EHR/EMR as a Data Source: Clinical Granularity Without Drift

Define the EHR cohort reproducibly. Specify care settings (inpatient, ED, ambulatory), required data density, and enrollment proxies (active patient flags, visit cadence). Use a new-user design for exposures where prior use would bias effect estimates; implement washout periods with payor coverage context to reduce left-censoring. For outcomes, prefer validated EHR algorithms or chart-review subsamples to assess positive predictive value, especially when codes were created for billing, not science.

Orders vs. results and the reality of workflows. EHRs often contain both the order (intent) and result (fact). For endpoints that rely on a lab value, ingest the result with LOINC, unit, reference range, and specimen metadata; for interventions, clarify whether an order counts as exposure or whether administration/dispense confirmation is required. Keep method metadata (assay, device model/firmware, collection time, body site) to explain outliers and enable sensitivity analyses.

Time, clocks, and clinical context. Store local and UTC timestamps to reconcile cross-site daylight savings and time-zone changes. Capture encounter context (admission–discharge–transfer) so longitudinal analyses can disambiguate pre-admission vs. inpatient events. When telehealth is in scope, record visit modality and identity assurance (e.g., 2-factor check) to support data integrity assertions.

Exposure construction. Medication exposure from EHR requires decisions about administered vs. prescribed vs. dispensed. For parenteral therapies recorded as administrations, start exposure at first administration with clinically sensible grace periods. For oral therapies, align ePrescribe events with pharmacy fill data if linked; otherwise, treat prescriptions as intent and test adherence assumptions in sensitivity analyses. For devices and procedures, anchor to procedure timestamps and verify with anesthesia or operative reports when outcomes are procedure-proximal.

Text mining with restraint. Natural language processing can recover smoking status, ECOG, or symptom severity, but free text is PHI-rich and idiosyncratic. Use NLP to suggest fields that are then persisted as structured variables with provenance. Run redaction before export; document model versions and known limitations; and keep the raw notes in a restricted enclave with minimum-necessary access.

Handling missingness and measurement error. In EHRs, “missing” often means “not observed yet” or “not measured because not clinically indicated.” Treat missing lab values and vitals with informative missingness strategies (indicator variables, multiple imputation with auxiliary variables) and report robustness. For measurement error (e.g., height/weight swaps), automate rule checks (unit plausibility, biologic ranges) and route flags to data managers with short, human-readable rationales.

Federated networks and site effects. When data cannot leave institutions, use a common data model and ship algorithms to sites. Keep a manifest of execution environments (terminology versions, algorithm hashes) and save site-level diagnostics (completeness, unit normalization, coding mix) so heterogeneity is transparent. In meta-analysis, consider random effects when site practice patterns differ meaningfully.

Audit-friendly lineage. For every EHR-derived variable, store the code and parameter hash, input tables, and run manifest. Monitors and reviewers should be able to click from a Kaplan–Meier point to the exact lab value or administration record that justified the event, with the locale, unit, and device context visible.

Claims and Linkage: Population Scale With Chronology You Can Defend

What claims measure well—and what they do not. Adjudicated medical and pharmacy claims reliably capture billed encounters, procedural exposure, dispenses, costs, and chronology of care. They are weak for clinical severity, in-hospital administrations not separately billed, and outcomes not tied to reimbursement. Treat diagnosis codes as signals whose meaning depends on setting and count; increase specificity with algorithms that require temporal patterns (e.g., repeated outpatient codes plus confirmatory imaging/procedure).

Membership and continuity. Establish continuous enrollment windows to ensure observable person-time; document medical and pharmacy coverage components separately. Record eligibility gaps, line-of-business switches, and carve-outs (e.g., behavioral health) because they explain missingness. For multi-payer linkages, maintain payer-source provenance to avoid person-time double counting.

Exposure from pharmacy claims. Dispense dates, quantities, and days’ supply enable robust on-treatment definitions. Define permissible gaps and stockpiles; adjust days’ supply for titrations and long-acting formulations. For specialty drugs, integrate buy-and-bill J-codes and NDCs; switch to administration-based exposure when billing reflects infusions rather than retail fills. Pre-declare switch/augmentation rules so censoring is not outcome-dependent.

Outcome ascertainment. For acute events (MI, stroke), use inpatient primary diagnosis codes with procedure corroboration (thrombolysis, PCI). For safety signals (bleeding), combine site-of-service rules with transfusion/procedure codes. For mortality, link to external death indices where legal; otherwise, treat discharge status cautiously. Always run negative control outcomes to probe residual systematic bias that design choices did not remove.

