Comparative_Analysis_of_CNN-CBAM_Fusion_Model_and_ViT

The code for comparative analysis of CNN+CBAM and ViT for processing ECG image data in joint fusion with EHR data for post-PCI prognosis.

Background

Assessing the patient's prognosis after undergoing PCI (Percutaneous Coronary Intervention) is vital for understanding their recovery trajectory and potential risks post-procedure. This evaluation not only aids in gauging the intervention's efficacy but also in predicting the probability of future cardiovascular incidents. We have proposed a multimodal, joint-fusion based multibranch CNN+CBAM model targeting the prediction of three different clinical endpoints i.e. heart failure hospitalization, all-cause mortality, and ischemic stroke (6-months post-PCI).

Data

We have considered Electrocardiogram (ECG) and Electronic Health Reacords (EHR) data of patients.

Data preprocessing - ECG Cropping

The ECG data images are preprocessed before being sent to the model

Proposed Models

We have implemented a hybrid fusion model consisting of multibranched CNN with Attention (CBAM) for processing ECG data and a feed forward neural network for processing the Electronic Health Record(EHR) data.

For comparative benchmarking, we have used a Vision Transformer (ViT) model instead of CNN+CBAM branch for processing ECG images while keeping same feed forward neural network architecture for EHR data processing.

CNN+CBAM Fusion Model Architecture

Overall Architecture

Spatial Attention Module

Channel Attention Module

ViT Fusion Model Architecture

How to Implement?

Run the main.py file using the code present below, it takes a CSV file containing EHR data as an input along with the model you want to consider for processing the ECG data branch i.e. either CNN+CBAM or ViT. The EHR data was named as "Share_data_outcome_deidentified.csv" in our case

!python3 main.py -i <input-csv-file> -m <model-name>

where <input-csv-file> represents the CSV file containing the EHR data, in our case it's named as "Share_data_outcome_deidentified.csv". The <model-name> argument indicates the model which you wish to run, which could be either CNN or ViT.

Example...

!python3 main.py -i Share_data_outcome_deidentified.csv -m CNN

This will automatically save the ROC curves along with confidence intervals (CI) in "ROC_Outcomes" directory. The data we worked on has not been uploaded yet.

Result

The results obtained and the comparisons have been mentioned in our manuscript. Following table displays the obtained ROC curves along with CI.