World Sleep Industry ENCYCLOPEDIA

Real-World Research And Real-World Data Services

6.5.3

Definition, boundaries, and service types

Real-world data are health-related data collected during routine health care, payment, registries, public-health activity, patient reporting, home life, digital health technology use, or other ordinary processes. Real-world evidence is a conclusion produced by analyzing suitable data with a stated method to answer a defined question. The terms are not interchangeable. A collection of electronic health records, claims, or wearable logs remains a data source; its size, everyday setting, or algorithmic processing does not automatically make it reliable evidence, and RWD does not replace every randomized trial.

Real-world research and data services may cover question formulation, feasibility, source selection, protocol and statistical analysis plan development, ethics and data-use materials, extraction and linkage, quality assessment, programming, interpretation, and reporting. A provider should disclose whether it is a data holder, study sponsor, contract research organization, analyst, or technology platform and whether regulatory-submission support is included. Access to data, completion of a report, or a statistically significant result does not establish that a study meets a particular authority’s evidentiary standard or that a product will be authorized.

Questions, plans, and governance

At initiation, a business request should be converted into a testable question specifying the target population, exposure or intervention, comparator, outcome, time horizon, estimand, and context of use. The team should determine whether a descriptive, diagnostic, predictive, causal, safety-surveillance, or health-services question requires a particular design. Before viewing results, the protocol should fix primary and secondary objectives, eligibility, index time, follow-up, variable definitions, sample size, missing-data approach, subgroups, sensitivity analyses, and amendment rules. Exploratory analyses must be labeled and not retrospectively presented as prespecified confirmation.

Governance should identify the sponsor, data controllers and processors, scientific lead, statistical and clinical reviewers, ethics or privacy review, patient involvement, conflicts, and publication rights. Consent, waiver, authorization, data-use agreement, cross-border transfer, and study-registration requirements depend on the actual jurisdiction and purpose. De-identification can reduce risk but must not be described as eliminating it. Contracts should also govern algorithms, code, derived data, intellectual property, reuse, retention, destruction, incident notice, and audit rather than use an “anonymous data” label to avoid applicable duties.

Data-source inventory and fitness for purpose

Source assessment should document why the data were created, institutions and dates covered, enrollment mechanism, population representativeness, update lag, coding systems, data model, available fields, linkage keys, source-system changes, and access conditions. EHRs reflect clinical documentation, claims primarily reflect billing, registries follow specific inclusion rules, and patient-reported or device data depend on participation, wear, and technical conditions. Any source may miss out-of-network care, nonreimbursed medicine, deaths, social factors, or unrecorded outcomes. Linkage should quantify matching errors and selective linkage, and repeated records from one person must not be treated as independent people.

Fitness for purpose is question-specific. Before extraction, the team should determine whether key exposures, outcomes, confounders, and time information are sufficiently complete, accurate, consistent, plausible, and timely. Feasibility counts are preliminary; a study still requires validation of code sets and phenotypes, source-record review or comparison with credible references where appropriate, and measurement of missingness, anomalies, duplicates, migrations, and institutional variation. Common data models and standard terminologies support reuse and multidatabase studies but do not repair source bias, so mapping loss, versions, and local rules require documentation.

Study design, bias, and causal interpretation

Observational research must address selection bias, confounding, misclassification, information bias, reverse causation, and immortal-time bias. Cohort, case-control, cross-sectional, self-controlled, natural-experiment, pragmatic-trial, or target-trial-emulation designs should follow the question and data-generating process rather than the fields already available. Comparative-effectiveness work should align eligibility, treatment assignment, index time, and start of follow-up and state the causal diagram or conceptual model, confounder rationale, overlap and positivity, treatment consistency, and no-unmeasured-confounding assumptions.

Propensity scores, weighting, matching, standardization, instrumental variables, or machine learning do not eliminate bias by name. Reports should show balance, overlap, extreme weights, model diagnostics, and the estimand. Negative controls, quantitative bias analysis, alternate phenotypes, grace periods, missingness mechanisms, and unmeasured-confounding analyses test robustness rather than provide a menu from which to select the most favorable result. Associations must be interpreted within design limits. Discrimination, calibration, and external validation of a prediction model likewise do not show by themselves that acting on its predictions improves health outcomes.

Data processing, analysis, and reproducibility

The data pipeline should preserve lineage from source fields through extraction, cleaning, unit and time-zone harmonization, code mapping, derived variables, frozen datasets, and analysis tables. Every run should record the data cutoff, software and code versions, parameters, environment, random seeds, quality checks, exclusion counts, and logs, with least-privilege separation of development, production, and release. Multicenter work should harmonize phenotypes and programs in advance while retaining site differences. Federated or distributed analyses also require confirmation that every node ran successfully, outputs meet disclosure controls, and combination methods are valid.

Results should show participant flow, baseline characteristics, follow-up and data density, missingness and censoring, effect sizes, and uncertainty rather than significance alone. Primary, secondary, exploratory, and post hoc findings should be separated, with disclosure of multiplicity, subgroup heterogeneity, model changes, protocol deviations, and analyses producing null results. Independent review should reproduce tables and figures from frozen data and controlled code. When individual-level data cannot be shared for privacy reasons, protocols, variable definitions, code lists, analysis logic, synthetic test data, and execution records can still support scrutiny.

Sleep data and digital measurement

Sleep RWD may come from outpatient and inpatient records, polysomnography, home sleep apnea testing, actigraphy, consumer wearables, bedside or contactless sensors, application logs, sleep diaries, patient reports, prescriptions, claims, and environmental information. A study should identify whether each signal is a raw measurement, device output, or algorithm-derived value and lock the device model, firmware, algorithm, wear location, sampling rate, valid-night rule, time zone, day boundary, and upgrade date. Terms such as sleep duration, awakening, and sleep stage may represent different constructs across systems and should not be merged without validation.

A device or algorithm should demonstrate fitness in the target population, setting, and endpoint, including accuracy, precision, failure rate, wear adherence, data loss, and drift after updates. Assessment should consider variation associated with age, skin tone, body size, comorbidity, medication, shift work, shared devices, and network access. Consumer-derived sleep staging should not automatically be treated as a clinical diagnosis, and a longitudinal association alone does not prove causality or treatment effect. Screening, diagnosis, treatment recommendations, and medical-product regulatory uses require additional clinical, technical, and regulatory evidence appropriate to those uses.

Delivery, transparency, and continuing oversight

Deliverables should match intended use and may include a data-asset and gap inventory, feasibility report, protocol, statistical analysis plan, ethics and governance package, data dictionary, phenotypes and code sets, quality report, analysis code, results report, and auditable archive. Reports should state coverage and cutoff dates, versions, applicable population, estimand, limitations, conflicts, and funding and should register studies and disclose results when required. A provider must not sell data of unverifiable provenance, selectively suppress null findings, or substitute labels such as “regulatory grade,” “whole population,” or “zero bias” for evidence.

After launch, the study team should monitor changes in sources, codes, institutional workflows, devices, and algorithms and assess effects on comparability, drift, missingness, and conclusions. Repeated analyses need explicit version relationships rather than silent replacement. At termination, access should be revoked, copies recovered, data retained or destroyed under the agreement, and open decisions and audit records handed over. Quality is demonstrated by a clear question, fit data, defensible design, reproducible analysis, protected privacy, and transparent limitations—not by data volume, processing speed, or the number of statistically significant findings.