Ideas

Medical Curriculum Intelligence: Analysis, Visualization, and Curriculum Development

Context

We have a substantial medical-school didactic curriculum corpus already extracted into a machine-readable form.

The source material includes, depending on the event/source:

  • slide decks
  • PDFs
  • slide/page layouts
  • learning objectives
  • full text extracted at the individual slide/page level
  • instructional events/talks/sessions
  • Bloom taxonomy assignments for events/objectives
  • basic catalog/metadata
  • likely course/event/instructor/date/time relationships

The current corpus is already represented in DuckDB with full-text search.

The goal at this stage is not to prescribe a particular application or methodology. First inspect the actual available data, schema, provenance, granularity, and completeness. Then determine which of the ideas below can be implemented reliably and which require additional extraction/inference.

The intended audiences are:

  1. Medical-school leadership — understand the curriculum as a whole.
  2. Course directors/curriculum leaders — understand and improve their portion of the curriculum.
  3. Curriculum developers/faculty — use the system during creation and revision of educational material.

The eventual opportunity is larger than a curriculum catalog: potentially a curriculum intelligence and curriculum-development system.

1. First task: characterize the existing data

Before designing the application, inspect the corpus and produce a concise data inventory.

Determine:

Entities available

For example:

  • programs
  • courses
  • blocks/modules
  • events/sessions
  • talks
  • lectures
  • slide decks
  • PDFs
  • slides/pages
  • learning objectives
  • Bloom levels
  • instructors
  • dates/times
  • instructional modality
  • assessments
  • references/resources
  • departments
  • competencies
  • topics/concepts

Do not assume these exist simply because they would be useful.

Relationships

Determine which relationships are explicit versus inferred:

course → event
event → slide/page
event → objective
event → instructor
event → Bloom level
event → assessment
concept → event
objective → concept

Provenance

For every important derived field, determine whether we can retain:

  • source document
  • source page/slide
  • extraction method
  • confidence
  • human-entered vs inferred
  • model/version used for inference

The system should ultimately be able to say:

“Why did you conclude this?”

and navigate back to the underlying slide/page/objective.

This evidence/provenance layer is important for faculty trust.

2. Core analytical opportunity

The central conceptual distinction is:

What the curriculum says it teaches

versus

What the underlying instructional material appears to teach.

Because the corpus includes page/slide-level text, we can potentially analyze both.

The system should explore the curriculum at several levels:

Institution
  ↓
Program/year
  ↓
Course/block
  ↓
Event/session
  ↓
Objective
  ↓
Slide/page
  ↓
Concept/content

Do not assume a graph database is necessary. A relational/DuckDB representation plus semantic indexes may be sufficient.

3. Leadership-level “Curriculum Observatory”

Explore a high-level view answering:

What does our curriculum actually look like?

Potential dimensions:

  • courses
  • instructional events
  • instructional hours
  • learning objectives
  • distinct concepts/topics
  • instructors
  • departments
  • instructional modalities
  • Bloom/cognitive level
  • longitudinal distribution
  • assessment distribution
  • basic science vs clinical content, if this can be reliably classified
  • other meaningful categories supported by the data

Possible visualizations:

Curriculum landscape

Matrix:

              M1 Fall   M1 Spring   M2 Fall   M2 Spring   M3
Concept A       ███        ██         █          █        ███
Concept B       ██         █         ███        ██        █
Concept C       ███        ███        ░          ░        ██

Intensity could represent:

  • instructional time
  • event count
  • slide count
  • objective count

These should not be conflated.

Time/curriculum distribution

Explore:

  • instructional hours by course/year
  • modality by course/year
  • Bloom distribution by course/year
  • topic/concept distribution by course/year

Curriculum concentration

Identify topics receiving disproportionately large amounts of instructional attention.

Do not call concentration “bad” automatically. Some topics should receive substantial attention.

4. Longitudinal concept trajectories

This may be one of the most valuable analyses.

For a concept/topic, reconstruct the learner’s exposure over time.