Lag, lookback, and channel bias. Price in claims lag (30–180+ days) when building dashboards and interim analyses. Align lookback windows for comorbidity and prior therapy across cohorts; misaligned baselines produce spurious imbalance. Recognize channeling: new agents may be used in different lines or risk strata; mitigate with active comparators, line-of-therapy proxies, and high-dimensional propensity scores.

Linkage—done once, done right. Most submission-grade RWE links claims with EHRs and registries. Use privacy-preserving tokenization or deterministic keys under a documented legal basis. Store linkage quality metrics (match rates, duplicates, conflicts) and a crosswalk manifest under access control. Never embed identifiers in filenames or logs; treat service accounts as identities with least privilege and immutable logging.

Bias diagnostics and sensitivity. Present covariate balance (standardized mean differences) pre- and post-weighting/matching, falsification endpoints, and quantitative bias analyses (e.g., E-values). Re-run analyses with alternative outcome windows, inpatient-only definitions, or stricter specificity to show robustness. Report intention-to-treat and on-treatment results side by side when clinically meaningful.

Economics and HTA readiness. Claims enable budget impact and cost-effectiveness work. Make unit costs, price year, and perspective explicit; separate allowed vs. paid amounts; and document assumptions for rebates and patient assistance. Use sealed data cuts so re-runs for health technology assessment reviewers reproduce exactly.

PROs and ePRO: Capturing What Matters to Patients—With Psychometric Rigor

Why PROs matter. Many outcomes that sway clinical and payer decisions—symptom burden, fatigue, function, role participation—do not appear in EHR or claims. Patient-reported outcomes fill that gap. They also contextualize safety (e.g., tolerability) and help explain divergence between utilization and well-being. But PROs are only as defensible as their instruments and administration.

Instrument selection and licensing. Choose instruments with demonstrated validity, reliability, and responsiveness for the population and language. Record licensing terms, scoring manuals, and permitted modifications. For custom items, state the construct, response options, recall period, and a plan to establish measurement properties (cognitive interviews, pilot testing) before pivotal use.

Administration discipline. Standardize when and how instruments are delivered (visit-anchored, time-anchored, or event-triggered), including reminders, grace windows, and allowable modes (web, app, SMS, IVR, paper). If mixed modes are unavoidable, test for mode effects and adjust or stratify. Preserve the version and language per submission, with timestamps and identity checks. For decentralized programs, capture device class and app version; provide offline capture with secure sync and hash-checked receipts.

Scoring and missing data. Implement instrument-specific scoring rules transparently (e.g., handling of skipped items and reverse scoring). For partial completion, follow manual-stated thresholds rather than ad-hoc imputation; where instrument guidance is silent, prespecify statistically principled methods and sensitivity checks. Present both change scores and responder analyses (with minimal clinically important difference rationale) so reviewers see magnitude and meaning.

Linking PROs to clinical and claims contexts. PROs gain interpretability when paired with EHR/claims events. Link questionnaires to visits, therapies, and adverse events via timestamps; analyze trajectories around therapy initiation or switching; and test coherence with utilization (e.g., ED visits falling as symptom scores improve). Keep arm-silent presentations in blinded programs to prevent leakage through dashboards.

Privacy, consent, and governance. Keep PRO platforms on least-privilege, token-based access; minimize on-device PHI; and log every export with business justification and watermarking. Consent should state what is collected, recall period burden, whether recontact is possible, and with whom results may be shared. Make language and reading-level appropriate; provide accessibility features and allow assisted completion with reason documentation where permitted.

Quality dashboards and KRIs. Monitor completion rates by wave, device/app versions in use, time-to-completion, item-level missingness, and “straight-lining” heuristics. Promote consequential indicators to Quality Tolerance Limits (e.g., “≥10% of PRO waves below 70% completion,” “≥5% unacknowledged exports,” “five-minute retrieval pass rate <95%”). Crossing a limit triggers dated containment (e.g., pause reports, retrain sites) and a corrective plan with owners.

Packaging for inspection and publication. Maintain a compact dossier: instrument evidence, licenses, administration SOPs, scoring code with hashes, language versions, pilot validation, completion/retention metrics, and linkages to clinical/claims contexts. Publish algorithms (code lists, windows) and share change logs; list deviations from the SAP with impact rationale. The same discipline accelerates regulatory queries, payer reviews, and peer-reviewed manuscripts.

Bottom line. EHR/EMR, claims, and PROs are complementary lenses on health. Treated as a small, disciplined system—clear estimands, harmonized vocabularies, privacy-preserving linkage, psychometric rigor, and audit-ready lineage—they produce RWE that withstands scrutiny and guides decisions that matter to patients, clinicians, regulators, and payers alike.