Acute Kidney Injury (AKI) Predictor

A real-time prediction system to detect Acute Kidney Injury (AKI) from clinical temporal changes of creatinine levels using a Machine Learning model.
Trained on 7000+ clinical samples, this ML model can correctly predict the presence of AKI with >96% accuracy.

The System processes real-time Health Level 7 (HL7) messages sent from the Hospital's Clinical Systems (Patient Administration System (PAS) and Laboratory Information Management System (LISM)).
For every new blood test result, the system will run the prediction using the pre-trained model and, if AKI is detected, the system will page the Hospital's Clinical Response team in less than 3 seconds from the test result receipt.

Requirements

The project is written in python. The following third party packages are required to ensure full project functionality:

• pandas: python package for data analysis, particularly suitable for handling relational and labelled data. Actively maintained (Last Updated: September 2024) and open-source.

• scikit-learn: python package for data modeling and machine learning algorithms. Actively maintained (Last Updated: January 2025), open-source and commercially supported.

• joblib: python package for fast and efficient python pipelines. Actively maintained (Last Updated: May 2024) and commercially supported

• hl7: python package for parsing Health Level 7 (HL7) messages.

• requests: python package to send HTTP requests. Actively maintained (Last Updated: May 2024) and commercially supported

All requirements (along with the used versions) can be found in the requirements.txt file.

System Design

The full system design specifications can be found here.

Data

All patient data from the Hospital is stored in a database, with separate tables for Patient Data and Blood Test Results data.
To enable prediction, the input clinical data must include:

• Sex (column 'sex): as 'f' (female) or 'm' (male)
• Age (column 'age'): an Integer
• Dates of blood tests taken (columns 'creatinine_date_x'): in the format 'DD/MM/YY HH:MM:SS'
• Results of each blood test (columns 'creatinine_result_x'): indicates the measured creatinine level, as an Integer

The database is initially populated with historical patient data from a given .csv file.

Example Usage

Clone this repository on your local device. Enter the directory storing the 'main.py' file ('/system/main.py').
Start a terminal at the folder and install necessary requirements:

../bin/pip3 install -r requirements.txt

Terminal Commands

To simulate the client side of the system (without activating the hospital server side):

python3 main.py

To simulate real-time stream of messages from the hospital server side, open another terminal at the directory storing the 'simulator' file ('/simulator/simulator.py') and run:

python3 simulator.py

Important

The server side MUST be activated before the system side, otherwise the connection will not be established!

Docker Commands

Enter the general directory. To build the docker image from the Dockerfile, run:

docker build -t {image_name} .

with image_name being any chosen name for the project.
To run the image:

docker run --net=host {image_name}

Optionally, the input address for the MLLP and Pager system can be specified as:

docker run -e MLLP_ADDRESS="{chosen_mllp_address}" PAGER_ADDRESS='{chosen_pager_address}' {image_name}

with chosen_mllp_address and chosen_pager_address being in the format 'localhost:8440', '0.0.0.0:8440' or similar.

Authors

• Cesare Bergossi (cb624)
• Ginevra Cepparulo (gc1424)
• Belfiore Asia (ab6124)
• Lisa Faloughi (lf524)