Example:

ECG

M1
  Introduction
       ↓
M1
  Basic interpretation
       ↓
M2
  Arrhythmias
       ↓
M3
  Clinical application
       ↓
M4
  Advanced/clinical context

Potentially calculate:

  • first exposure
  • number of subsequent exposures
  • spacing between exposures
  • instructional modality at each exposure
  • Bloom level at each exposure
  • whether the concept is subsequently assessed
  • whether cognitive complexity increases over time

This could support a concept-level model such as:

Introduction
→ reinforcement
→ application
→ analysis
→ assessment

Do not assume this progression is educationally optimal. Present it as an observable curriculum trajectory and allow faculty to interpret it.

5. Bloom/cognitive-level analysis

Bloom information is already available and should be exploited.

Potential analyses:

Bloom distribution

By:

  • institution/program
  • year
  • course
  • event
  • concept/topic

Bloom × time

Explore whether the distribution of cognitive levels changes through the curriculum.

Bloom × concept

For important concepts:

Concept: X

Remember       █████
Understand     ███████
Apply          █████
Analyze        ██
Evaluate       ░
Create         ░

Cognitive trajectory

For a concept, show the sequence of Bloom levels encountered over time.

Important caveat:

Bloom assignments should be treated as metadata with uncertainty.

The wording of an objective does not necessarily establish the actual cognitive demand of the instructional activity.

If Bloom classifications were generated automatically, assess their provenance and confidence before presenting them as facts.

6. Curriculum redundancy / overlap

Use the slide/page text and embeddings/semantic similarity to identify potentially redundant material.

Potential levels:

Near duplicates

Multiple sessions appear to teach nearly identical content.

Conceptual redundancy

Different terminology appears to cover substantially overlapping concepts.

Objective redundancy

Learning objectives across courses are highly similar.

Instructional redundancy

Different sessions appear to perform substantially the same instructional function.

Do not label these as “bad redundancy.”

Some repetition is intentional and educationally useful.

Use language such as:

Potential overlap

or

Potential redundancy for faculty review.

Every finding should expose its supporting source material.

7. Potential contradictions / content inconsistency

This is a particularly interesting LLM-assisted analysis.

Look for cases where different sessions appear to make inconsistent claims.

Example:

Session A:
  "X is generally first-line..."

Session B:
  "X is generally reserved for..."

This should be treated as a candidate inconsistency, not an automatically validated contradiction.

The UI should show:

  • statement A
  • source slide/page
  • statement B
  • source slide/page
  • relevant dates/version
  • confidence
  • optional LLM-generated explanation

This could become a high-value curriculum QA function.

8. Curriculum “debt”

Explore whether the corpus can support detection of:

Orphan objectives

Objectives with little/no identifiable supporting instructional content.

Untaught concepts

Important concepts/competencies with little evidence of instruction.

This requires an authoritative concept/competency inventory if available.

Unassessed objectives

Objectives for which no corresponding assessment can be identified.

Assessment without obvious preparation

Assessment content that has weak identifiable instructional support.

Cognitive mismatch

For example:

Objective: Analyze...
Instruction: predominantly Remember/Understand
Assessment: Recall

Again, these should be presented as potential alignment issues, not automated judgments.

Stale content

Potentially identify:

  • old references
  • old terminology
  • outdated clinical recommendations
  • content not revised for long periods

This should be treated separately from ordinary version age. A five-year-old anatomy slide may be perfectly appropriate; a five-year-old treatment recommendation may not be.

Ownership gaps

Concepts appearing across multiple courses without obvious ownership.

9. Instructional modality analysis

If modality can be reliably derived, build matrices such as:

Topic          Lecture   Case   Simulation   Small Group   Clinical   Assessment
Sepsis            4       2        1            2            6           8
Genetics          6       1        0            1            1           3
Communication     1       3        5            6            8           4

Potential questions:

  • Which concepts are primarily delivered through lectures?
  • Which concepts receive application/practice?
  • Where are clinical contexts introduced?
  • Where are concepts assessed?
  • Does instructional modality change longitudinally?

Again: descriptive first, evaluative second.

10. Student journey / learner experience

Build a longitudinal view for a concept or topic.

