Reconciling openFDA and AccessData device records is not a matter of forcing two systems to return the same number. The sources expose different backends, refresh schedules, and record groupings, so a defensible reconciliation treats each as a separate feed with its own contract. The core task is to identify stable keys, document extraction time and scope, and apply explicit merge rules for additions, changes, and blank fields. openFDA annotates original FDA records with harmonized fields but does not rewrite them, while AccessData interfaces such as MAUDE may group follow-ups or apply redactions differently [1][4]. A practical workflow therefore begins with source acceptance conditions, then moves to key-based matching and a concrete evaluation example. The goal is analytical lineage, not a single reconciled total that erases the differences between feeds.

Why the Two Sources Disagree

openFDA and AccessData are different systems, and inconsistency is expected when comparing their outputs. A Stack Exchange discussion of MAUDE queries reported that the openFDA API returned 32 reports for a stent search while the MAUDE web interface returned 59, and another query returned 3 API reports against 2 UI records [4]. The explanation offered was that the systems may use different fields or synonyms, de-duplicate or group follow-ups differently, or operate on slightly different refresh cadences and redactions [4]. This means a reconciliation cannot assume one source is simply a delayed copy of the other.

The openFDA documentation states that openFDA does not rewrite original records; it adds harmonized fields as annotations in a special openfda dictionary [1]. Harmonization requires an exact match, so some records will have openfda sections while others will never receive them if no match occurred [1]. Searching harmonized fields therefore returns only a subset of records from a given endpoint [1]. AccessData records may not carry those same annotations, so a join on a harmonized identifier can silently drop rows. Reconciliation must distinguish between a missing value and a failed match.

Establishing Feed Contracts Before Matching

A reconciliation guide from IntuitionLabs proposes five acceptance conditions for each feed: scope, access, timing, coverage, and contract [5]. Scope records which product category and historical window the source actually exposes. Access distinguishes an interactive search, a limited export, an API response, and a complete bulk snapshot. Timing preserves extraction time with an explicit time-zone offset. Coverage tests the chosen cohort using source report keys rather than a displayed headline total. Contract records undocumented fields as unknown until a device-scoped definition and sample support the mapping [5].

These controls make the evidence needed for a switch explicit. The same guide notes that the MDR report key links MAUDE master and child records, downloaded files have monthly updates, and openFDA documents weekly updates [5]. For the device registration and listing API, openFDA states that the time period covered runs from 2007 to 2026-11-02 and updates occur monthly [2]. A reconciliation that compares a monthly AccessData export against a weekly openFDA snapshot without recording those dates cannot explain count differences. The contract step should also note that registration and listing does not denote approval or clearance of a firm or device [3].

Recording a Reconciliation Run

For each run, store the source endpoint, query string, extraction timestamp with time-zone offset, and the report keys returned. Compare incoming keys against existing keys before classifying records as new, and preserve the preceding observation when a change is detected [5]. Keep identifier value, issuer, and jurisdiction together, and record whether a UDI was machine-read or entered manually where that information is available [5]. Treat undocumented fields as unknown until a device-scoped definition and sample support the mapping [5].

Key-Based Matching and Merge Rules

A defensible merge design follows proposed rules for baseline, adds, changes, and precedence [5]. Baseline loads compatible historical and current snapshots while retaining source file and batch identity. Adds compare incoming report keys against existing keys before classifying records as new. Changes resolve the affected master or child identity and preserve the preceding observation. Precedence applies a documented source-specific rule, with the FDA pipeline serving as a reference implementation [5]. The MDR report key is the documented link between MAUDE master and child records [5].

Blank fields require a specific rule. The IntuitionLabs guide cautions that evidence about one implementation is not a universal rule that every blank field means retain the old value [5]. For device identifiers, the guide proposes retaining identifier value, issuer, and jurisdiction together, and preserving whether a UDI was machine-read or entered manually where available [5]. Manufacturer-name consolidation can change aggregation while leaving the underlying report unchanged, so the reported name, the source-provided consolidated name, and the internally curated organization identity should be kept as separate fields [5]. These rules prevent a merge from silently overwriting source values.

A Concrete Evaluation Example

Consider a hypothetical reconciliation of a coronary stent cohort. An analyst pulls openFDA device event records for a date range and brand name, then runs the same search in the AccessData MAUDE interface. The openFDA API returns 32 reports while the UI returns 59 [4]. The analyst records the extraction timestamps, the exact query strings, and the report keys returned by each source. The first check is whether the 59 UI records include follow-ups that the API grouped under a master key, or whether the API snapshot predates records that have since flowed into the UI backend [4].

The analyst then joins on MDR report key where available and flags records present in only one source. For records present in both, the analyst compares device identifiers and manufacturer names. If the AccessData record shows a consolidated firm name while openFDA retains the reported name, both values are stored separately rather than merged [5]. The final output is a lineage table showing source, extraction time, report key, and match status. This is a suggested workflow, not a tested procedure, and it does not establish that the two sources will ever produce identical counts for the same cohort.

What Reconciliation Cannot Establish

Reconciliation does not turn openFDA or AccessData into a single authoritative record. openFDA states that users should not rely on it to make decisions regarding medical care [1][2]. The Library of Congress report notes there is no single searchable platform on the FDA website that allows users to search across the agency's entire catalog of databases [3]. Data availability and accessibility vary depending on scope, types of data, search capabilities, and terminology used in the individual database [3]. A reconciled dataset is therefore a composite with documented provenance, not a replacement for the underlying sources.

Some device views remain unverified. The IntuitionLabs guide states that a stable device export schema, MDR-key equivalence, device API contract, and refresh guarantee for AEMS remain unverified, and that live device records and exports could not be independently inspected [5]. That absence of verification does not establish their absence, but it means a reconciliation involving those feeds must validate the actual device view before treating it as interchangeable [5]. For AccessData and openFDA specifically, the documented differences in refresh cadence and grouping are enough to require per-cohort validation rather than a one-time mapping.