Frequently Asked Questions (FAQ)

Using Seismometer without an events dataset

In case your predictions dataset already includes a specified target, you can avoid merging in any events data and still leverage the model performance evaluations available in seismometer.

Update the usage_config.yml’s data_usage section, removing any lines that reference the events or event_table definitions. Also validate that the primary_target matches the name of the target column in your dataset.

# usage_config.yml
data_usage:
   entity_id: patient_nbr
   context_id: encounter_id
   primary_output: LGBM_score
   primary_target: Readmitted within 30 Days
   predict_time: ScoringTime
   features:
      - admission_type_id
      - num_medications
      - num_procedures
   cohorts:
      - source: age
        display_name: Age
      - source: race
        display_name: Race
      - source: gender
        display_name: Gender
   event_table:
      type: Type
      time: EventTime
      value: Value
   events:
      - source: TargetLabel
        display_name: Readmitted within 30 Days
        window_hr: 6
        offset_hr: 0
        usage: target
   censor_min_count: 10

Then, adjust your config.yml to remove any reference to an events dataset.

# config.yml
other_info:
   usage_config: "usage_config.yml"
   template: "binary"
   info_dir: "outputs"
   event_definition: "dictionary.yml"
   prediction_definition: "dictionary.yml"
   data_dir: "data"
   event_path: "events.parquet"
   prediction_path: "predictions.parquet"
   metadata_path: "metadata.json"

Note

You can still use an events dataset in combination with a target in your predictions dataset. Don’t tag any events in the data_usage configuration with usage: target and everything else should work as intended.

See also

Create Configuration Files

Adding risk groups as a cohort

The following configuration file demonstrates an example of adding risk groups as a possible cohort for analysis. Make sure your cohort defines a display_name for the new cohort that is different from the underlying score column, otherwise you will replace the existing column in the loaded dataframe.

# usage_config.yml
data_usage:
  # other configured data usage ...
  cohorts:
    # other configured cohorts ...
    - source: Score
      display_name: Risk Group
      splits:
        - 0.08
        - 0.15
        - 0.30

Viewing feature reports in Visual Studio Code

By default, Visual Studio Code does not come with a way to view rendered HTML. To support this, navigate to the Visual Studio Code extension marketplace and ensure you have the Live Preview extension installed and enabled:

With the extension enabled, you can right-click on HTML files in the Visual Studio Code file explorer and select Show Preview to view the file as rendered HTML.

Selectors don’t show in Visual Studio Code

Occasionally, using notebooks in Visual Studio Code will result in broken selectors, widgets, or visualizations.

Visual Studio Code relies on support files from a third party website that need additional configuration.

First, navigate to the Visual Studio Code extension marketplace and ensure you have the Jupyter extension installed and enabled:

Then, open your Visual Studio Code extension settings (Ctrl+, on Windows or Cmd+, on Mac), search for jupyter.widgetScriptSources, and add jsdelivr.com and unpkg.com to the list.

See also

Visual Studio Code’s support for IPyWidgets