Contents
On this page
- From electronic record adoption to data usability
- Documentation burden and interventions with measured effects
- Fitness of routinely collected data for a research question
- Benefits and costs of common data models
- Variability in reported decision-support outcomes
- Patient-generated and remotely collected data
- Reporting data provenance for peer review
- Appraising claims about the future of data collection
- Key points
- Application to your study
- Frequently asked questions
Academy · Research Data Analytics & AI · Data collection, next decade
Medical Data Collection in Healthcare: Challenges, Strategies and Future Evidence
The central question in medical data collection is no longer whether records are electronic but whether their contents can be used. Although adoption has largely been achieved, problems of usability, documentation burden and fitness for research remain. Current evidence supports a narrow set of interventions with measured effects and a considerably larger set without such evidence.
Hospital record systems have been digitised. Annual electronic health record adoption among eligible short-term acute care hospitals rose from 3.2 per cent before the Health Information Technology for Economic and Clinical Health (HITECH) Act to 14.2 per cent by 2015, and such systems are now effectively universal in US acute care. Use did not follow automatically: fewer than half of clinicians regularly draw on outside information during an encounter, even though around 70 per cent of hospitals participate in all four core exchange functions. This article examines where the principal constraints now lie, which strategies have measured effects, and how a genuine advance may be distinguished from a well-funded pilot.
Learning objectives
- The reasons adoption statistics ceased to be an informative measure, and the metrics that replaced them
- The evidence on documentation burden and on the interventions that reduce it, expressed in minutes per day
- Methods for assessing whether routinely collected data are fit for a given research question before the study is designed
- The circumstances in which common data models are beneficial and those in which they result in an unnoticed loss of information
- The reasons alert and decision-support optimisation studies report widely varying results, and the implications of that variance for study planning
- Approaches to writing the data-provenance section of a protocol so that a reviewer can assess it
Common misconceptions
- “Universal adoption means the data is available for research.” Adoption measures installation. Availability depends on exchange, structure and permission, and the gap between them is wide: participation in all four core exchange functions is close to 70 per cent of hospitals, whereas fewer than half of clinicians regularly use outside information during encounters.
- “Routinely collected data is observational data, so the analysis is the hard part.” The principal difficulty usually lies in provenance. Data recorded for billing and care behave differently from data recorded for research: the presence of a measurement is itself informative, coding reflects reimbursement incentives, and absence rarely indicates a normal finding.
- “Artificial intelligence will remove the documentation burden.” The evidence supports a more modest effect. With an ambient scribe, median daily documentation time fell by 6.89 minutes, after-hours time by 5.17 minutes and total daily record time by 19.95 minutes. These reductions are measured and meaningful, but they do not eliminate the problem.
- “Decision support reduces alert fatigue.” The effect varies considerably. Across studies of optimised medication alerts, alert burden fell by between 14 and 90 per cent and positive predictive value ranged from 9 to 100 per cent; moreover, only 2 of 10 studies implemented the system in practice.
- “A predictive model that performs well on records will perform well in the clinic.” Discrimination and deployment are distinct achievements. Deep learning on whole records from 216,221 adult hospitalisations reached an area under the receiver operating characteristic curve (AUROC) of 0.93 to 0.94 for in-hospital mortality but 0.75 to 0.76 for 30-day readmission; neither figure indicates whether acting on the prediction improved patient outcomes.
From electronic record adoption to data usability
For approximately fifteen years the principal metric in health data was penetration: the proportion of hospitals and practices with any electronic record. That metric has served its purpose. The annual adoption rate among eligible short-term acute care hospitals rose from 3.2 per cent before the HITECH incentives to 14.2 per cent by 2015, and US acute care is now effectively fully digitised 35.
The relevant metrics are now whether data are exchanged and whether they are used. Here the picture is less favourable. Around 70 per cent of hospitals participate in all four core exchange functions (sending, receiving, finding and integrating), yet fewer than half of clinicians regularly use outside information during a clinical encounter 42. Capability and behaviour have diverged; this is a different problem from the one that incentive programmes were designed to address, and it does not respond to the same instruments.
The international framing has also changed: more than 120 WHO Member States have developed eHealth strategies or policies, and in that work health data are classified as sensitive personal data predominantly generated by the public sector 29. This is relevant to multi-country research, because the governing instrument is increasingly a national strategy rather than an institutional agreement. The Academy article on electronic health records addresses the record itself; the present article addresses data collection and its trajectory.
Adoption and exchange capability both climbed. Clinician use of outside information did not follow: fewer than half use it regularly during an encounter.
Documentation burden and interventions with measured effects
Documentation burden is the constraint most likely to determine which data are available, because it governs what clinicians actually enter. It is also an area in which recent evidence is unusually concrete, which makes it a useful test case for the interpretation of claims in this field.
Ambient artificial-intelligence scribes have been evaluated by before-and-after measurement rather than by assertion. Median daily documentation time fell by 6.89 minutes, after-hours record time by 5.17 minutes, and total daily time in the record by 19.95 minutes 10. In paired surveys of 38 physicians using the same tool, task load fell by 24.42 points, burnout by 1.94 and usability improved by 10.9, all at p<0.001 9.
