outofray/kardionet

Kardio-Net: A Deep Learning Model for Predicting Serum Potassium Levels from Apple Watch ECG in ESRD Patients

I. Input preparation

1. Prerequisites

   Ensure that all ECG files are saved in *.npy format with the following shapes:

   • (5000, 12) for 12-lead ECGs
   • (500 * 30 seconds, 1) for Apple Watch ECGs

   Store these files in a single flat directory, and include a manifest CSV file in the same location to accompany them.

2. Manifest File Format

   The manifest CSV file should include a header. Each row corresponds to one ECG file with the following columns:

   • filename: Name of the .npy file (without the extension).
   • label: serum potassium label.

II. Inference

For 12-Lead ECG

1. use predict_potassium_12lead.py to get the potassium level prediction.

2. Edit data_path and manifest_path and run the predict_potassium.py script.

3. Upon completion, a file named "dataloader_0_predictions.csv" will be saved in the same directory. This file contains the inference results "preds" from the model.

4. Use generate_result.py to get the performance metric and figure.

For Single Lead / Apple Watch ECG

A. ECG preprocessing and segmentation

1. Use preprocessing.py for denoise, normalize, and segment ECG into 5-second for input

2. Set the following paths in preprocessing.py:

   • raw_directory = "path/to/raw_data_directory" #raw ecg folder for target task
   • output_directory = "path/to/output_directory" #output folder for normalize ECG
   • manifest_path = "/path/to/manifest.csv" # Manifest file path
   • output_path = "path/to/output_path" # Output path for segmented ECGs

3. Execute predict.py.

4. If the ECG files were already normalize, can execute the segmentation function only.

B. potassium regression model

1. use predict_potassium_1lead.py to get the potassium level prediction.

2. Edit data_path and manifest_path and run the predict_potassium.py script.

3. Upon completion, a file named "dataloader_0_predictions.csv" will be saved in the same directory. This file contains the inference results "preds" from the model.

4. Use generate_result.py to get the performance metric and figure.

Paper Link: https://www.sciencedirect.com/science/article/pii/S2405500X24007527

Citation: I-Min Chiu, Po-Jung Wu, Huan Zhang, J. Weston Hughes, Albert J. Rogers, Laleh Jalilian, Marco Perez, Chun-Hung Richard Lin, Chien-Te Lee, James Zou, David Ouyang, Serum Potassium Monitoring Using AI-Enabled Smartwatch Electrocardiograms, JACC: Clinical Electrophysiology, 2024, ISSN 2405-500X, https://doi.org/10.1016/j.jacep.2024.07.023.