Syllabus
Intended learners
Medical students, clinical researchers and doctoral researchers who want to judge data from clinical care and translational research and carry out their own analyses in R. Part I is the entry point; Parts III and IV deepen modelling and translation. Teaching teams can adopt single workshops, for example W3 on diagnostic accuracy or W7 on data protection and security; the workshop identifiers W1–W8 and the unit numbers 01–32 stay citable.
Entry requirements
Basic clinical concepts are enough. A first R session with a data table, a figure and four expressions is part of the first workshop. Prior experience with Python, Julia or machine learning is not needed. A laptop with a local R installation is recommended; every task can also be completed with the printed outputs on paper.
Scope and workload
| Component | Hours |
|---|---|
| Contact time: 8 workshops × 4 teaching units of 45 minutes = 32 units | 24 |
| Preparatory reading (core readings of the companion book) | 8 |
| Guided R tasks | 8 |
| Group project | 16 |
| Individual revision | 4 |
| Total | 60 |
Four parts
| Part | Level and guiding question | By the end, learners can … | Product and transition |
|---|---|---|---|
| I · Clinical questions and data (W1–W2) | Entry. What exactly is being claimed here? | explain purpose, target population, reference, independent unit and permissible data flow; check a generated claim against its original source. | A case card, a small data dictionary and an evidence card. These determine the test in Part II. |
| II · Testing and comparing models (W3–W4) | Application. Does the evidence hold in my setting? | build a baseline, split data by group, explain performance and uncertainty, and design a counter-check for imaging or movement data. | A reproducible comparison table, an error pattern and a first model passport. Part III examines which assumption matters most. |
| III · Justifying complex models (W5–W6) | Advanced. Which biology or process lies behind the data? | distinguish repeated measurements, subtypes and mechanisms; run a sensitivity analysis for one model family and name the limit of transfer. | An R report with baseline, counter-check and next experiment. Part IV uses it as an appendix to the decision. |
| IV · Translating models into practice (W7–W8) | Independent application. Which decision is defensible today? | work through legal and ethical questions for a case, check data protection and security separately, handle an incident and defend the project. | A model passport, evidence of reproducibility and a decision with conditions, responsibility and review date. |
Workshops and required reading
| Workshop | Teaching units | Required reading (core readings of the companion book) | Lab or case sheet |
|---|---|---|---|
| W1 · From clinical case to testable question | 01–04 | B01, B03 | Lab W1 |
| W2 · Language models and scientific evidence | 05–08 | B07, B08 | Lab W2, case sheet |
| W3 · Diagnostics, prediction and external evidence | 09–12 | B06, B09 | Lab W3 |
| W4 · What imaging, gait and tissue models measure | 13–16 | B10, B19 | Lab W4 |
| W5 · Causal questions, trajectories and phenotypes | 17–20 | B02, B13 | Lab W5 |
| W6 · Pharmacometrics, forecasting and simulation | 21–24 | B17 and B15 or B18 | Lab W6 |
| W7 · Regulation, data protection and a security incident | 25–28 | B22, B23 | Case sheets W7 |
| W8 · Model colloquium, decision and review | 29–32 | B27, B28 | Colloquium material |
Assessment
One project accompanies all four parts. Each group hands in a one-page decision sheet, a one-page model passport and a commented R script with a start command (project brief). The project is assessed with a published rubric: seven criteria with weights of 15, 25, 15, 10, 10, 10 and 15%, each scored from 0 to 3.
Individual examination. In unit 31, 20 case-based single-best-answer questions in 25 minutes.
Weighting. For graded runs, the course proposes 60% for the individual examination and 40% for the group project. In a stand-alone run, the course team announces the assessment and participation conditions before the course starts. Institutions that recognise single workshops set their own rules for credit. Peer feedback and self-assessment serve as feedback only.
Examination blueprint. The 20 questions are distributed as follows. The questions themselves are kept outside this public repository; the self-test offers newly developed practice cases instead.
| Area | Questions |
|---|---|
| Foundations and source criticism | 4 |
| Testing models | 5 |
| Complex modelling | 4 |
| Ethics | 2 |
| Regulation | 2 |
| Data protection | 1 |
| Information security | 2 |
| Total | 20 |
Teaching organisation
- Group size. The plan assumes six teams of three to four learners.