These figures illustrate a general principle. The effects are genuine but modest: a reduction of approximately twenty minutes a day in total record time is a meaningful improvement in a clinician’s working life but does not remove documentation work. A vendor claim of transformation and a measured reduction of twenty minutes may describe the same intervention, but only the latter provides a sound basis for planning.
Ambient documentation support: measured rather than projected effects
| What was measured | Change | How it was measured |
|---|---|---|
| Median daily documentation time | −6.89 minutes | Before-and-after record timing |
| After-hours time in the record | −5.17 minutes | Before-and-after record timing |
| Total daily time in the record | −19.95 minutes | Before-and-after record timing |
| Task load | −24.42 | Paired surveys, 38 physicians, p<0.001 |
| Burnout | −1.94 | Paired surveys, 38 physicians, p<0.001 |
| Usability | +10.9 | Paired surveys, 38 physicians, p<0.001 |
As reported in the evaluation of an ambient artificial-intelligence scribe.
Fitness of routinely collected data for a research question
Data collected for care and billing are not neutral raw material. Three properties recur, and each should be examined before, rather than after, a study is designed around the data.
First, informative presence. Whether a test was ordered reflects clinical suspicion, so the existence of a value carries information and missingness is rarely at random. Second, coding behaviour. Administrative codes are assigned under reimbursement rules, so their accuracy varies by condition and by the financial consequence of recording it. Third, free text. Much of the clinically meaningful information is not held in structured fields at all: a fine-tuned clinical language model classifying infection type from free-text antibiotic indications achieved an average F1 of 0.97 internally and 0.98 externally, and identified 31 per cent more cases than the structured diagnostic codes for the same patients 13. A phenotype that depends on codes alone may omit cases on approximately this scale.
Reproducibility work offers some reassurance at the aggregate level. When results were reproduced for 150 real-world evidence studies published in peer-reviewed journals, original and reproduction effect sizes were positively correlated at a Pearson coefficient of 0.85, with a median relative effect of 1.0, although reporting quality across the 250 studies assessed was the limiting factor in whether reproduction could be attempted at all 24,34. The implication is not that routine data are unreliable but that reporting must be sufficiently detailed to permit verification.
Information omitted by a code-only phenotype, measured in the same patients.
Benefits and costs of common data models
Mapping local records to a shared schema makes multi-site research on routine data practicable, and the Observational Medical Outcomes Partnership (OMOP) model has become the common choice. The approach entails both genuine gains and genuine losses, and both should be stated in a protocol.
The principal gain is portability: the same analytic code runs at every site, and the cohort definition is explicit rather than reconstructed by each team 18,26. An OMOP-based rare-disease extension tested on a haematology cohort drawn from 61,697 patients mapped genomic elements at a 100 per cent success rate on a 1,674-record phenotype-genotype dataset, which is a strong result for a domain that standard models handle poorly 16,19.
The cost is that mapping loses information in ways that are not visible in downstream analyses 23,26. Local detail that has no target concept is collapsed, free text usually does not survive the mapping at all, and the vocabulary version used becomes part of the method. The model, the vocabulary version and the mapping date should be stated in the same way as a software version.
Variability in reported decision-support outcomes
Alert optimisation is the clearest example of a field in which the average result is uninformative and the spread constitutes the finding. Across studies of artificial-intelligence-optimised medication alerts, alert burden decreased by between 14 and 90 per cent, and positive predictive value ranged from 9 to 100 per cent 1,16,19.
A range of this width is not random noise; it reflects genuinely different interventions, baselines and definitions reported under a single label. An institution with a poorly tuned baseline has substantial scope for improvement, whereas one that already filters aggressively has little. Positive predictive value depends on the definition of a true positive, which varies between studies.
A more important finding of that review is that only 2 of the 10 included studies implemented the system in practice 19. Most of this literature reports how a model would have performed on retrospective data, which is a legitimate first step but does not constitute evidence of clinical effect. In appraising an alert-optimisation claim, the first question is whether the system was deployed in practice. The same distinction applies to machine learning and clinical impact generally.
When the spread is the finding: optimised medication alerts across 10 studies
| Reported quantity | Range across studies | What the range tells you |
|---|---|---|
| Reduction in alert burden | 14 per cent to 90 per cent | Baseline tuning differs substantially between institutions; a single average would be uninformative |
| Positive predictive value | 9 per cent to 100 per cent | The definition of a true positive differs between studies |
| Studies actually implemented in practice | 2 of 10 | Most of this literature is retrospective simulation, not evidence of clinical effect |
Systematic review of artificial-intelligence-optimised medication alerting.
Patient-generated and remotely collected data
Data collected outside the clinic represent the principal genuine expansion in data collection, and the evidence, although preliminary, is no longer absent. A scoping review of integrating mobile health application data into electronic medical and health record systems included 19 studies covering 113,135 participants, with 13 reporting improved patient outcomes 11.