Example:

M1 Week 3
  Anatomy
      ↓
M1 Week 5
  Physiology
      ↓
M1 Week 8
  Basic ECG
      ↓
M1 Week 12
  Heart failure
      ↓
M2
  Antiarrhythmics
      ↓
M3
  Cardiology clerkship

Each node should link directly to:

  • event
  • objectives
  • slide/page
  • source document

This bridges the gap between a formal curriculum map and the actual learner experience.

12. Course-director workbench

Leadership and course-director interfaces should be different.

A course director might see:

Course: Cardiovascular

42 sessions
68 instructional hours
217 objectives

Coverage
  92% objectives have supporting content
  13 potentially weakly supported
   7 with no identified assessment

Cognitive profile
  Remember      18%
  Understand    31%
  Apply         37%
  Analyze       12%
  Evaluate       2%

Potential issues
  3 potential redundant sessions
  5 objective/content alignment questions
  2 potential inconsistencies
  4 topics introduced but not obviously revisited
  6 assessment alignment questions

Every metric should be drillable to its underlying evidence.

13. AI explanation layer

Use LLMs primarily as an interpretation layer over retrieved evidence, not as the authoritative curriculum database.

Example:

Why does the system think heart failure is overrepresented?

The agent should retrieve:

Course A — 4 sessions
Course B — 3 sessions
Course C — 5 sessions
Course D — 2 sessions

11.3 instructional hours

Potentially overlapping sessions:
A.03
B.17
C.04

Then summarize cautiously:

These sessions contain substantial semantic overlap around HFrEF pathophysiology and treatment. The overlap may represent intentional reinforcement; faculty review is needed.

The evidence should remain visible.

14. “Before creating a new lecture” workflow

This is potentially the most interesting curriculum-development application.

A faculty member enters:

I want to teach X.

The system should first retrieve:

  1. Existing sessions covering X.
  2. Related concepts.
  3. Existing objectives.
  4. Prerequisite material.
  5. Downstream applications.
  6. Existing assessments.
  7. Bloom distribution.
  8. Potential overlap.
  9. Potential gaps.
  10. Relevant curriculum trajectory.

Then assist with creating/revising the new session.

Possible workflow:

Faculty intent
     ↓
Existing curriculum analysis
     ↓
Identify prerequisites / overlaps / gaps
     ↓
Draft objectives
     ↓
Select cognitive level
     ↓
Select instructional method
     ↓
Create/revise content
     ↓
Create assessment
     ↓
Check curriculum alignment

This should be treated as a design assistant, not an autonomous curriculum authority.

15. Instructional-design skill collection

Investigate existing agent skill collections rather than building every capability from scratch.

One promising general educational collection is:

claude-education-skills

It reportedly contains skills covering areas such as:

  • learning-target authoring
  • curriculum alignment
  • coverage auditing
  • scope/sequence
  • instructional design
  • assessment

Determine which skills are actually useful and compatible with the agent environment.

Potentially create a local medical-school-specific skill layer on top of them.

16. Medical education / medical curriculum skills

Investigate medical-specific agent skills, including projects such as:

med-mentor

Treat these as candidate components, not authoritative sources.

Determine whether they contain reusable patterns for:

  • medical teaching
  • clinical reasoning
  • lesson structure
  • learner progression
  • quizzes/assessment

The objective is to find reusable instructional patterns, not necessarily to adopt an entire external framework.

17. Medical/scientific diagramming skills

Investigate:

BioRender mechanism figure skill

Potentially useful for:

  • mechanisms
  • pathways
  • biological processes
  • experimental workflows
  • graphical abstracts
  • biomedical explanatory figures

Excalidraw diagram skills

Investigate skills such as:

  • excalidraw-diagram-skill
  • mcp_excalidraw
  • related educational Excalidraw skills

Particularly valuable features include:

  • diagram type selection
  • structured layouts
  • visual self-validation
  • iterative correction
  • editable output

Diagram Design

Investigate diagram-design-type skill collections that support multiple diagram types and are compatible with coding agents/Pi.

