Configure MPI module

This page explains how to configure the Aidbox MPI module by adding a matching model. It includes an example model for testing, and notes on performance and model tuning.

The current implementation only supports the Patient resource. If you need support for additional resources, please contact us.

The example in the next section provides a basic model that allows you to start the MPI module and test its functionality. For a detailed explanation of all model elements and matching logic, see Matching Model Explanation.

Add model to Aidbox

To add a matching model, you need to create a custom resource named AidboxLinkageModel.

Example of creating an AidboxLinkageModel:

POST /fhir/AidboxLinkageModel
content-type: application/json
accept: application/json

{
  "id" : "model",
  "resourceType" : "AidboxLinkageModel",
  "resource" : "Patient",
  "thresholds" : {
    "auto" : 25,
    "manual" : 16
  },
  "vars" : {
    "family" : "(#.resource#>>'{name,0,family}')",
    "given" : "(#.resource#>>'{name,0,given,0}')",
    "name" : "#.#family || ' ' || #.#given",
    "gender" : "(#.resource#>>'{gender}')",
    "dob" : "(#.resource#>>'{birthDate}')",
    "address" : "(#.resource#>>'{address,0,text}')",
    "addressQuery" : "array(select jsonb_array_elements_text(jsonb_path_query_array( #.resource, '$.telecom[*] ? (@.value != \"\").value')))"
  },
  "features" : {
    "fn" : [ {
      "expr" : " ( l.resource->'name' IS NULL OR r.resource->'name' IS NULL )",
      "bf" : 0.0
    }, {
      "expr" : "l.#name = r.#name",
      "bf" : 13.336495228175629
    }, {
      "expr" : "r.#given = l.#family AND l.#given = r.#family",
      "bf" : 13.104401641242227
    }, {
      "expr" : "levenshtein(l.#name, r.#name) <= 2",
      "bf" : 9.288385498954133
    }, {
      "expr" : "r.#given = l.#given AND string_to_array(l.#family, ' ') && string_to_array(r.#family, ' ')",
      "bf" : 10.36329167966839
    }, {
      "expr" : "r.#family = l.#family AND string_to_array(l.#given, ' ') && string_to_array(r.#given, ' ')",
      "bf" : 10.36329167966839
    }, {
      "expr" : "r.#given = l.#given",
      "bf" : 2.402276401131933
    }, {
      "else" : -12.37233293924643
    } ],
    "dob" : [ {
      "expr" : " ( l.#dob  IS NULL OR r.#dob IS NULL )",
      "bf" : 0.0
    }, {
      "expr" : "l.#dob = r.#dob",
      "bf" : 10.59415069916466
    }, {
      "expr" : "levenshtein(l.#dob, r.#dob) <= 1",
      "bf" : 3.9911610470417744
    }, {
      "expr" : "levenshtein(l.#dob, r.#dob) <= 2",
      "bf" : 0.5164298695732575
    }, {
      "else" : -10.322063538772698
    } ],
    "sex" : [ {
      "expr" : " ( l.#gender IS NULL OR r.#gender IS NULL )",
      "bf" : 0.0
    }, {
      "expr" : " l.#gender = r.#gender",
      "bf" : 1.8504082299552485
    }, {
      "else" : -4.842034404727677
    } ],
    "ext" : [ {
      "expr" : "l.#address = r.#address",
      "bf" : 7.465648574292063
    }, {
      "expr" : "l.#addressQuery  && r.#addressQuery",
      "bf" : 9.236771286242664
    }, {
      "else" : -10.517360697819983
    } ]
  },
  "blocks" : {
    "fn" : {
      "var" : "name"
    },
    "dob" : {
      "var" : "dob"
    },
    "addr" : {
      "sql" : "l.#address = r.#address"
    }
  }
}

Matching Model Tuning

The example model is intended for testing and demonstration purposes and may not deliver optimal results out of the box.

For production use and reliable, accurate matching on your data, you should:

• Adapt the model to reflect your data specifics and your definition of a correct match.

• Calibrate feature weights using your real-world data. This step typically involves machine learning and manual expert tuning.

We offer a professional service for model training and expert tuning. If you need assistance, please contact us.

Performance considerations

For fast and accurate matching, consider the following:

• Database indexes: If you are working with large volumes of patient records, ensure proper database indexes are created to keep matching fast and scalable.

• Data normalization: Matching quality depends heavily on well‑normalized input data. Avoid using placeholders like "UNKNOWN" or "not provided" for names, addresses, or birthdates, as they negatively impact results.