Two cautions follow for study design. Integration, rather than capture, is the principal difficulty: devices readily produce data, but entering those data into the record in a form on which a clinician will act and which a researcher can analyse is where the work lies 11,25. In addition, the denominator changes: patients who use and persist with an application differ systematically from those who do not, so a cohort assembled in this way carries a selection structure that institutional data do not. Such data should be treated as a distinct source with its own biases rather than as additional records.
Reporting data provenance for peer review
Most protocols using routine data describe the origin of the data inadequately, and reviewers increasingly request this information. An adequate provenance paragraph covers six elements: the source systems and the period; whether the extract is structured, free text or both; the coding systems and their versions; the common data model and vocabulary version if one is used, with the mapping date; what was excluded by the extract and why; and the permission basis under which the data are held and shared 37.
The last element has become increasingly important. With health data classified as sensitive personal data in national eHealth strategies across more than 120 Member States, the permission basis frequently determines which analyses are possible, and identifying a constraint only after the analytic plan has been written is costly 29. The template below sets this out as a fillable stub.
A collective example is instructive: a perspective drawing on the experience of 15 academic medical centres using records to support a public health response showed that the binding constraints were governance and data definition rather than analysis 31. This ordering is typical.
The data-source paragraph of a protocol using routine records
Names the source and stops
Data will be extracted from the hospital electronic health record for all adult patients admitted during the study period. Standard data cleaning will be performed and missing values will be handled using multiple imputation.
“The hospital electronic health record” does not identify a system, a period or whether free text is included, so a reviewer cannot judge whether the phenotype is achievable from it.
No coding systems or versions. A cohort defined on ICD-10 is not the same cohort under a different vocabulary release, and the difference is not recoverable later.
“Missing values will be handled using multiple imputation” assumes that data are missing at random, an assumption that routine laboratory data violate by construction: a test that is absent because the condition was not suspected represents a clinical decision rather than a missing value.
No permission basis, which is the constraint most likely to stop the analysis after it has been designed.
States provenance a reviewer can assess
Data will be extracted from [system] inpatient records for admissions between [dates], comprising structured fields and free-text antibiotic indications. Diagnoses are coded in ICD-10 [version]; medications in [vocabulary, version]. The extract will be mapped to the OMOP CDM [version] using vocabulary release [date]; concepts without a target mapping will be retained in a site-local table and reported. Laboratory values are absent when not ordered rather than when normal, so presence of a measurement will be modelled as informative rather than imputed as missing at random. Data are held under [legal basis] and shared under [agreement]; free text does not leave the institution.
Every element a reviewer needs is present and checkable: systems, period, modality, coding systems with versions, model and vocabulary with a mapping date.
Retaining unmapped concepts in a site-local table and reporting them makes the information lost in mapping visible rather than concealed.
Treating the presence of a measurement as informative states a modelling position explicitly and enables the reviewer to contest it, which is the purpose of a methods section.
Naming the legal basis and stating that free text remains within the institution addresses the governance question pre-emptively and establishes which multi-site analyses are possible.
Reviewers of routine-data studies recognise that provenance failures cannot be remedied at the analysis stage; precision in this section therefore increases confidence in the rest of the protocol.
Appraising claims about the future of data collection
Four criteria distinguish a measured advance from a funded aspiration: whether the intervention was implemented in practice or evaluated retrospectively; whether the effect is reported in units on which a reader can act (minutes, percentage points, or a predictive metric with its interval) rather than in adjectives; whether the baseline is described, so that the size of the improvement can be interpreted; and whether it has been measured that acting on the output changed a patient outcome, as distinct from whether the output was accurate.
Applied rigorously, these criteria leave a short list: ambient documentation support with measured minute-level effects, free-text classification that materially outperforms structured codes, common data models that make multi-site work feasible, and mobile data integration with early outcome signals. This constitutes a substantive agenda. It is narrower than the wider discourse suggests, and it is the portion on which a protocol can be based.
Key points
Key points
Digitisation is complete but has not delivered usability: hospital adoption climbed from 3.2 to 14.2 per cent annually across the HITECH period and around 70 per cent of hospitals now exchange across all four core functions, yet fewer than half of clinicians regularly use outside information in an encounter. The remaining constraints are documentation burden, provenance and governance. The interventions with measured effects are narrow and should be characterised precisely: ambient scribes reducing total record time by roughly twenty minutes a day, free-text classification identifying 31 per cent more cases than structured codes, common data models enabling multi-site analysis at a stated cost in lost granularity, and mobile data integration with early outcome signals across 19 studies and 113,135 participants. Routine data should be treated as recorded for another purpose: the presence of a measurement is informative, codes follow reimbursement, and the permission basis often determines the analysis. Provenance should be reported precisely enough for a reviewer to reproduce the extract.
Application to your study
Application to your study
- Replace adoption questions with questions about use. Establish what proportion of the required fields are populated, by whom, and whether clinicians act on them, rather than whether the institution has a record system.
- Profile the data before designing the study around it. For each variable in the phenotype, establish completeness, who enters it, whether absence indicates a normal or an unmeasured value, and whether the signal resides in free text instead.