- Per workshop: an R script, a synthetic data file or generator, pre-computed outputs on the lab page, instructor notes and links to the companion book.
- Technical hurdles. Learners who cannot run R start with the figure and the printed output and then move to a partner’s computer. No task requires an external AI account or any software other than R.
- Team teaching. Clinical–methodological tandems lead W5 and W6. W7 benefits from contributions from data protection, medical device regulation and information security. These roles are suggestions.
- Material status. The editorial team documents permission for any external figure. Real case histories and confidential examination questions are kept only in environments intended for them; the public course uses new cases.
Evaluation
W1 and W8 address the same competence dimensions with two short cases of equal difficulty, developed for this course. The course team reports the share of correct denominators, recognised group leakage and reasoned incident decisions, alongside an anonymous self-assessment and open feedback (evaluation form). A published scale for non-experts’ AI literacy (SNAIL; Laupichler et al., 2023, doi:10.1016/j.chbr.2023.100338) is an option once the language version and permission to use are clarified. A difference between before and after is used to revise the course; a causal claim about the course’s effect would need a suitable study design.
R as the common language
The required path is in R. Base R covers the two-by-two table, identifiers, simulation and beta quantiles; packages are used where they add something concrete: glmnet for penalised regression, lme4 for mixed models, deSolve for differential equations and simmer for discrete-event simulation. All labs share one figure style with the companion book (r/style.R).
Where a Python or Julia implementation offers a concrete advantage, a box on the lab page names the original implementation, the advantage, and what must be identical for a fair comparison (input table, split, unit, metric, tolerance). These comparisons are conceptual unless stated otherwise; no task requires them.
| Situation | Second implementation | Advantage and limit of the comparison |
|---|---|---|
| Grouped split, W3/W4 | scikit-learn GroupKFold | Folds can be checked within an existing Python image pipeline. Shared target: zero shared identifiers; fold order may differ. |
| Image models, W4 | scikit-image and PyTorch transfer learning | Connects to pre-trained networks and documented image operators. Pre-processing and weights must match. |
| Pharmacometrics, W6 | Julia SciML and SciPy solve_ivp | Choice of solver and parameter sensitivity; concentrations at identical time points checked against the analytic solution. |
| Event simulation, W6 | SimPy alongside simmer | Process generators make single events traceable. With identical arrivals and service times, compare the event order first. |
Professional anchors
| Dimension | What the course asks learners to do | Primary source or professional orientation |
|---|---|---|
| Ethics | Record who is affected, avoidable harm and autonomy. Show errors for clinical subgroups with their denominators. Justify the 3Rs and transferability in preclinical projects. | WHO guidance on large multi-modal models (2024) · NC3Rs: the 3Rs · ARRIVE guidelines 2.0 |
| Regulation | Determine intended purpose and actors for the case. Article 4 of the AI Act, as amended by Regulation (EU) 2026/1744, requires measures to support AI literacy. Look up dates of application and specific obligations separately. | AI Act, Regulation (EU) 2024/1689 · Regulation (EU) 2026/1744 · MDR · MDCG 2019-11 Rev. 1 (guidance) |
| Data protection | For health data, work out the legal basis under Article 6 and the exception under Article 9 GDPR separately. Check Articles 5, 25, 28, 32 and 35 along the data flow and Chapter V for third countries. Document national and institutional rules for research. | GDPR, Regulation (EU) 2016/679 · § 27 Federal Data Protection Act (Bundesdatenschutzgesetz, BDSG) |
| Information security | Examine clinical harm from manipulated outputs or missing availability as well as breaches of confidentiality. Test measures in the tabletop exercise and document residual risk. | BSI: Generative KI-Modelle, version 2.0 (2025) · NIST AI 100-2e2025 |
Accessibility and paper alternative
Every figure has a text description. Every lab page prints its complete results, so all tasks can be done on paper; the case sheets for W2 and W7 need no computer at all. The site uses a light and a dark theme and can be printed from the browser. Please tell the course team early if you need other formats or adjustments.
Data and licence
All examples use synthetic data generated with a fixed seed, or explicitly cited published figures. No number describes a commercial product, a real hospital or unpublished project results. Texts are licensed under CC BY 4.0, code under the MIT licence.