Enhanced feature discussion - Atlas/WebAPI cohort characterization with custom features

[rkboyce]

This "issue" is a placeholder for further discussion on the direction that Atlas/WebAPI should go for providing users with an enhanced ability to create and use custom features within cohort characterization and other study designs.

Current functionality

• Within Atlas users can create a custom feature using multiple approaches:

  • Using CRITERIA - a powerful approach using cohort-building UI forms. Some documentation in the Book of OHDSI but not allot of details
    • can generate count and averaged data over simple cohort selections or multiple sub-group analyses
  • Using CUSTOM - SQL created within Atlas using SQL
    • uses a parameterized SQL with few pre-set parameters
    • There has to be a valid concept_id for the new feature which can be 0
  • In either case, counts are aggregated and stored in the cc_results table. Row-level feature results are only stored during execution and then deleted.
    • Metadata in cc_analysis, fe_analysis, cc_cohort, cohort_characterization
    • row level data is only stored temporarily and then deleted
    • This means that carefully Atlas constructed features are not presently accessible within R e.g., using Feature Extraction

• Users can also use the R Feature Extractor approach

  • some documentation in in the Book of OHDSI
  • multiple approaches

    • Entirely custom

      • Powerful but difficult to write
      • Produces an R Adromeda object that wraps a sqllite database (on disk) using RSQLite

    • No use of Atlas unless the custom query SQL is using the cohort or other results tables

    • Useful for row-level or summary stats

    • Feature Extraction using Cohort definitions

      • Easier to write and uses Atlas/WebAPI cohorts and produces row-level or aggregated data

    • Feature Extraction using Cohort Attributes

      • Broken : no cohort_attributes exists in the CDM as of >= v5.3

• In any case, the output of an Andromeda object could potentially be used to apply functions from other packages (e.g., a risk scoring function) and the results either updated in the Andromeda object or used to create a new object

  • this implies a potential approach where Atlas/WebAPI is used to create cohorts and then to design a workflow to arrive at final feature object that is then used in other studies
    • e.g., population estimation or patient-level prediction
    • Note: there is no way to go from an Andromeda object back into Atlas/WebAPI
      • for example, to provide additional features for a cohort characterization, incidence rate or cohort pathways

Desired functionality

• User friendly integration between Atlas/WebAPI and R methods for creating and using/re-using custom features
  • Atlas/WebAPI custom features --> R study package
  • R Feature Extraction custom features --> Atlas/WebAPI analysis and specifications (e.g., Estimation or PLP)
  • Easier integration of non-OHDSI R packages that make for useful features
    • e.g., risk scores
• Better documentation and vignettes
  • Book of OHDSI is scant on details for Atlas custom features
  • the cohort_attribute method is deprecated but still promoted on github
  • Writing CUSTOM in Atlas requires some skill and could probably be assisted by AI
    • similarly for the fully custom R Feature Extraction method
• Revised architecture for synergy with new changes happing in the community
  • Use of Strategus
  • Alternative cohort and study characterization approaches

These are some thoughts to get the conversation going.