Technical specialist lab · Module 1 · 35 min

From Clinical Problem to Prediction Problem

Before any code: turn “reduce readmissions” into a prediction task with a defined population, index time, horizon, outcome and intended action — and separate predicting risk from claiming an intervention works.

After this module you can

Take an operational ambition — “reduce readmissions” — and turn it into a prediction task precise enough that a dataset can be assembled, a model can be evaluated and a reviewer can tell you where it would fail.

State population, index time, prediction horizon, outcome and intended action without temporal ambiguity
Decide what information is legitimately available at the prediction moment
Separate a predictive association from a claim about the effect of intervening
Expose the ambiguities hidden inside a phrase like “30-day readmission”
Recognise label definitions that would invalidate everything downstream
Write a concise model development brief a clinical and data team can both sign

No Python is required in this module, and none is used. Module 2 examines the data that this brief implies; the pandas and scikit-learn work starts in Module 3. Those modules are not yet released.

Step 1

“Can you build us something that reduces readmissions?”

Meridan Hospital (fictional) has two sites and an unplanned readmission rate its board considers too high. A transformation lead asks the data team for “an AI model that reduces 30-day readmissions”. There is an extract of roughly 3,000–5,000 adult admissions with discharge-time variables, a readmission prevalence somewhere around 12–15%, and an enthusiastic sponsor. There is also a post-discharge follow-up team with capacity for about 15 extra patients a week.

Who is being predicted for?

All adult inpatients? Emergency admissions only? Does a same-day discharge count? Are maternity, paediatric, psychiatric or day-case admissions in scope? The population determines the prevalence, and prevalence materially affects positive and negative predictive value and the alert burden the service carries. Sensitivity, specificity and AUROC are not determined by prevalence alone, but a different population can still shift them, because it is a different case mix.

At which moment?

A risk score computed on admission, at the discharge decision and after discharge summaries are typed are three different models with three different feature sets. Only one of them can be acted on by a discharge planner.

Predicting what, exactly?

“Readmission” could mean any return to this hospital, any unplanned admission anywhere in the region, or an unplanned admission related to the index condition. Each definition produces a different label column and a different rate.

So that someone does what?

A score with no attached action changes nothing. The follow-up team's capacity of roughly 15 patients a week is a hard constraint on how many patients can usefully be flagged, and therefore on the operating point.

Specification is the cheapest place to fix a modelling project and the most expensive place to get it wrong. A brief written after the first model is fitted tends to be a description of the model rather than a statement of the clinical need.

Four reporting and appraisal frameworks are used throughout this course as structure for the questions worth asking — TRIPOD+AI for what a prediction model study should report, PROBAST+AI for risk of bias and applicability, FUTURE-AI for lifecycle concerns such as fairness, traceability and robustness, and DECIDE-AI for early clinical evaluation. They organise scrutiny. Following them does not by itself demonstrate that a model is safe, effective or fit for a particular service.

From Module 2 onwards the course uses a synthetic dataset built for teaching: roughly 3,000–5,000 admissions, readmission prevalence in the 12–15% range, two sites, and fields such as age, prior admissions, length of stay, selected comorbidities, medication count, creatinine, haemoglobin and discharge destination. It deliberately contains missingness, repeated patients, temporal change, subgroup differences and at least one leakage trap. It is not real patient data, and no model built on it is clinically validated or fit for use on patients.

Lab progress

0 of 6 steps complete

Sources & evidence · 5 sources

This module cites scholarly literature, technical documentation.

Content reviewed: September 2026. Publication dates of the individual sources are shown in each citation.

Collins GS, Moons KGM, Dhiman P, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378.
Reporting items for prediction model studies, including participants, predictors, outcome definition and timing.
Open source
Moons KGM, Damen JAA, Kaul T, et al. PROBAST+AI: an updated quality, risk of bias, and applicability assessment tool for prediction models using regression or artificial intelligence methods. BMJ. 2025;388:e082505.
Separates model development from model evaluation, with domains for participants and data sources, predictors, outcome and analysis.
Open source
Lekadir K, Frangi AF, Porras AR, et al. FUTURE-AI: international consensus guideline for trustworthy and deployable artificial intelligence in healthcare. BMJ. 2025;388:e081554.
Lifecycle guidance across fairness, universality, traceability, usability, robustness and explainability.
Open source
Vasey B, Nagendran M, Campbell B, et al. Reporting guideline for the early-stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI. Nature Medicine. 2022;28:924–933.
Early-stage clinical evaluation, human factors and performance in the intended workflow.
Open source
scikit-learn user guide.
Technical documentation for the library used from Module 3 onwards. Cited for library behaviour only.
Open source