Module 2 · 120 min
Healthcare Data
EHR structures, imaging, notes and wearables; provenance, time windows, informative missingness, dataset shift, target leakage, FHIR versus SNOMED CT, and what the European Health Data Space changes.
- Describe the main healthcare data modalities and interrogate the provenance of each.
- Specify a prediction task with an explicit index time, observation window, prediction window and outcome window.
- Recognise informative missingness and explain why 'missing' is itself a clinical measurement.
- Judge whether a dataset represents the population the model will actually serve, and name the shift you fear.
- Detect target leakage in a proposed feature set before it reaches a model.
- Explain what FHIR and SNOMED CT each contribute, and why exchange structure/interoperability and clinical terminology/semantics are complementary but distinct concerns.
- Explain what the European Health Data Space changes about secondary use of health data.
- Run a structured data readiness assessment and defend a go / fix-first / no-go recommendation.
Why it matters
Continuing from Module 1: that module established what a model is, what a label commits you to and when AI is the wrong instrument. Every one of those distinctions is a claim about data before it is a claim about a model, so this module examines the data itself.
Data is the asset that makes any of the value in Module 1 possible. A longitudinal record lets a decision be made earlier and with the whole history in view rather than one encounter at a time. Interoperable exchange removes duplicate requesting, re-typing and chasing, and lets one workflow serve several systems. Shared terminology means a definition built once can be reused across sites instead of rebuilt locally. Governed secondary use lets an organisation validate and improve on its own population rather than borrowing someone else's evidence. None of that is automatic — but it is what the discipline in this module is for.
The discipline matters because many healthcare AI failures begin upstream of the model. Ambiguous outcome definitions, unrepresentative cohorts, target leakage, changing workflows and poorly understood provenance can all undermine impressive offline performance long before an algorithm is chosen.
As a decision-maker you are often shown summary model metrics — most commonly an AUC — without enough detail about the data and evaluation chain behind them. Discrimination alone does not establish data quality, calibration, clinical utility or transportability to your setting.
A disciplined set of questions can expose important weaknesses early — and, just as usefully, confirm that a proposal rests on data worth investing in: When is the index time? What was actually recorded and visible at that moment? Who is included in the cohort, who is excluded, and why? Why might a value be absent? Do the distributions and the underlying care processes differ across time periods or sites in ways that would matter here?
Sources & evidence · 19 sources
This module cites primary legal, standards and specifications, public or consensus guidance, scholarly literature, technical documentation, other cited references.
Content reviewed: September 2026. Publication dates of the individual sources are shown in each citation.
Regulation (EU) 2025/327 of the European Parliament and of the Council on the European Health Data Space.
The primary legal text for both primary and secondary use of electronic health data in the EU. In force since 26 March 2025, with staged application: implementing acts due 2027, main provisions from 2029 and remaining categories from 2031 (status checked August 2026).
Open sourceEuropean Commission — European Health Data Space: regulation and implementation timeline.
Official overview of the EHDS regime, its phased application dates and secondary-use arrangements including the individual opt-out.
Open sourceFDA, Health Canada and MHRA — Good Machine Learning Practice for Medical Device Development: Guiding Principles (October 2021).
Ten guiding principles published jointly in October 2021 by the U.S. FDA, Health Canada and the UK MHRA, covering intended-population representativeness of datasets and independence of training and test data. The IMDRF later issued a related draft on good machine learning practice in June 2024; check both before citing.
Open sourceRegulation (EU) 2016/679 (General Data Protection Regulation), in particular Articles 5, 9, 32 and 89.
Lawful basis and safeguards for processing special-category health data, including research purposes.
Open sourceRegulation (EU) 2024/1689 (Artificial Intelligence Act), Articles 10 and 15.
Data and data-governance requirements for high-risk AI systems: relevance, representativeness, error-freedom to the extent possible, and examination for bias. Consolidated text as amended by Regulation (EU) 2026/1744 (in force 27 July 2026).
Open sourceHL7 FHIR — R5 (v5.0.0, STU) is the current published release; R4 and R4B remain widely deployed (version status checked 10 September 2026).
The resource and profiling model for health data exchange; supports RESTful APIs and search as well as document, messaging and service-based exchange paradigms. Which release a system implements is a procurement question rather than a fixed fact — verify the published current version before relying on it.
Open sourceSNOMED International — SNOMED CT starter guide and editorial guide.
The clinical terminology: concepts, hierarchies, post-coordination and value sets.
Open sourceRegenstrief Institute — LOINC.
Identifiers for laboratory tests, measurements and clinical observations.
Open sourceOHDSI — The Book of OHDSI (OMOP Common Data Model).
Reference for standardising observational health data so studies replicate across sites.
Open sourceCollins 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.
The reporting standard to demand for any prediction model — including cohort, index time, predictors and missing-data handling.
Open sourceGianfrancesco MA, Tamang S, Yazdany J, Schmajuk G. Potential biases in machine learning algorithms using electronic health record data. JAMA Intern Med 2018;178(11):1544–1547.
The canonical short account of missingness, misclassification and informed presence in EHR-derived data.
Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias in an algorithm used to manage the health of populations. Science 2019;366(6464):447–453.
How a cost proxy for health need produced large-scale racial bias — the standard example of label choice as an equity decision.
Wong A, Otles E, Donnelly JP, et al. External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Intern Med 2021;181(8):1065–1070.
What happens when a widely deployed model meets external data and real alert burden.
Finlayson SG, Subbaswamy A, Singh K, et al. The clinician and dataset shift in artificial intelligence. N Engl J Med 2021;385(3):283–286.
A practical taxonomy of dataset shift written for clinicians rather than statisticians.
Kaushal A, Altman R, Langlotz C. Geographic distribution of US cohorts used to train deep learning algorithms. JAMA 2020;324(12):1212–1213.
Concrete evidence of how narrow the geographic base of published clinical AI actually is.
Sáez C, Gutiérrez-Sacristán A, Kohane I, García-Gómez JM, Avillach P. EHRtemporalVariability: delineating temporal dataset shifts in electronic health records. GigaScience 2020;9(8):giaa079.
Methods for detecting the temporal shifts caused by system, coding and protocol changes.
Groenwold RHH. Informative missingness in electronic health record systems: the curse of knowing. Diagn Progn Res 2020;4:8.
Why the act of measuring is itself data, and what that means for imputation.
Rajkomar A, Oren E, Chen K, et al. Scalable and accurate deep learning with electronic health records. npj Digital Medicine 2018;1:18.
A reference point for FHIR-based feature pipelines and rigorous temporal cohort construction.
Kapoor S, Narayanan A. Leakage and the reproducibility crisis in machine-learning-based science. Patterns 2023;4(9):100804.
A taxonomy of leakage types across scientific fields, with the specific errors reproduced in the Leakage Detective.
Capstone Board Pack
Add what you just learned to your own strategy document while it is fresh.