giovabaj / ecg-cnn-xgb-lr

Code for the paper "Comparison of discrimination and calibration performance of ECG-based machine learning models for prediction of new-onset atrial fibrillation"

Comparison of discrimination and calibration performance of ECG-based machine learning models for prediction of new-onset atrial fibrillation

Python code for the paper "Comparison of discrimination and calibration performance of ECG-based machine learning models for prediction of new-onset atrial fibrillation ", Baj Giovanni et al., BMC Medical Research Methodology (2023).

Data organization

Data has to be organized as follows:

• information about ECGs has to be saved in a csv file. Necessary columns are:
  • ExamID key that uniquely identifies the ECG
  • KEY_ANAGRAFE patient's registry key
  • FA binary variable that indicates if the patient developed atrial fibrillation within 5 years after the exam
• ECG signals have to be stored in a single directory as numpy files, and named with their key (ExamID in our case). Each ECG is a 12-lead signal sampled at 1000 Hz for 10s.

Scripts

run_all.py runs all the experiments reported in the paper (dependence on the sample size and dependence on the balancing ratio).

plot_results.py loads the results from previous computations and makes the plots reported in the paper. To use this script one has to create a json file where directories whose results have to be analyzed are listed (for more details look at the code).

make_dataset.py creates a single dataset that ca be used to train the models. It takes as input the number of ECGs to be included in the dataset, the path of the csv file where ECGs information is stored, and the path where ECG signals are stored. Once the dataset is created, the script saves in the "datasets" directory a compressed numpy file (.npz) that contains two numpy arrays: a matrix of shape (N, 12, 5000) with the ECG signals (where N is the number of ECGs), and a second array with the corresponding exam IDs.

cnn_training.py trains CNN models in a cross-validation setting, for a specified configuration

xgb_lr_training.py trains XGB or LR models with the same dataset and cross-validation folds of a previously trained CNN model.

hyperparameters_search.py performs hyperparameters search for XGB and LR models

xgb_lr_training.py trains XGB or LR model