- Quantify what a code-only definition would miss. Compare a code-based phenotype against chart or free-text review on a sample before adopting it, since the gap has been measured at around 31 per cent of cases in one antibiotic-indication study.
- State the model and vocabulary version. When mapping to a common data model, record the model version, the vocabulary release and the mapping date, and report concepts that had no target rather than allowing them to disappear.
- Establish whether the system was implemented. Before planning around any decision-support or predictive-model result, determine whether the study implemented the system prospectively or simulated it retrospectively; in one review, only 2 of 10 had implemented it.
- Establish the permission basis first. Determine the legal basis, the sharing agreement and whether free text may leave the institution before the analytic plan is written, not after.
Frequently asked questions
Not in high-income acute care, where it is essentially complete. Annual adoption among eligible short-term acute care hospitals rose from 3.2 per cent before HITECH to 14.2 per cent by 2015, and the installed base is now near-universal. The informative measures today are exchange and use, and these diverge sharply: around 70 per cent of hospitals participate in all four core exchange functions while fewer than half of clinicians regularly use outside information during an encounter. Adoption remains a live measure in lower-resource settings, where more than 120 WHO Member States have eHealth strategies at varying stages of implementation.
Yes, by a measured and moderate amount. In evaluation, median daily documentation time fell by 6.89 minutes, after-hours record time by 5.17 minutes, and total daily time in the record by 19.95 minutes. In paired surveys of 38 physicians, task load fell by 24.42 points, burnout by 1.94 and usability improved by 10.9, all significant at p<0.001. A reduction of approximately twenty minutes a day in total record time is a genuine improvement in working conditions. It does not eliminate documentation work, and claims of transformation should be assessed against these figures.
In routine data a value is usually absent because the test was not ordered, not because the measurement failed. Ordering reflects clinical suspicion, so the presence of a measurement carries information about the patient’s state, a property usually termed informative presence. Standard multiple imputation assumes that data are missing at random conditional on the observed data, an assumption that this situation violates by construction. A preferable approach is to model presence explicitly as a variable and to state this in the protocol, which allows a reviewer to evaluate the assumption rather than discover it.
It depends on the condition, and it can be substantial. A fine-tuned clinical language model classifying infection type from free-text antibiotic indications achieved an average F1 of 0.97 on internal data and 0.98 on external validation, and identified 31 per cent more cases than the structured diagnostic codes for the same patients. That figure is specific to that phenotype and should not be generalised, but it indicates the order of magnitude. The practical step is to validate the code-based definition against chart review on a sample before the study is built on it.
For multi-site work, usually yes, because the same analytic code runs everywhere and the cohort definition becomes explicit rather than reconstructed locally. An OMOP-based rare-disease extension tested on a haematology cohort from 61,697 patients mapped genomic elements at a 100 per cent success rate on a 1,674-record phenotype-genotype dataset, which is a strong result for a difficult domain. The cost is loss of information: local granularity without a target concept is collapsed, free text generally does not survive mapping, and the vocabulary version becomes part of the method. The model version, vocabulary release and mapping date should be recorded, and unmapped concepts reported.
The studies measure different interventions against different baselines under a single label. Across studies of artificial-intelligence-optimised medication alerts, alert burden fell by between 14 and 90 per cent and positive predictive value ranged from 9 to 100 per cent. An institution with a poorly tuned baseline has far more scope for improvement than one already filtering aggressively, and the definition of a true positive differs between studies. The more important finding of that review is that only 2 of the 10 studies implemented the system in practice, so most of the literature describes retrospective simulation rather than clinical effect.
At the aggregate level the record is reasonably encouraging. When results were reproduced for 150 real-world evidence studies published in peer-reviewed journals, original and reproduction effect sizes correlated at a Pearson coefficient of 0.85 with a median relative effect of 1.0. The binding constraint was reporting: of 250 studies assessed for reporting adequacy, many could not be attempted at all because the published description was insufficient to rebuild the cohort. The implication concerns individual reports as much as the field: a study that cannot be reproduced because of insufficient reporting is indistinguishable from one whose findings would not reproduce.
Six elements, stated briefly and specifically: the source systems and the period covered; whether the extract includes structured fields, free text or both; the coding systems in use with their versions; the common data model and vocabulary version if one is used, with the mapping date; what the extract excludes and why; and the legal basis under which the data are held and may be shared. The permission basis is the element most often left until last and the one most likely to constrain the analysis, particularly where health data are classified as sensitive personal data under a national eHealth strategy.
