Wrapping a FHIR API Around a BERT Classification Model

In the world of healthcare and medicine, the ability to quickly and accurately classify queries as medical questions is an invaluable tool. As Natural Language Processing (NLP) and machine learning technology continue to develop, these tasks are becoming increasingly more attainable. This blog post will explain how to wrap a Fast Healthcare Interoperability Resources (FHIR) API around a BERT Classification model to answer the question, "Is this a medical question?"

To make this possible, we will be using FHIR Questionnaire and QuestionnaireResponse Resources, and the Bi-directional Encoder Representations from Transformers (BERT) Classification model. The FHIR standard allows for streamlined healthcare data exchange, while the BERT model, a transformer-based machine learning technique for NLP tasks, will provide us with the ability to effectively classify queries as medical questions.

The Questionnaire Resource

The FHIR Questionnaire resource defines metadata about a document, such as a structured set of questions intended to guide the collection of answers. Below is an example of the Questionnaire resource we will be using:

jsonCopy code{
    "id": "is-medical-question-questionnaire",
    "item": [
        {
            "answerOption": [
                {
                    "valueCoding": {
                        "code": "(0)",
                        "display": "This is not a medical question.",
                        "system": "http://fly.health/medicalquestions"
                    }
                },
                {
                    "valueCoding": {
                        "code": "(1)",
                        "display": "This is a medical question.",
                        "system": "http://fly.health/medicalquestions"
                    }
                }
            ],
            "linkId": "is-medical-question",
            "text": "Is this a medical question?",
            "type": "choice"
        },
        {
            "linkId": "user-question",
            "text": "What is your question?",
            "type": "string"
        }
    ],
    "name": "Medical Question Questionnaire",
    "resourceType": "Questionnaire",
    "status": "active",
    "title": "Medical Question Questionnaire"
}

This Questionnaire presents the user with two questions: "Is this a medical question?" and "What is your question?" and it provides two answer options for the first question: "This is not a medical question" and "This is a medical question."

The FHIR API

Our API is invoked by a POST request to /QuestionnaireResponse with the following payload in the body:

jsonCopy code{
    "resourceType": "QuestionnaireResponse",
    "questionnaire": "Questionnaire/is-medical-question-questionnaire",
    "status": "in-progress",
    "item": [
        {
            "linkId": "user-question",
            "text": "What is your question?",
            "answer": [
                {
                    "valueString": "Whats a good care plan for Diabetes?"
                }
            ]
        }
    ]
}

The payload includes the user's question and indicates that the status of the questionnaire is "in-progress".

Wrapping the BERT Model

Upon receiving a request, the API will invoke our BERT Classification model. This model will classify the question as a medical or non-medical question. The BERT model has been trained on a dataset of medical and non-medical questions, and it is highly accurate at distinguishing between the two types.

The API Response

Once the BERT model has made its classification, the API will return a 201 Created response, along with the FHIR QuestionnaireResponse. This will look something like this:

jsonCopy code{
    "meta": {
        "profile": [
            "http://hl7.org/fhir/us/core/StructureDefinition/us-core-questionnaireresponse"
        ]
    },
    "resourceType": "QuestionnaireResponse",
    "questionnaire": "Questionnaire/is-medical-question-questionnaire",
    "status": "completed",
    "subject": {
        "reference": "Patient/Unknown",
        "display": "Unknown"
    },
    "item": [
        {
            "linkId": "user-question",
            "text": "What is your question?",
            "answer": [
                {
                    "valueString": "Whats a good care plan for Diabetes?"
                }
            ]
        },
        {
            "linkId": "is-medical-question",
            "text": "Is this a medical question?",
            "answer": [
                {
                    "valueCoding": {
                        "system": "http://fly.health/medicalanswers/answer-codes",
                        "code": 1,
                        "display": "This is a medical question."
                    }
                }
            ]
        }
    ]
}

The returned QuestionnaireResponse confirms the user's question and provides the answer to "Is this a medical question?".

Conclusion

By wrapping a FHIR API around a BERT Classification model, we can utilize the power of machine learning and the interoperability of FHIR to accurately classify queries as medical questions. This system could have numerous applications in healthcare, from triaging patient questions to aiding research tasks. It demonstrates the exciting possibilities that emerge when we combine advanced NLP techniques with healthcare-focused data standards.