D002 数据结构与关联关系分析

references/schema-linkage-contract.md

SchemaLinkageProfile is D002's aggregate-only structural description of a medical-data package.

Sensitivity

The output is still restricted_local because table/column names and aggregate counts may reveal internal hospital system structure. It intentionally contains no sampled patient values.

Table object

Key fields:

• table_ref: D002-local stable reference. Virtual tables may use file_ref#sheet, file_ref#sqlite_table, or file_ref#FHIRResourceType.
• source_ref: original D001 file reference.
• source_format: CSV/TSV/JSON/JSONL/XLSX/SQLite/FHIR JSON.
• profile_status: complete | partial | skipped | error.
• rows_profiled: number of rows whose values contributed to statistics.
• row_count_observed: count observed by the profiler. When capped, it is the observed lower bound, not necessarily total rows.
• row_count_complete: whether the observed count is a complete total.
• columns: aggregate column profiles.
• primary_key_candidates: single-column keys supported by observed uniqueness/null statistics.
• composite_key_candidates: small plausible key combinations.
• inferred_grain: structural grain hypothesis with confidence and rationale.
• risks: table-local structural issues.

Column profile

A column profile may include:

• name
• normalized name
• inferred type
• non-null/null counts and rates
• distinct count/rate where safe
• identifier role
• PHI-risk flag
• time-field candidate flag
• key suitability

No examples, top_values, min_string, max_string, or raw sample value fields are permitted.

Relationship object

A relationship is empirical unless constraint_source says a database FK was actually observed.

Coverage is based on tokenized distinct identifiers:

• left_coverage = overlap / distinct_left
• right_coverage = overlap / distinct_right

Cardinality is inferred from uniqueness on each side:

• 1:1
• 1:N
• N:1
• N:M_or_ambiguous

The result is a candidate relationship, not a clinical assertion.

Linkage graph

linkage_graph summarizes structural anchors:

• patient anchors
• encounter anchors
• specimen/order anchors
• disconnected tables

It must preserve ambiguity instead of forcing a single root.