References
- Cunningham J, Davies N, Devaney S, Harding M, Holm S, Neumann V, Smith C, Ainsworth J. A Review of Medical Data Sharing Initiatives With a Focus on the Use of Blockchain Technologies: Qualitative Comparative Analysis. Journal of Medical Internet Research. 2026;28:e63575. doi:10.2196/63575 read 9 October 2026. doi.org/10.2196/63575
- Lieberwirth JK, Mittermaier M, Stern AD. Challenges and potential of using digital biomarkers in healthcare and clinical trials. Communications Medicine. 2026;6(1):151. doi:10.1038/s43856-026-01450-8 read 9 October 2026. doi.org/10.1038/s43856-026-01450-8
- Lemieux J, Pfaff ER, Moffitt RA, Nakashima J, Duncan R, Galvez E, Giannini JP, McMurry JA, Walden A, Haendel M. From clinics to discoveries: harnessing national EHR interoperability for research. JAMIA Open. 2026;9(4):ooag137. doi:10.1093/jamiaopen/ooag137 read 9 October 2026. doi.org/10.1093/jamiaopen/ooag137
- European Data Protection Board. Guidelines 1/2026 on processing of personal data for scientific research purposes (version for public consultation). European Data Protection Board. 2026. read 9 October 2026. edpb.europa.eu
- Office of the National Coordinator for Health Information Technology. Health Level 7 (HL7) Fast Healthcare Interoperability Resources (FHIR). HealthIT.gov, Office of the National Coordinator for Health Information Technology. 2026. read 9 October 2026. healthit.gov
- European Medicines Agency. Real-world evidence (Data Analysis and Real World Interrogation Network, DARWIN EU). European Medicines Agency. 2026. read 9 October 2026. ema.europa.eu
- Mononen N, Salo MM, Timonen D, Pohjanoksa-Mantyla M, Airaksinen M. Receipt of Medicines Information From the Internet and Other Information Sources Among Adult Medicine Users in Developed Economies, 2010-2025: Systematic Review. Journal of Medical Internet Research. 2026;28:e71984. read 9 October 2026. jmir.org
- Jaffe DH, Malin BA, Hendricks-Sturrup RM. A real-world data challenge: guidance for aligning data privacy compliance and fit-for-purpose usability. Health Affairs Scholar. 2025;3(11):qxaf210. doi:10.1093/haschl/qxaf210 read 9 October 2026. doi.org/10.1093/haschl/qxaf210
- Shah SJ, Devon-Sand A, Ma SP, Jeong Y, Crowell T, Smith M, Liang AS, Delahaie C, Hsia C, Shanafelt T, Pfeffer MA, Sharp C, Lin S, Garcia P. Ambient artificial intelligence scribes: physician burnout and perspectives on usability and documentation burden. Journal of the American Medical Informatics Association. 2025;32(2):375-380. doi:10.1093/jamia/ocae295 read 9 October 2026. doi.org/10.1093/jamia/ocae295
- Ma SP, Liang AS, Shah SJ, Smith M, Jeong Y, Devon-Sand A, Crowell T, Delahaie C, Hsia C, Lin S, Shanafelt T, Pfeffer MA, Sharp C, Garcia P. Ambient artificial intelligence scribes: utilization and impact on documentation time. Journal of the American Medical Informatics Association. 2025;32(2):381-385. doi:10.1093/jamia/ocae304 read 9 October 2026. doi.org/10.1093/jamia/ocae304
- Lin J, Bates SM, Allen LN, Wright M, Mao L, Kidd M. Integrating Mobile Health App Data Into Electronic Medical or Health Record Systems and Its Impact on Health Care Delivery and Patient Health Outcomes: Scoping Review. JMIR mHealth and uHealth. 2025;13(1):e66650. doi:10.2196/66650 read 9 October 2026. doi.org/10.2196/66650
- Cheng AC, Shyr C, Lewis A, Delacqua F, Bosler T, Banasiewicz MK, Taylor R, Lindsell CJ, Harris PA. Opportunities, barriers, and remedies for implementing REDCap integration with electronic health records via Fast Healthcare Interoperability Resources (FHIR). JAMIA Open. 2025;8(5):ooaf111. doi:10.1093/jamiaopen/ooaf111 read 9 October 2026. doi.org/10.1093/jamiaopen/ooaf111
- Yuan K, Yoon CH, Gu Q, Munby H, Walker AS, Zhu T, Eyre DW. Transformers and large language models are efficient feature extractors for electronic health record studies. Communications Medicine. 2025;5(1):83. doi:10.1038/s43856-025-00790-1 read 9 October 2026. doi.org/10.1038/s43856-025-00790-1
- Heumos L, Ehmele P, Treis T, Upmeier zu Belzen J, Roellin E, May L, Namsaraeva A, Horlava N, Shitov VA, Zhang X, Zappia L, Knoll R, Lang NJ, Hetzel L, Virshup I, Sikkema L, Curion F, Eils R, Schiller HB, Hilgendorff A, et al., Theis FJ. An open-source framework for end-to-end analysis of electronic health record data. Nature Medicine. 2024;30(11):3369-3380. doi:10.1038/s41591-024-03214-0 read 9 October 2026. doi.org/10.1038/s41591-024-03214-0
