Which are some of your favourite places in the world?

ⓘ Start presenting to display the poll results on this slide.

There are no “good” mappings.

How can we make mapping quality more transparent?

OHDSI 2022 Symposium

https://bit.ly/ohdsi-sssom-slides

https://bit.ly/ohdsi-sssom-slido

​

Melissa Haendel, Nicolas Matentzoglu, Chris Mungall, Lisa Eskenazi, Davera Gabriel, Chris Roeder, Bryan Laraway, Siggie Gold, Joe Flack, Stephanie Hong, Tanner Zhang, Anita Walden, Shahim Essaid, Christopher Chute

Interoperability is in the eye

of the beholder

Legal/Licensing

Restrictively licensed data can only be combined with permissively licensed data

Regulatory

Access control must match provenanced regulatory permissions

System

Data

Platforms and tools often cannot talk to one another to move data and analyses

Data is often un-encoded or coded in different data models & terminologies, limiting search and integrated analytics

{ }

Mapping is all over the place, lossy, and lacks provenance

“Mappers”

Source terminologies

Coded Data

Codesets

Uncoded/locally

coded Data

Mapping COVID positivity across sites

Site:

Assessment of synonyms, hierarchies, and mappings across ontology sources for example diseases EDS and pancreatic cancer

http://bit.ly/annual-reviews-ontologies

Wide heterogeneity in:

• placement of diseases hierarchically, and therefore meaning,
• mapping to other diseases,
• the number and typing of synonyms

2018 statistics shown from this manuscript:

Mappings are insufficient

• Proliferation of mappings
• Too many combinations
• Frequently conflicting
• Frequently stale
• Semantics unclear
• Equivalent?
• Exact?
• Broad/Narrow/Related?
• Without precise equivalence mapping, merging is not possible

C1

C2

C3

C4

C6

C5

(N^2)-N sets of mappings (if each source provides their own mappings to all)

Are they the same disease?

Taking into account the classification

https://github.com/monarch-initiative/mondo/issues/61

Orphanet

NCIT

DO

• Different labels
• Different parents
• Different children
• Different synonyms
• Different text definitions

​

Are they equivalent?

Why mapping matters: We need to provide equal diagnostic power around the world

NCIT

DOID

GARD

Orphanet

OMIM

Just 5 sources comprise 10,577 unique rare disease concepts

(prior estimates ~7,500)

​

​

Only 333 shared disease concepts in all five sources

Many diseases are in only one source

Nature Reviews Drug Discovery (bit.ly/nature-rare-diseases)

Intersection size

5 selected sources

Schema alignment: Each source models things differently

For example, no direct link from Sample-to-Diagnosis in one model

Would need to “remodel” Sample-to-Case, and Diagnosis-to-Case to align with Sample-to-Diagnosis

Terminological alignment:

Each source uses different values

​

For example, one node encodes race like this:

• not reported
• white
• american indian or alaska native
• black or african american

​

While another does it like this:

• not allowed to collect
• unknown
• white
• native hawaiian or other pacific islander
• american indian or alaska native
• asian
• other
• black or african american

Addressing the “mapping problem” ⁠— both at the schema and value levels

Schema level

Value level

Example of a schema mapping

​

​

What are entity/field/value mappings?

​

​

“Friedreich's Ataxia”

OMOP:441554

What are entities? From strings to things

“Friedreich's Ataxia”

OMOP:441554

Entities are symbols, such as codes in a terminology, classes in an ontology, enums in a data model, or simply strings in a text field which are intended to refer to a real world thing.

What exactly is the “mapping” part of “entity mapping”?

​

​

“Friedreich's Ataxia”

OMOP:441554

• An association?
• An instruction for replacement?
• A rule/function?

Many different fields of engineering concerned with entity mappings use highly overlapping approaches…

Clinical terminology mapping

​

​

​

are insufficient

• Map codes from one clinical terminology to another
• Applications range from communication and migration of legacy codes to decision support
• There may be different maps between the same pair of code systems to meet different uses cases
• Example: 1:n “maps to” relationships mapping from ICD10CM to SNOMED

Read more about clinical mappings here: https://library.ahima.org/doc?oid=58226

Ontology alignment/matching

​

​

• A very active research area
• Goal: integrating two semantic spaces (ontologies, knowledge graphs) by computing semantic links
• Note: Ontology Matching is a generic activity that is not typically associated with a “purpose”

Ontology 1

Ontology 2

Seed mapping

Matching

Merged/integrated ontology

Value set mapping - Sex at birth example

​

​

Sex at birth mappings from CDM model to OMOP. Need to capture gender identity across the board along with sex at birth

Value set mapping - Race example

ONC race values are limited to White, Black or African American, Asian, American Indian or Alaska Native, or Native Hawaiian or other pacific Islanders. There is data loss, lossiness of specificity.