Mermaid

Consider for diagrams that need to live naturally inside Quarto/Markdown.

Prefer editable/vector/structured diagrams over generated raster images when possible.

18. Educational diagram design principles

Explore whether a diagram skill can enforce principles such as:

A diagram should explain something

not merely decorate a slide.

Isomorphism test

If textual labels were removed, would the spatial relationships still communicate the underlying concept?

Appropriate representation

Choose the visual representation based on the concept:

  • causal model
  • process
  • timeline
  • hierarchy
  • mechanism
  • decision tree
  • anatomical relationship
  • comparison
  • workflow
  • quantitative relationship

Do not default to generic flowcharts.

19. Potential technical architecture

Do not assume this architecture, but evaluate it against the existing system:

                     Curriculum Corpus
                            │
                     DuckDB / Parquet
                            │
              ┌─────────────┴─────────────┐
              │                           │
        Structured data              Text/semantic
              │                           │
        SQL analysis                 FTS + embeddings
              │                           │
              └─────────────┬─────────────┘
                            │
                    Curriculum analytics
                            │
          ┌─────────────────┼─────────────────┐
          │                 │                 │
      Leadership       Course Director      Search
      Observatory       Workbench          / Explore
          │                 │                 │
          └─────────────────┼─────────────────┘
                            │
                    AI interpretation
                            │
                    Curriculum design
                            │
               instructional/visual skills

Do not introduce a graph database unless the actual use cases demonstrate that DuckDB + relational relationships + semantic indexes are insufficient.

20. Existing applications to investigate

Consider existing curriculum-management systems, particularly:

  • Elentra
  • Medtrics
  • other medical-school curriculum-management systems

The purpose is not necessarily to replace the existing system.

Instead determine:

  1. What functionality they already provide.
  2. What data they expose/export.
  3. Whether they support APIs.
  4. Whether underlying curriculum mappings can be extracted.
  5. Whether they can coexist with this analytical layer.
  6. Whether their data model is richer or poorer than our current corpus.

The existing extracted slide/page-level corpus may provide a richer content intelligence layer than conventional curriculum-management systems.

21. Suggested initial outputs

Before building a polished application, produce a prototype set of analyses:

A. Curriculum inventory

Basic descriptive statistics and distributions.

B. Curriculum landscape

Topic/concept × course/year visualization.

C. Bloom landscape

Bloom × course/year and Bloom × concept.

D. Longitudinal concept trajectory

Select several important concepts and demonstrate how exposure changes over time.

E. Potential redundancy

Show a small number of high-confidence examples with underlying evidence.

F. Potential inconsistencies

Show several examples with source slide/page evidence.

G. Objective/content/assessment alignment

Only if the necessary assessment data actually exists.

I. Course-director view

One course rendered as an actionable curriculum workbench.

J. Faculty design assistant

Prototype the “before you create a new lecture” workflow.

These prototypes should make it possible to decide what is genuinely valuable before committing to a full application.

22. Important methodological principle

Do not turn every model-derived signal into a “score.”

Prefer:

Evidence
   ↓
Derived observation
   ↓
Faculty interpretation
   ↓
Potential action

rather than:

Model
  ↓
Curriculum score
  ↓
"Good/bad course"

The system should surface evidence and opportunities for review.

Faculty should remain responsible for curricular judgments.

23. Overall hypothesis

The working hypothesis is that the existing DuckDB corpus can evolve from:

a searchable catalog of curriculum materials

into:

a machine-readable institutional representation of the curriculum

and ultimately:

a curriculum intelligence and development environment.

The first phase should establish what the existing data actually supports.

The agent should then prioritize analyses/features based on:

  • data availability
  • reliability
  • faculty usefulness
  • leadership usefulness
  • implementation complexity
  • ability to provide source-level evidence
  • potential for iterative curriculum improvement

Do not prematurely optimize the application architecture or commit to a particular commercial platform.

The primary deliverable from this exploration should be a prioritized set of capabilities grounded in the actual corpus, with examples demonstrating what can already be done and a clear list of additional data/extraction required for higher-value analyses.