- Lacson R, Yu Y, Kuo TT, Ohno-Machado L. Biomedical blockchain with practical implementations and quantitative evaluations: a systematic review. 2024. doi:10.1093/jamia/ocae084 read 9 October 2026. doi.org/10.1093/jamia/ocae084
- Ahmadi N, Zoch M, Guengoeze O, Facchinello C, Mondorf A, Stratmann K, et al. How to customize common data models for rare diseases: an OMOP-based implementation and lessons learned. Orphanet Journal of Rare Diseases. 2024;19(1):298. doi:10.1186/s13023-024-03312-9 read 9 October 2026. doi.org/10.1186/s13023-024-03312-9
- Becker R, Chokoshvili D, Dove ES. Legal bases for effective secondary use of health and genetic data in the EU: time for new legislative solutions to better harmonize data for cross-border sharing? International Data Privacy Law. 2024;14(3):223-246. doi:10.1093/idpl/ipae014 read 9 October 2026. doi.org/10.1093/idpl/ipae014
- Reich C, Ostropolets A, Ryan P, Rijnbeek P, Schuemie M, Davydov A, Dymshyts D, Hripcsak G. OHDSI Standardized Vocabularies-a large-scale centralized reference ontology for international data harmonization. Journal of the American Medical Informatics Association. 2024;31(3):583-590. doi:10.1093/jamia/ocad247 read 9 October 2026. doi.org/10.1093/jamia/ocad247
- Graafsma J, Murphy RM, van de Garde EMW, Karapinar-Carkit F, Derijks HJ, Hoge RHL, Klopotowska JE, van den Bemt PMLA. The use of artificial intelligence to optimize medication alerts generated by clinical decision support systems: a scoping review. Journal of the American Medical Informatics Association. 2024;31(6):1411-1422. doi:10.1093/jamia/ocae076 read 9 October 2026. doi.org/10.1093/jamia/ocae076
- Ewoh P, Vartiainen T. Vulnerability to Cyberattacks and Sociotechnical Solutions for Health Care Systems: Systematic Review. Journal of Medical Internet Research. 2024;26(1):e46904. doi:10.2196/46904 read 9 October 2026. doi.org/10.2196/46904
- Parker G. Future of Electronic Health Records: A Challenge to Maximize Their Utility. SSRN Electronic Journal. 2023. doi:10.2139/ssrn.4457214 read 9 October 2026. doi.org/10.2139/ssrn.4457214
- Walker DM, Tarver WL, Jonnalagadda P, Ranbom L, Ford EW, Rahurkar S. Perspectives on Challenges and Opportunities for Interoperability: Findings From Key Informant Interviews With Stakeholders in Ohio. JMIR Medical Informatics. 2023;11(1):e43848. doi:10.2196/43848 read 9 October 2026. doi.org/10.2196/43848
- Oja M, Tamm S, Mooses K, Pajusalu M, Talvik HA, Ott A, Laht M, Malk M, et al. Transforming Estonian health data to the Observational Medical Outcomes Partnership (OMOP) Common Data Model: lessons learned. JAMIA Open. 2023;6(4):ooad100. doi:10.1093/jamiaopen/ooad100 read 9 October 2026. doi.org/10.1093/jamiaopen/ooad100
- Wang SV, Kattinakere Sreedhara S, Schneeweiss S; REPEAT Initiative. Reproducibility of real-world evidence studies using clinical practice data to inform regulatory and coverage decisions. Nature Communications. 2022;13(1):5126. doi:10.1038/s41467-022-32310-3 read 9 October 2026. doi.org/10.1038/s41467-022-32310-3
- Zeng B, Bove R, Carini S, Lee JSJ, Pollak JP, Schleimer E, Sim I. Standardized Integration of Person-Generated Data Into Routine Clinical Care. JMIR mHealth and uHealth. 2022;10(2):e31048. doi:10.2196/31048 read 9 October 2026. doi.org/10.2196/31048
- Kent S, Burn E, Dawoud D, Jonsson P, Ostby JT, Hughes N, Rijnbeek P, et al. Common Problems, Common Data Model Solutions: Evidence Generation for Health Technology Assessment. PharmacoEconomics. 2021;39(3):275-285. doi:10.1007/s40273-020-00981-9 read 9 October 2026. doi.org/10.1007/s40273-020-00981-9
- Sieck CJ, Sheon A, Ancker JS, Castek J, Callahan B, Siefer A. Digital inclusion as a social determinant of health. npj Digital Medicine. 2021;4(1):52. doi:10.1038/s41746-021-00413-8 read 9 October 2026. doi.org/10.1038/s41746-021-00413-8
- Sarma KV, Harmon S, Sanford T, Roth HR, Xu Z, Tetreault J, Xu D, Flores MG, Raman AG, Kulkarni R, Wood BJ, Choyke PL, Priester AM, Marks LS, Raman SS, Enzmann D, Turkbey B, Speier W, Arnold CW. Federated learning improves site performance in multicenter deep learning without data sharing. Journal of the American Medical Informatics Association. 2021;28(6):1259-1264. doi:10.1093/jamia/ocaa341 read 9 October 2026. doi.org/10.1093/jamia/ocaa341
- World Health Organization. Global strategy on digital health 2020-2025. World Health Organization. 2021. read 9 October 2026. who.int
