EMPOWER

This is the server-side part of the Empower platform.

Requirements

• Python3
• Pip3 to install dependencies
• A MongoDB server with a clean collection (No empty fields or documents)

Installation

If running this on a non-dedicated system, consider using a venv for installation and usage of this project.

sudo apt install python3 pip3
pip3 install scikit-learn=0.21.2 flask=1.0.3 pymongo=3.8.0 pandas=0.24.2 joblib=0.13.2 matplotlib=3.1.0 pandas_profiling=1.4.2 numpy=1.16.4 odo=0.5.0 
git clone [email protected]:patientempowerment/patientEmpowerment.git

Usage

The main functionalities are placed in the rest_server directory:

• empower_cli.py, giving access to the functions through the command-line, and
• empower_server.py, providing a REST-API for access through the web.

In the useful_scripts directory you can find some data cleaning utilities used in the project.

CLI

python3 empower_cli.py --help

In short:

• --database, --collection, and --labels are mandatory
• --database expects the name of a mongo db
• --collection is a collection in that db
• --labels expects names of columns to predict on, like "something somethingElse"
• Label-Columns' fields should contain exclusively 0 or 1

For example, if you want to predict if people have dogs, cats, and/or birds. You have a mongoDB called Citizens12, which has a collection called livingSituation. You would call:

python3 empower_cli.py -db Citizens12 -c livingSituation -l "has_cat has_dog has_bird"

Optionals are --output and --correlator.

--output generates a confusion matrix in /tmp/, where you can find a visual classification report.

--correlator superimposes the values of a column onto the graphs that are generated at each execution. For example, if you wanted visualize not only pet-distribution, but also get an idea for correlating house-size you could call -corr house-size-in-sqm.

Flask-Server

Runs a flask server offering a REST-API for the Empower application. All requests expect a JSON document.

Start the server by calling

python3 empower_server.py

The API currently provides 4 endpoints:

/predict (POST):

Allows for server-side prediction of different labels. Currently, Logistic Regression is used as a predictor. All fields used by the model during training need to be in the request body.

The request body should be formatted like this:

{
    "user_data": {
        <feature1>: <value1>,
        <feature2>: <value2>,
        ...
    },
    "labels": [
        <label1>,
        <label2>,
        ...
    ]
}

The server's response will look like this:

{
    <label1>: <prediction1>,
    <label2>: <prediction2>,
    ...
}

/models (POST):

Asks the server for its precomputed models to allow for on-device predictions. Currently, we use logistical regression, for which only

The request body should be formatted like this:

{
    "labels": [
        <label1>,
        <label2>,
        ...
    ]
}

The server's response will look like this:

{
    <label1>: {
        "features": {
            <feature1>: {
                "coef": <feature1-coef>,
                "mean": <feature1-mean>
            },
            <feature2>: {
                "coef": <feature2-coef>,
                "mean": <feature2-mean>
            },
            ...
        },
        "intercept": <label1-intercept>
    }
    ...
}

/retrain (POST):

Asks the server to retrain its models on a given database and collection and for a given set of labels contained in the collection.

The request body should be formatted like this:

{
    "db": <mongo db name>,
    "collection": <mongo collection name>,
    "labels": [
        <label1>,
        <label2>,
        ...
    ]
}

The server's response will look like this:

{
    <label1>: <prediction1>,
    <label2>: <prediction2>,
    ...
}

feature-config (POST):

Requests a configuration JSON for the user-facing data input fields. This JSON gets auto-generated from the fields in the collection specified by the user. By default, every field with values between 0 and 1 becomes a 'choice' field, represented by radio buttons in the application, every other field becomes a slider.

The request should be formatted like this, specifying the DB and collection the user wants the configuration JSON for:

{
  "db": <mongo db name>,
  "collection": <mongo collection name>
}

The response will look as follows, with <featureX> being the name of the feature in the collection and <featureX_title> being the human readable name to be displayed on the GUI of the requesting application.

{
    <feature1>: {
        "choices": {
            "No": 0,
            "Yes": 1
        },
        "title": <feature1_title>,
        "mean": <float>
    },
    <feature2>: {
        "slider_max": <value>,
        "slider_min": <value>,
        "title": <feature2_title>,
        "mean": <float>
    },
    ...
}

Contributing

License

MIT