All mapped to 38003573

Other approaches (not exhaustive, there are many)

​

​

​

are insufficient

• Thesaurus building involves assigning natural language strings (synonym) to a code in a knowledge organisation system (like a taxonomy, terminology, or ontology)
• Named Entity Recognition and Entity Linking involves recognising entities (such as diseases) in text and linking them to some identifier (like OMOP).
• Entity resolution/record linkage involves determining if records from different data sources represent, in fact, the same entity

What kinds of mappings do you care about?

ⓘ Start presenting to display the poll results on this slide.

Let’s look more closely at some examples.

Case Study 1: OMOP Non-standard to Standard Concepts

​

​

• Mapping algorithm
• Interesting challenges:
• Distinguishing exact from broad mappings
• Distinguishing broad mappings from complex mappings
• Risk of n:1 mappings
• The complexity of 1:n mappings

https://www.ohdsi.org/web/wiki/doku.php?id=documentation:vocabulary:mapping

OMOP: Examples

https://www.ohdsi.org/web/wiki/doku.php?id=documentation:vocabulary:mapping

Amebiasis (A06) ---> Amebic infection (438959)

ⓘ Start presenting to display the poll results on this slide.

Refractory anaemia with multi-lineage dysplasia (D46.A)--->Refractory anemia (4003185)

ⓘ Start presenting to display the poll results on this slide.

Unspecified open wound of finger with damage to nail (S61.30) ---> [Open wound of finger (4054063)], [Fingernail injury (4154164)]

ⓘ Start presenting to display the poll results on this slide.

Pedestrian on roller-skates injured in collision with pedal cycle in traffic accident ---> [Pedal cycle accident involving collision between pedal cycle and pedestrian, pedestrian injured], [Victim, pedestrian in vehicular AND/OR traffic accident]

ⓘ Start presenting to display the poll results on this slide.

How important is mapping precision for your use case?

ⓘ Start presenting to display the poll results on this slide.

Case Study 2: OMOP to OBO

​

​

OBO

Disclaimer: Lots of missing links, just for illustration!

Case Study 2: OMOP to OBO

​

​

OBO

Disclaimer: Lots of missing links, just for illustration!

Biomedical knowledge graphs for genomics data

Cross-species knowledge

Clinical data (condition occurrences)

Anatomical reference models

Diseases

Widely used for rare disease, est. 25% coverage in SNOMED

OMOP2OBO - Bridging the translational divide

• Ongoing initiative led by Tiffany Callahan and Andrew Williams: https://covid.cd2h.org/ontology
• https://github.com/callahantiff/OMOP2OBO
• https://doi.org/10.48550/arXiv.2209.04732
• Many related initiatives:
• Mondo to ICD10
• HPO to SNOMED
• HPO to LOINC

OBO Foundry: http://obofoundry.org/

Case Study 3: Registry data to OMOP

Dataset related to KIF1A Associated Neurological Disorder, Critical Path Institute

How do you deal with survey data?

ⓘ Start presenting to display the poll results on this slide.

Mapping are pivotal for healthcare, yet,

there are few “good” mappings. Why?

​

​

• Producing mappings is expensive…
• Figuring out the intended semantics of a code in a terminology is hard work
• Producing mappings requires costly expertise: semantics, data engineering and medical expertise
• …often too expensive
• The scale of the data integration problem mandates the use of automated mapping processes
• The mapping space is N², which is enormous
• Society should not have to afford the incredible duplication of effort we currently have
• Mappings are “fit” or “unfit for purpose” rather than “good” or “bad”...
• The source and target semantics are almost never 100% clear
• The contexts in which codes are used often affect their intended meaning
• …and as a consequence they are very hard to be re-used across use cases

..but we can do better

• Share mappings to avoid creating the same mapping over and over again
• Enrich the mapping with metadata that enable us to determine our level of trust in a mapping for a particular purpose
• Provide “justifications” for our mappings:
• Why have these entities been mapped?
• Make “confidence” explicit
• Document curation rules as part of the metadata
• How recent was the mapping? Could there be a better mapping in a more recent version?
• What was the expertise of the reviewer?
• Use automated matchers more effectively
• We can rarely just “trust” them - but we can ask them to produce justifications

​

What kind of metadata would you like to see for mappings?

