Project Details
Deep learning of echocardiography for risk prediction in atrial fibrillation
Applicant
Dr. Shinwan Kany
Subject Area
Cardiology, Angiology
Term
since 2023
Project identifier
Deutsche Forschungsgemeinschaft (DFG) - Project number 521832260
Atrial fibrillation (AF) is the most common arrhythmia in adults. Currently, clinical risk prediction models for incident AF risk or risk of stroke in AF patients offer poor discrimination for individual risk estimation. The phenotype of AF is a consequence of multiple different biological pathways. Machine Learning (ML) algorithm can be exploit information that is not accessible to the human eye. For instance, ML models can be trained to predict incident AF risk from electrocardiograms (ECG) of patients with sinus rhythm. Studies have shown that this information is complementary to clinical risk scores. Echocardiography (Echo) offers more information on cardiac function and structure than ECG.The Community Care Cohort Project (C3PO) cohort of the Mass General Brigham systems provides comprehensive data on over 500,000 patients with over 15 years of follow-up (FU). From within the C3PO, we plan to train ML models to predict incident AF risk in patients without prevalent AF and incident stroke risk in AF patients by using Echo videos. For this, we use an established pipeline of over 228,000 anonymized Echo’s from >88,000 patients of the Massachusetts General Hospital.In a first quality control (QC) step, we will train a model (QC Model 1) to recognize bad image quality to filter out videos of bad quality. Training will be done with annotation from two cardiologists.In a second step (QC model 2), we will train a model to predict Echo measurements from the Echo reports to check for discrepancies and allow for troubleshooting. Then, we will train a model on patients without prevalent AF where the diagnosis of AF in C3PO FU is known. The dataset will be split in a training set, a further set for finetuning and a test set that the ML model will only see once training is complete. In a separate step, we will train a ML model to predict incident stroke with Echo videos from AF patients following a similar process. The next validation step is to test the model in an independent Echo dataset of similar size from the Brigham and Women’s Hospital. Additionally, we will calculate polygenic risk scores from patients with available genetic data and calculate clinical risk scores (CHARGE-AF und CHA2DS2-VASc) for incident AF risk and incident stroke risk in AF patients. These risk scores will be compared with the ML-derived risk. Additionally, we will study combined risk prediction using all risk scores to exploit all available data to improve individual risk prediction as good as possible.
DFG Programme
WBP Fellowship
International Connection
USA