Chagas disease is a parasitic illness caused by the protozoan Trypanosoma cruzi, primarily induced by triatomine bugs. It is endemic to Central and South America but can also be found in other parts of the world through migration. The disease begins acutely with fever, fatigue, and swelling near the infection site. However, about 20–30% of infected individuals may develop chronic Chagas cardiomyopathies and require intensive care. Chagas disease has ben hypothesized to be detectable on a 12-lead ECG, and a fast deep learning model has the potential to promote widespread preliminary testing, efficient treatment, and vector control.
In response to the George B. Moody Physionet Challenge for 2025, this is our attempt to create a diagnostic model that classifies Chagas disease from 12 lead ECG data using a convolutional neural network.
We did this as final project for our CSCI 1470 at Brown university. Please see the poster, write up, and some technical details below. Special thanks to Professor Eric Ewing for a great semester!
For those interested, please take a look at an in depth write up of our project linked here
Here is a link to our second check in write-up
Here is a link to our third check in write-up
In accordance with the rules of physionet, we used the CODE-15% dataset, here are some instructions for downloading and preprocessing the data
These instructions use code15_input as the path for the input data files and code15_output for the output data files, but you can replace them with the absolute or relative paths for the files on your machine.
-
Download and unzip one or more of the
exam_partfiles and theexams.csvfile in the CODE-15% dataset. -
Download and unzip the Chagas labels, i.e., the
code15_chagas_labels.csvfile. -
Convert the CODE-15% dataset to WFDB format, with the available demographics information and Chagas labels in the WFDB header file, by running python prepare_code15_data.py
-i code15_input/exams_part0.hdf5 code15_input/exams_part1.hdf5
-d code15_input/exams.csv
-l code15_input/code15_chagas_labels.csv
-o code15_output/exams_part0 code15_output/exams_part1
Each exam_part file in the CODE-15% dataset contains approximately 20,000 ECG recordings. You can include more or fewer of these files to increase or decrease the number of ECG recordings, respectively. You may want to start with fewer ECG recordings to debug your code.
After downloading and preprocessing the data we set up a data file structure with a test_data folder and train_data folder in our working directory. We partitioned the data manually into a 75:25 train test split in these folders. As a result of the size of our dataset, we chose to do this for considerations of local memory. With more compute, it would be more ideal to randomize the train/test split every time to obtain more representative results.
Intrigued? Please email stephen_c_yang@brown.edu, brandon_lien@brown.edu, mason_usher@brown.edu, and/or manan_pancholy@brown.edu with any questions or comments :)