- Froelicher D, Troncoso-Pastoriza JR, Raisaro JL, Cuendet MA, Sa Sousa J, Cho H, Berger B, Fellay J, et al., Hubaux JP. Truly privacy-preserving federated analytics for precision medicine with multiparty homomorphic encryption. Nature Communications. 2021;12(1):5910. doi:10.1038/s41467-021-25972-y read 9 October 2026. doi.org/10.1038/s41467-021-25972-y
- Madhavan S, Bastarache L, Brown JS, Butte AJ, Dorr DA, Embi PJ, Friedman CP, Johnson KB, Moore JH, Kohane IS, Payne PRO, Tenenbaum JD, Weiner MG, Wilcox AB, Ohno-Machado L. Use of electronic health records to support a public health response to the COVID-19 pandemic in the United States: a perspective from 15 academic medical centers. Journal of the American Medical Informatics Association. 2021;28(2):393-401. doi:10.1093/jamia/ocaa287 read 9 October 2026. doi.org/10.1093/jamia/ocaa287
- Peloquin D, DiMaio M, Bierer B, Barnes M. Disruptive and avoidable: GDPR challenges to secondary research uses of data. European Journal of Human Genetics. 2020;28(6):697-705. doi:10.1038/s41431-020-0596-x read 9 October 2026. doi.org/10.1038/s41431-020-0596-x
- Drosatos G, Kaldoudi E. Blockchain Applications in the Biomedical Domain: A Scoping Review. 2019. read 9 October 2026. client.nup.ac.cy
- Rajkomar A, Oren E, Chen K, Dai AM, Hajaj N, Hardt M, Liu PJ, Liu X, Marcus J, Sun M, Sundberg P, Yee H, Zhang K, Zhang Y, Flores G, Duggan GE, Irvine J, Le Q, Litsch K, Mossin A, Tansuwan J, Wang D, Wexler J, Wilson J, Ludwig D, Volchenboum SL, Chou K, Pearson M, Madabushi S, Shah NH, Butte AJ, Howell MD, Cui C, Corrado GS, Dean J. Scalable and accurate deep learning with electronic health records. npj Digital Medicine. 2018;1(1):18. doi:10.1038/s41746-018-0029-1 read 9 October 2026. doi.org/10.1038/s41746-018-0029-1
- Adler-Milstein J, Jha AK. HITECH Act Drove Large Gains In Hospital Electronic Health Record Adoption. Health Affairs. 2017;36(8):1416-1422. doi:10.1377/hlthaff.2016.1651 read 9 October 2026. doi.org/10.1377/hlthaff.2016.1651
- Mandel JC, Kreda DA, Mandl KD, Kohane IS, Ramoni RB. SMART on FHIR: a standards-based, interoperable apps platform for electronic health records. Journal of the American Medical Informatics Association. 2016;23(5):899-908. doi:10.1093/jamia/ocv189 read 9 October 2026. doi.org/10.1093/jamia/ocv189
- Wilkinson MD, Dumontier M, Aalbersberg IJ, et al. (53 authors; Mons B last). The FAIR Guiding Principles for scientific data management and stewardship. Scientific Data. 2016;3:160018. doi:10.1038/sdata.2016.18 read 9 October 2026. doi.org/10.1038/sdata.2016.18
- Adler-Milstein J, DesRoches CM, Kralovec P, Foster G, Worzala C, Charles D, Searcy T, Jha AK. Electronic Health Record Adoption in US Hospitals: Progress Continues, But Challenges Persist. Health Affairs. 2015. doi:10.1377/hlthaff.2015.0992 read 9 October 2026. doi.org/10.1377/hlthaff.2015.0992
- Adler-Milstein J, Jha A. Health information exchange among U.S. hospitals: who's in, who's out and why? Healthcare. 2014;2(1):26-32. read 9 October 2026. help.senate.gov
- Kuraitis V, McNicoll I. read 9 October 2026. e-caremanagement.com
- Journal of the American Medical Informatics Association. doi:10.1093/jamia/ocab289 read 9 October 2026. doi.org/10.1093/jamia/ocab289
- Assistant Secretary for Technology Policy (ASTP). Interoperable Exchange of Patient Health Information Among US Hospitals, 2023. read 9 October 2026. techtarget.com
- Directorate-General for Health and Food Safety, European Commission. Reuse of health data (European Health Data Space: secondary use of health data). European Commission. read 9 October 2026. health.ec.europa.eu
Every source above is a peer-reviewed article, a publisher’s record of one, or a standard issued by the body that owns it. Each carries the date it was last read. Where a standard moves, this article is revised rather than left up with a note.
How to cite this article
Pick your referencing style and copy it. Your university or target journal may vary the punctuation — check its guide if it has one.
Pubrica Academy. (2026, October 9). Medical Data Collection in Healthcare: Challenges, Strategies and Future Evidence. Pubrica. https://pubrica.com/academy/medical-writing/medical-data-collection-challenges-strategies-and-future-impacts-in-healthcare/
Pubrica Academy. “Medical Data Collection in Healthcare: Challenges, Strategies and Future Evidence.” Pubrica, 9 Oct. 2026, pubrica.com/academy/medical-writing/medical-data-collection-challenges-strategies-and-future-impacts-in-healthcare/.
