Skip to content

Setup a new AlgoMarker

AlgoMarker Configuration Guide

A sample packaging script can be found in GastroFlag:
https://github.com/Medial-EarlySign/AM_LGI/blob/main/Extended_model/scripts/pack_algomarker.sh

The setup script defines variables for the model, repository, reference matrix, deployment path, and more.

Typical Steps

  1. Model Adjustments (Optional but Recommended)

    • Uses tools Flow, adjust_model, and change_model operations to modify the model as needed. For example:
      • Flow --export_production_model: Enables verbose outlier reporting (useful for production, but may slow down predictions and increase memory usage).
      • adjust_model: Perform operations such as adding virtual signals (e.g., "eGFR") or adding checks (e.g., ensuring a CBC exists on the prediction date and storing the result as an attribute for eligibility filtering).
      • change_model: Enable explainability output for AlgoMarker by setting store_as_json=1.
    • These steps are optional and depend on your production requirements. If you do not need outlier documentation or explainability, you may skip them.

  2. Copy Model Files

    • Copy the binary model file to the AlgoMarker directory.
    • Document the original model path and its MD5 checksum (recommended).

  3. Build and Copy the Shared Library

    • Build the libdyn_AlgoMarker.so shared library as described in AlgoMarker Library Setup.
    • Copy the compiled library to the AlgoMarker folder under "lib" folder.

  4. Prepare the Repository Configuration

    • Generate a repository config file containing only the required signals and dictionaries. This defines the signal names, types, and categorical mappings for AlgoMarker.
    • You can copy the config files from the repository used during model training, but it is recommended to clean them up and remove unused signals. A script is available to assist with this from GastroFlag.

  5. Create the AlgoMarker Configuration File (amconfig)

    • See the Configuration File section for details and a full example.
    • Keep TYPE as MEDIAL_INFRA
    • NAME is free text to store name for AlgoMarker and will be returned in discovery call request.
    • The FILTER rows allow you to define eligibility rules and warnings. For testing, you may skip these filters.

    Example snippet: 

    1
2
3
4
5
6
7
8
9
    #################################################################################
# MedialInfra AlgoMarker config example file
#################################################################################
TYPE    MEDIAL_INFRA
NAME    PRE2D
REPOSITORY  rep/pre2d.repository
TIME_UNIT   Date
MODEL   resources/pre2d.model
# Eligibility filters and other settings follow...

    • Filters can enforce rules such as minimum/maximum number of results, allowed values, age ranges, dictionary validation, and outlier limits.
    • See the full example in Eligibility rules configuration section
    • REPOSITORY and MODEL refer to file paths: REPOSITORY is the path to the repository config file (used in Step 4), and MODEL is the path to the model binary (used in Step 2).

    It's recommended to use relative paths and place both files in the same folder hierarchy for simplicity.

  6. Generate Explainability Config (Optional)

    • If your model supports explainability, create a configuration file (e.g., resources/explainer.cfg) to control display settings for each group, such as which signals to show and how many values to fetch.

  7. Create Leaflet/Cutoff Threshold Table (Optional)

    • Generate a cutoff threshold table config file (e.g., from bootstrap_app output) if you want to store thresholds in the AlgoMarker.

  8. Copy Test Script Template (Optional)

    • Optionally, include a script template for testing AlgoMarker deployments.