ⓘ Start presenting to display the poll results on this slide.

SEMANTIC

* SSSOM can be pronounced: Sesom

https://doi.org/10.1093/database/baac035

https://w3id.org/sssom

The anatomy of a mapping

​

​

​

are insufficient

mapping_justification: semapv:LexicalMatching

Example SSSOM TSV file

​

​

Can be exported to JSON, RDF, etc.

#mapping_set_id: MGI_Full_MP_HPO

#mapping_set_title: All mappings of MP terms to HPO terms generated by MGI

#mapping_set_description: "Consolidated list of all HPO to MP mappings done by MGI…."

#creator_id:

# - orcid:0000-0003-4606-0597

# - orcid:0000-0002-6490-7723

# - orcid:0000-0003-2307-1226

# - ror:021sy4w91

# - wikidata:Q1951035

#license: https://creativecommons.org/licenses/by/4.0/

#object_source: obo:hp

#subject_source: obo:mp

#curie_map:

# HP: http://purl.obolibrary.org/obo/HP_

# MP: http://purl.obolibrary.org/obo/MP_

https://bit.ly/ohdsi-sssom-example

Mapping Table

Licensing information in header

Provenance and descriptions

Dereferencable identifiers

​

​

• Most metadata elements in SSSOM require the use of “entity references” rather than simple strings: source references, preprocessing techniques, etc
• An “entity reference” is a globally unique, persistent (and resolvable) identifier (GUPRI)
• The entity reference itself is usually recorded as a Compact URI, or CURIE, which can be resolved to a URI using a special “curie_map”.

​

MONDO:0035401

http://purl.obolibrary.org/obo/MONDO_0035401

semapv:MappingChaining

https://w3id.org/semapv/vocab/MappingChaining

Work in progress: identifiers for specific curation rules!

Semantic mapping predicates

​

​

Orphanet:563609

MONDO:0035401

ICD10CM:Q00.0

skos:broadMatch

1. Mapping chains

skos:exactMatch

2. Walk-back (symmetric)

3. Walk-back (inverse)

skos:narrowMatch

skos:exactMatch

skos:broadMatch

It does not matter but the concept is (isolated) anencephaly

Tutorial: https://mapping-commons.github.io/sssom/mapping-predicates/

Mapping justifications

:A

:B

skos:exactMatch

Lexical matching

“S”

rdfs:label

skos:prefLabel

subject_match_field

object_match_field

match_string

mapping_justification

Manual mapping curation

mapping_justification

orcid:123

author_id

A mapping can have more than one justification!

confidence

0.7

Other examples of justifications:

• Lexical similarity
• Semantic similarity
• Mapping Chaining

The bigger vision

​

​

mapping

mapping set

mapping server

mapping commons

makes available & trusts

makes available & trusts

maintains

part of

https://fairsharing.org/bsg-s001618

A group of mappings that belong together, e.g. an ontology alignment, a value set mapping, a terminology mapping

An individual entity mapping including justification, connecting two codes/values or entities.

an SSSOM file

A community effort seeking to collect and reconcile mappings from multiple sources.

A suite of tools and services for accessing APIs in a standardised manner

Summary: Why SSSOM?

​

​

• Shared provenance model to enable cross-purpose re-use:
• Who did the mapping, when, for what purpose and versions of which sources?
• Shared model for computable justifications and confidence to increase trust, enable automated conflict resolution and explainable mapping decisions.
• This mapping was established because of a synonym match (70% confidence)
• This mapping was reviewed by a medical professional
• This mapping was established as a consequence of applying curation rule X
• Shared model for semantic predicates to enable cross-walking and semantically consistent interpretation of mappings
• A -[maps to]-> C, because A -[maps to]-> B and B -[maps to]-> C
• The ability to cross-walk drastically drops costs of producing mappings (mitigate n-squared problem)
• Distinguish complex from broad matches (not yet finally defined in SSSOM)
• Development of shared infrastructure for mapping integration, quality control and delivery (mapping servers).

