The embeddings.csv file is required to run the models and analysis. It is located in the Embeddings.zip file located on Google Drive. The specific file used in the archive was Embeddings/brset/embeddings.csv.
The labels.csv file is taken from the BRSET dataset and can be downloaded from Physionet. The images are also required in order to run image_analysis.ipynb. Running this ipynb file will also require Model Weights for the Convextv2 model.
Python 3.10.14 was used. To install the required libraries, run the following command:
pip install -r requirements.txtresplit_data.ipynb contains the code to resplit the data into training, validation, and test sets based on patient_id so no patient is in more than one set.
image_analysis.ipynb contains the code for analyzing the BRSET dataset using a Convextv2 model with weights provided by David Restrepo. The model is used to predict diabetic retinopathy from the images in the dataset.
text_analysis.ipynb contains the code to analyze the tabular data from labels.csv file, which where used to generate synthetic notes for the text embeddings.
fusion.ipynb contains the code for the fusion models, which combine the image and text embeddings to predict diabetic retinopathy.
calc_stats.ipynb contains the code to calculate the statistics for the models based on saved prediction probabilities generated from the above notebooks.