Master thesis - Feasibility of deep learning-based methods for the detection of simulated ECGs

Repository contains the code to train and test the models developed during the master thesis. The thesis report can be accessed via the xAI chairs webpage.

Structure:

.
├── classification              # Files describing the models, how they where trained and tested
│   ├── CNN-LSTM                # All files regarding the CNN-LSTM model
│   └── CNN-Transformer         # All files regarding the CNN-Transformer model
├── generative_approach         # Contains all models, loss functions, training and demo scripts regarding the generative approach

Reqired Packages:

• Defibrillator-Preprocessing (https://github.com/Detecting-Simulated-Real-ECG-Signals/preprocessing)
• LocalDataManagementTool (https://github.com/Detecting-Simulated-Real-ECG-Signals/localdatamanagementtool)
• jupyterlab
• scikit
• sklearn
• torch (https://pytorch.org/get-started/locally/)
• torchmetrics
• tqdm

Important scripts

CNN-Transformer

Training

console python classification/CNN-Transformer/train_CNNTransformer.py -h, --help show this help message and exit -p PREPROCESSING, --preprocessing PREPROCESSING If preprocessing is not defined, script uses default 'base' preprocessing. --device DEVICE Device to run the model on. Options: cuda, cpu -lr LEARNING_RATE, --learning-rate LEARNING_RATE Learning rate used to optimize the model. --d_model D_MODEL d_model size of the transformer model. --stride STRIDE Stride used by the CNN backbone. --kernel_size KERNEL_SIZE Kernal size of the CNN backbone. --nhead NHEAD Number of multi headed attention used by the transformer. --num_layers NUM_LAYERS Number of Transformer layers. --width_multiplier WIDTH_MULTIPLIER Channel multiplier used by the CNN backbone. --dropout DROPOUT Dropout percentage. --FFN_dim_hidden_layers [FFN_DIM_HIDDEN_LAYERS ...] Number of neurons used by the classification component used as hidden layer --validation-batches VALIDATION_BATCHES Amount of validation batches used to validate the models performance. --epochs EPOCHS Number of training epochs. --num-workers NUM_WORKERS Number of workers used to preprocess data. -b BATCH_SIZE, --batch-size BATCH_SIZE Batch size. --mlflow-tracking-uri MLFLOW_TRACKING_URI MLflow tracking uri. --mlflow-experiment MLFLOW_EXPERIMENT MLflow experiment name.

Testing console python classification/CNN-Transformer/test_CNNTransformer.py -h, --help show this help message and exit --device DEVICE Device to run model on. Options: cuda, cpu -p [PREPROCESSING ...], --preprocessing [PREPROCESSING ...] If preprocessing is not defined, script uses default 'base' preprocessing. -d DATABASE, --database DATABASE The used database must contain the preprocessing selected with '--preprocessing'. This argument requires a path to the preprocessing files. -r RUN_ID, --run-id RUN_ID MLflow run id -o OUTPUT, --output OUTPUT Output folder of the testing performance. Default: 'output/{run-id}' --num-workers NUM_WORKERS Number of workers used to preprocess data. -b BATCH_SIZE, --batch-size BATCH_SIZE Batch size --mlflow-tracking-uri MLFLOW_TRACKING_URI MLflow tracking uri --mlflow-experiment MLFLOW_EXPERIMENT MLflow experiment name you can state multiple preprocessings at one, this will return the classification results for each preprocessing. The output folder should be set, because if the path already exists, a error will occur.

CNN-LSTM

Training

console python classification/CNN-LSTM/train_CNNLSTM.py -h, --help Show this help message and exit -p PREPROCESSING, --preprocessing PREPROCESSING If preprocessing is not defined, script uses default 'base' preprocessing. --device DEVICE Device to run the model on. Options: cuda, cpu -lr LEARNING_RATE, --learning-rate LEARNING_RATE Learning rate used to optimize the model. --bidirectional BIDIRECTIONAL Set the model to uni or bidirectional. --validation-batches VALIDATION_BATCHES Amount of validation batches used to validate the models performance. --epochs EPOCHS Number of training epochs. --num-workers NUM_WORKERS Number of workers used to preprocess data. -b BATCH_SIZE, --batch-size BATCH_SIZE Batch size. --mlflow-tracking-uri MLFLOW_TRACKING_URI MLflow tracking uri. --mlflow-experiment MLFLOW_EXPERIMENT MLflow experiment name.

Testing console python classification/CNN-Transformer/test_CNNTransformer.py -h, --help Show this help message and exit --device DEVICE Device to run the model on. Options: cuda, cpu -p [PREPROCESSING ...], --preprocessing [PREPROCESSING ...] If preprocessing is not defined, script uses default 'base' preprocessing. -d DATABASE, --database DATABASE The used database must contain the preprocessing selected with '--preprocessing'.This argument requires a path to the preprocessing files. -r RUN_ID, --run-id RUN_ID MLflow run id -o OUTPUT, --output OUTPUT Output folder of the testing performance. Default: 'output/{run-id}' --num-workers NUM_WORKERS Number of workers used to preprocess data. -b BATCH_SIZE, --batch-size BATCH_SIZE Batch size --mlflow-tracking-uri MLFLOW_TRACKING_URI MLflow tracking uri --mlflow-experiment MLFLOW_EXPERIMENT MLflow experiment name you can state multiple preprocessings at one, this will return the classification results for each preprocessing. The output folder should be set, because if the path already exists, a error will occur.