Pubrica Academy (2026) Medical Data Collection in Healthcare: Challenges, Strategies and Future Evidence. Pubrica. Available at: https://pubrica.com/academy/medical-writing/medical-data-collection-challenges-strategies-and-future-impacts-in-healthcare/ (Accessed: 15 October 2026).
Pubrica Academy. “Medical Data Collection in Healthcare: Challenges, Strategies and Future Evidence.” Pubrica. October 9, 2026. https://pubrica.com/academy/medical-writing/medical-data-collection-challenges-strategies-and-future-impacts-in-healthcare/
Pubrica Academy, ‘Medical Data Collection in Healthcare: Challenges, Strategies and Future Evidence’ (Pubrica, 9 October 2026) <https://pubrica.com/academy/medical-writing/medical-data-collection-challenges-strategies-and-future-impacts-in-healthcare/> accessed 15 October 2026.
Pubrica Academy. Medical Data Collection in Healthcare: Challenges, Strategies and Future Evidence. [Internet]. Pubrica; 2026 Oct 9 [cited 2026 Oct 15]. Available from: https://pubrica.com/academy/medical-writing/medical-data-collection-challenges-strategies-and-future-impacts-in-healthcare/
Pubrica Academy, “Medical Data Collection in Healthcare: Challenges, Strategies and Future Evidence,” Pubrica, Oct. 9, 2026. [Online]. Available: https://pubrica.com/academy/medical-writing/medical-data-collection-challenges-strategies-and-future-impacts-in-healthcare/
@misc{medical_data_collection_challenges_str_2026,
author = {{Pubrica Academy}},
title = {Medical Data Collection in Healthcare: Challenges, Strategies and Future Evidence},
howpublished = {Pubrica},
year = {2026},
month = {oct},
url = {https://pubrica.com/academy/medical-writing/medical-data-collection-challenges-strategies-and-future-impacts-in-healthcare/},
note = {Accessed: <span data-cite-acc="iso">2026-10-15</span>}
}This article is written to be learned from. Cite it if you are referring to what it says; if you are using the evidence behind it, cite the sources listed above instead.
Data provenance and fitness-for-purpose stub
The provenance paragraph that reviewers of routine-data studies now expect, together with a per-variable fitness check to be completed before the analytic plan is written. The blanks should be completed from the extract specification.
DATA PROVENANCE (for the methods section)
SOURCE
System(s): _______________________________________________
Period covered: ____________ to ____________
Setting: [ ] inpatient [ ] outpatient [ ] ED [ ] primary care
Modality: [ ] structured fields [ ] free text [ ] imaging [ ] device
CODING
Diagnoses: ______________________ version ____________
Procedures: _____________________ version ____________
Medications: ____________________ version ____________
Laboratory: _____________________ version ____________
COMMON DATA MODEL (omit this block if not mapping)
Model: ________________ version ____________
Vocabulary release: ____________ mapping date: ____________
Concepts with no target mapping: ______ retained locally? [ ] yes [ ] no
Does free text survive the mapping? [ ] yes [ ] no
EXCLUSIONS BUILT INTO THE EXTRACT (not the analysis)
______________________________________________________
Reason: ______________________________________________
PERMISSION
Legal basis: _________________________________________
Sharing agreement: __________________________________
May free text leave the institution? [ ] yes [ ] no
Approval reference: _________________ date: ____________
FITNESS CHECK -- run once per variable in your phenotype
Variable: ____________________________________________
Completeness in the extract: ______ %
Entered by: [ ] clinician [ ] nurse [ ] coder [ ] device [ ] patient
Entered for what purpose: [ ] care [ ] billing [ ] research [ ] audit
When absent, does that mean:
[ ] normal [ ] not measured [ ] cannot tell
If "not measured": presence modelled as informative? [ ] yes [ ] no
Is the signal also, or only, in free text? [ ] yes [ ] no
Validated against chart review on a sample? [ ] yes [ ] no
n reviewed: ______ agreement: ______ %
BEFORE PLANNING AROUND ANY MODEL OR ALERT RESULT
Was the system implemented prospectively, or simulated retrospectively?
[ ] implemented [ ] retrospective only
Is the baseline described well enough to interpret the improvement?
[ ] yes [ ] no
Effect reported in actionable units (minutes / percentage points /
metric with interval) rather than adjectives? [ ] yes [ ] no
Was any patient outcome measured, as distinct from model accuracy?
[ ] yes [ ] no
Pubrica AcademyWritten and maintained by Pubrica’s subject matter experts. Articles are revised when the underlying standards change, and each article states the date on which it was last reviewed.
How Pubrica can help
Pubrica’s biostatistics and data teams profile a routine-data extract before the protocol is written, when provenance problems remain inexpensive to correct. The teams perform a per-variable fitness assessment, validate code-based phenotypes against chart review on a sample, and draft the data-source section so that a reviewer can assess it rather than query it.