OCL Mapper information architecture

[paynejd]

Proposal:
Elevate Match Algorithms and Concepts (from the target repo) to 1st-class entities in OCL Mapper and separate $lookup from $match

Why?

• Decoupling candidates from concepts means that an algorithm only needs to return a code/concept ID and match metadata (score, matched fragments, etc.)
• Using $lookup (in addition to $match) means that a user may configure an authoritative concept lookup source, if it is not preferable or not possible to rely on a $match algorithm to provide full concept details
  • e.g. $lookup could point to a different repo in OCL, an external FHIR service
• OCL Mapper will load a single canonical representation for each concept that is available in the candidate pool, even if the same code is returned by more than one algorithm
• Algorithms that don't return full concept details (e.g. ocl-scispacy-loinc and external ICD-11 algorithms) will be linked to the canonical concept representation, so a user will still be able to view full concept details
• Formal algo definitions will allow many benefits:
  • No hard-coded algo definitions in the code
  • Configurable and expandable set of algorithms for each mapper instance
  • Treat external match algos as first-class algos
  • Automate optimization of candidate retrieval by using algo attributes to manage batch size, parallelization, etc.

Current OCL Mapper Information Architecture

flowchart TD
    A[Input Dataset] --> B[Mapping Project]
    B --> C[Match Candidates]
    C --> B
    B --> D[Target Repository]

Loading

Planned OCL Mapper Information Architecture

flowchart LR
  A[Input Dataset] --> B[Mapping Project]
  E[Match Algorithms] --> B
  B --> C[Match Candidates]
  D[Target Repository] --> F[Concepts]
  C --> F
  B --> D

Loading

Requirements

Decouple Candidates from Concepts

• Candidates = algo output; code and match metadata (score, matched fragments, etc.) are required; additional attributes are optional – can be minimalist, or fully enriched
• Concepts = single source of truth for the definition of a concept that is shared across algorithms
  • Retrieved via a dedicated $lookup operation, not an algorithm response
  • As an optional optimization, $match algorithms may return a full concept definition, but that is not on the critical path
• The decoupled approach means that ocl-scispacy-loinc now points to a fully specified concept
  • Mapper can show a unified view of a concept in the candidates tab (e.g. in Match Quality view), where there is only a single row that was returned by more than one algo
• Users are mapping to Concepts, not Candidates -- meaning Candidates and Match Metadata are linked to the Mapped Concept, but are not directly part of it
• Re-ranking will be applied to the Concept Pool (both bridge and target concepts) not to the Candidate Pool
• Updated Retrieval workflow:
  • Create Candidate Pool across all algorithms -->
  • Generate Concept Pool consisting of both bridge and target codes (codes need to maintain bridge/target relationships) -->
  • Populate the Concept Pool with full details
  • Re-rank the Concept Pool to get Unified scores for all bridge and target codes

Separate $lookup from $match

• Enables a canonical concept representation without expecting that an algorithm provides this info
• Required when decoupling candidates from concepts
• Mapper should be smart enough to configure $lookup on its own when it can (e.g. user selects ocl-semantic or ocl-search algos)
• User should have the option to configure $lookup manually when they want to
• $match still able to return a full concept definition, but no longer default behavior

Formalize Match Algorithms to be a first-class trackable entity

• Requirements for this are in this ticket: Implement generic match algorithm definition config for OCL Mapper #2301

[paynejd]

Draft of updated schema from @snyaggarwal

[paynejd]

A few comments on the schema:

• Candidate is linked to project algo
• Suggest adding Target Repo
• Concept is linked to both Candidate AND Target Repo
• States need proper definition -- @snyaggarwal please add all state variables to the diagram and all possible coded values in a comment
• Events may need parent-id option, i.e. bundling auto-match sub-events into a single auto-match event? (or do you have a suggestion for how to do this in a different way?)

[snyaggarwal]

@paynejd To unblock the deployment:

• Two native algos -- ocl-search, ocl-semantic must be part of the OS of oclmap
• OCL online API service (new) will have capabilities to provide more algorithms to ocl-online mapper tool