, because CME they don't dread is genuinely valuable to them as practitioners and to their patients. The economics are favorable: clinicians already pay for CME, and they pay willingly when the alternative is grinding through narrated PowerPoints.

Product shape​

A case-based, interactive CME platform where each "module" is a diagnostic case the clinician works through:

  • The case. A synthetic patient (LLM-generated, no real PHI) presents with a chief complaint. Clinical history, current meds, family history, prior visits — all generated as a coherent record.
  • The conversation. The clinician asks questions; an LLM role-plays the patient or their family, with medically-consistent (but possibly misleading or red-herring) answers.
  • The workup. The clinician orders labs, imaging, biopsies. The system returns LLM-generated lab values, imaging reports, EKGs that are medically plausible and consistent with the case's underlying truth.
  • The diagnosis. The clinician proposes a diagnosis and treatment plan. The system grades against the case's ground truth.
  • The teaching point. Each case is built around a teaching point: a recently published paper, a new drug indication, a rare presentation, an updated guideline. The post-case debrief surfaces the literature that informed the case and counts toward CME credit.

This is fundamentally different from current CME because the diagnostic reasoning is the curriculum. The clinician doesn't sit through a lecture; they practice the work.

Why it produces real CME, not theater​

CME accreditation bodies (ACCME for US physicians) accredit content providers and require evidence of educational design. The content here:

  • Is built from current peer-reviewed literature.
  • Has measurable learning outcomes (case completion, diagnostic accuracy, post-case knowledge checks).
  • Provides evaluative feedback to the learner.
  • Generates an audit-trail of completion, time spent, and performance.

ACCME accreditation is a gating step for shipping this as paid CME. It's a real process, not a paperwork exercise. The accreditation work is the moat.

Why this becomes interesting because of Starlight Practice​

The case-generation engine is something we already have to build for HIPAA-safe testing of Starlight Practice (see Compliance · Synthetic Data Program). We need to generate realistic synthetic patients with histories, conditions, family structures, visits, vitals, notes, and labs — at scale, with edge cases, and with literature seeding for unusual presentations. That's the same engine.

What CME adds on top of the synthetic-data engine:

  • Live LLM role-play loop during case-work (already needed for the post-v1 parent triage tool).
  • Structured artifact generation (labs, imaging reports) — natural extension of synthetic data.
  • Grading and feedback engine — net new.
  • Literature ingestion pipeline — net new (PubMed Central, recent guideline publications, drug-approval feeds).
  • CME accreditation paperwork and audit-trail recording — net new but well-defined.

Cadence and freshness​

CME content needs to feel current. The library wants to refresh as clinical knowledge moves:

  • New drug approval → cases where the new drug is the right answer.
  • New paper on rare presentation → case built around that presentation.
  • Updated guideline → cases that test the updated decision criteria.

Building a literature-ingestion pipeline that converts new publications into seeded case prompts is the long-term competitive moat. Every new Claude release, every new published paper, makes the library better.

Pricing model (sketch)​

CME is sold to clinicians directly or to practices/health systems on bulk subscription. Both work:

  • Direct-to-clinician: subscription-based access to the full library; e.g., $X/month or $X/year covering N CME credits.
  • Practice-bulk: flat per-clinician annual fee, paid by the practice as a benefit (DPC practices already do this for malpractice and other operating costs).

Pricing should be less than what clinicians currently pay for paid CME while delivering a better experience. The market is willing to pay for CME — just nobody enjoys the current product.

Open questions​

  • ACCME accreditation timeline and cost. This is the gating regulatory work and needs scoped early if we want this real.
  • Specialty coverage. Start with primary care (family med, internal medicine, pediatrics) — overlap with Dr. P and the Starlight Practice domain. Specialist CME (cardiology, dermatology, oncology) is a much larger build.
  • Outcome measurement. What does "we made a clinician better" look like? Pre/post diagnostic accuracy? Confidence calibration? Patient outcomes? Set the metric early.
  • AI-generated content disclosure. Some accreditation bodies are wary of AI-generated educational content. Need to understand current ACCME stance and design content provenance to satisfy it (case authored/reviewed by licensed clinicians; LLM is the rendering engine, not the curriculum author).

Cross-references​