Do you care about re-using existing mappings in your work?

ⓘ Start presenting to display the poll results on this slide.

Step-by-step: a Five-Star system for mappings

​

​

• Standardisation of mappings is costly
• Sometimes, perfect is the enemy of the good enough
• As a community we need to find a good middle ground for what is “good enough”.
• See “handout”

Please lobby all mapping providers out there to publish mappings using CC-0 or CC-BY licenses under a public URL!

Acknowledgements (SSSOM team)

​

​

Funding:

​

Phenomics First (NIH / NHGRI #1RM1HG010860-01): Spec, Mondo integration, sssom-py CLI��Monarch (NIH / OD #5R24OD011883): Cross-species mappings, outreach, knowledge graph integration

​

Bosch Gift to LBNL: sssom-py IO, testing, converters, tutorials�

​

​

​

​

​

​

​

Community contributions: https://w3id.org/sssom

Core team

• Charlie Hoyt (Harvard Medical School)
• Chris Mungall (LBL)
• David Osumi-Sutherland (EMBL-EBI)
• Ernesto Jimenez-Ruiz (City, Univ. of London)
• Harshad Hegde (LBL)
• Henriette Harmse (EMBL-EBI)
• James McLaughlin (EMBL-EBI)
• John Graybeal (Stanford)
• Kim, Hyeongsik (Bosch)
• Melissa Haendel (CU Anschutz)
• Nicolas Matentzoglu (EMBL-EBI)
• Simon Jupp (SciBite)
• Thomas Liener (Pistoia Alliance)
• Tiffany Callahan
• William Duncan (LBL)
• ...many more contributors, see publication

Database (Oxford), Volume 2022, baac035, https://doi.org/10.1093/database/baac035

Do you care about Open Science?

ⓘ Start presenting to display the poll results on this slide.

Exercise Scenario

• You are a data steward of a research company for rare disease that seeks to integrate public genomics data from a major research institute with your local patient data (OMOP CDM).
• The research institute uses OBO ontologies to annotate their data - http://obofoundry.org/.
• The goal is to integrate known gene associations with specific conditions, as well as affected anatomical entities and diagnostic procedures that were recorded as part of the metadata.
• As a proof of concept, you are tasked to map 10 OBO ontology concepts to OMOP standard concepts.

Exercise instructions

1. Access the exercise skeleton at (https://bit.ly/ohdsi-sssom-exercise or QR code) and copy the data to your table editor of choice.
2. Fill in the existing metadata elements in the “Mapping set” sheet as well as you can using Athena and OLS.
3. Familiarise yourself with https://mapping-commons.github.io/sssom/ to get a sense of existing metadata elements in SSSOM
4. Enrich your “Mapping set” with as much metadata as you find useful.
5. Maintain a critical eye. What is missing? Which elements are not clear to you? Feed this information back to the team.
6. Fill in the “Mapping set metadata” sheet.
7. Enrich your “Mapping set metadata” with as much information as you find useful.

The SSSOM metadata model

Rich YAML schema powered by

Shex shapes for validating rdf

JSON Schema

Markdown docs

- subject_id

- subject_label

- subject_category

- predicate_id

- predicate_label

- object_id

- object_label

- object_category

- match_type

- creator_id

- creator_label

- license

- subject_source

- subject_source_version

- object_source

- object_source_version

- mapping_provider

- mapping_cardinality

- mapping_tool

- mapping_date

- confidence

- subject_match_field

- object_match_field

- match_string

- subject_preprocessing

- object_preprocessing

- match_term_type

- semantic_similarity_score

- see_also

- other

- comment

https://github.com/mapping-commons/sssom-py/tree/master/schema

Value Set Mapping- Unit of measurement mappings

Units are captured in string format in EHR system and the text used for unit varies from site to site.

Laboratory result categorical answer codes are not always in LOINC code

​

​

�Answer to a laboratory result may contain

text instead of the LOINC coded answers.

�Example: Below assay limitation due to possible

cross reactivity with other HLA-B antigens.

Must be determined by HLA-B Hr-SSP

Molecular Typing Method

​

​

​

Value Set Mapping - EHR Encounter type example

Visit types