Maternal Health Risk Predictor

Introduction

According to the World Health Organization (WHO):

“Maternal health refers to the health of women during pregnancy, childbirth and the post-natal period. Each stage should be a positive experience, ensuring women and their babies reach their full potential for health and well-being. Although important progress has been made in the last two decades, about 295 000 women died during and following pregnancy and childbirth in 2017. This number is unacceptably high. The most common direct causes of maternal injury and death are excessive blood loss, infection, high blood pressure, unsafe abortion, and obstructed labour, as well as indirect causes such as anemia, malaria, and heart disease. Most maternal deaths are preventable with timely management by a skilled health professional working in a supportive environment. Ending preventable maternal death must remain at the top of the global agenda. At the same time, simply surviving pregnancy and childbirth can never be the marker of successful maternal health care. It is critical to expand efforts reducing maternal injury and disability to promote health and well-being. Every pregnancy and birth is unique. Addressing inequalities that affect health outcomes, especially sexual and reproductive health and rights and gender, is fundamental to ensuring all women have access to respectful and high-quality maternity care.”

The goal of the project is to apply what has been learned during the MLOps Zoomcamp course to build a MLOps pipeline for woman health risk prediction during pregnancy.

Dataset

The dataset used to feed the MLOps pipeline has been downloaded from Kaggle and contains data collected from several hospitals, community clinics and maternal health cares through an IoT-based risk monitoring system. The dataset is updated daily and is characterized by the following features:

MLOps pipeline

Architecture

Deployment

The MLOps pipeline is fully dockerised and can be easily deployed via the following steps:

1. Clone the maternal-health-risk repository locally:

   $ git clone https://github.com/Peco602/maternal-health-risk.git
   

2. Install the pre-requisites necessary to run the pipeline:

   $ cd maternal-health-risk
   $ sudo apt install make
   $ make prerequisites
   

   It is also suggested to add the current user to the docker group to avoid running the next steps as sudo:

   $ sudo groupadd docker
   $ sudo usermod -aG docker $USER
   

   then, logout and log back in so that the group membership is re-evaluated.

3. [Optional] Configure the development evironment:

   $ make setup
   

   This is required to perform further development and testing on the pipeline.

4. [Optional] Insert Kaggle credentials in the .env file to allow the automatic scheduled dataset update:

   # Kaggle credentials
   KAGGLE_USERNAME=*****
   KAGGLE_KEY=*****
   

   In case the credentials are not available, the training dataset data/data.csv must be updated manually.

5. Pull the Docker images:

   $ make pull
   

6. Launch the MLOps pipeline:

   $ make run
   

   Once ready, the following services will be available:

   Service	Port	Interface	Description
   Web Application	80	0.0.0.0	Prediction web service (see picture below)
   Prefect	4200	127.0.0.1	Training workflow orchestration
   MLFlow	5000	127.0.0.1	Experiment tracking and model registry
   MinIO	9001	127.0.0.1	S3-equivalent bucket management
   Evidently	8085	127.0.0.1	Data and target drift report generation (/dashboard route)
   Grafana	3000	127.0.0.1	Data and target drift real-time dashboards

   

Training

Once the MLOps pipeline has been started, the prediction web service can already work thanks to a default pre-trained model available in the Docker image. In order to enable pipeline training workflow it is necessary to create a scheduled Prefect deployment via:

$ make deployment

The training workflow will be then automatically executed every day. It will download the latest dataset (if the Kaggle credentials have been provided), search the best model in terms of accuracy among XGBoost, Support Vector Machine and Random Forest and finally will store it in the model registry. It is worth noting the training workflow can also be immediately executed without waiting the next schedule:

$ make train

Once the updated model is ready, it can be moved to production by restarting the pipeline:

$ make restart

the web service will automatically connect to the registry and get the most updated model. If the model is still not available, it will continue to use the default one.

Monitoring

It is possible to generate simulated traffic via:

$ make generate-traffic

Then, the prediction service can be monitored via:

• Grafana (in real-time): http://127.0.0.1:3000
• Evidently (for report generation): http://127.0.0.1:8085/dashboard

Disposal

The MLOps pipeline can be disposed via:

$ make kill

while the Docker volumes used for persistence can be removed via:

$ make clean

GitHub Actions

• Continuous Integration: On every push and pull request on main and dev branches, the Docker images are built, tested and then pushed to DockerHub.
• Continuous Deployment: On every push and pull request on main branch, only if the Continuous Integration workflow has been successful successful, the updated pipeline is deployed to the target server and run.

Applied technologies

Disclaimer

This prediction service has been developed as the final project of the MLOps Zoomcamp course from DataTalks.Club. It does not provide medical advice and it is intended for informational purposes only. It cannot be considered a substitute for professional medical advice, diagnosis or treatment. Never ignore professional medical advice in seeking treatment because of something you have read here.