Disentangled knowledge representations for patient stratification in heart failure
Dr Andrew King
Reader in Medical Image Analysis
School of Biomedical Engineering & Imaging Science
Faculty of Life Sciences & Medicine
King’s College London
Prof Reza Razavi, Dr Bram Ruijsin, Dr Esther Puyol-Antón
Under current clinical guidelines, all patients with heart failure are essentially treated in the same way. All receive treatment with beta-blockers, angiotensin-converting-enzyme (ACE-)inhibitors, aldosterone antagonists, and second line treatment using angiotensin receptor neprilysin inhibitors and possibly cardiac resynchronisation therapy. However, in reality there are large differences between patients, both in terms of the disease underlying heart failure, as well as co-morbidities. This variation causes large differences in treatment response between patients, with some patients gaining little or no improvement in clinical outcome as a result of their treatment. Therefore, patients would benefit from more personalised treatment regimes. However, currently evidence is lacking on tailoring treatments to patient subgroups, because finding these subgroups is complex due to the many factors involved (cardiac function, co-morbidities, patient characteristics and underlying disease). Our aim in this project is to address this limitation in current treatments for heart failure.
We aim to develop artificial intelligence (AI) based techniques for forming knowledge representations that can discriminate between different patient subgroups. We will form the representations from multiple sources of data, including magnetic resonance imaging (MRI), echocardiography (echo) and electrocardiogram (ECG) data. We will develop multiview learning approaches, on the assumption that the different modalities represent different ‘views’ of the underlying disease characteristics. We will use deep learning models and the latest disentanglement techniques to derive biomarkers that can not only be used for improved patient stratification but are also interpretable to cardiologists. The representations and derived biomarkers will be linked, using supervised machine learning techniques, to outcome data (major adverse cardiovascular events, or MACE endpoints, which are a combined outcome measure based on mortality and hospitalisation for heart failure, myocardial infarction, stroke or serious arrhythmias), enabling the development of an AI-based personalised treatment regime recommendation tool.
Key to this ambitious but transformative aim will be the availability of clinical data for training purposes. To this end, we will use retrospective patient data from the Guy’s and St. Thomas’ Trust (GSTT) clinical imaging database. This contains detailed data from over 6,000 heart failure patients, including imaging and ECG data, as well as treatment regime and outcome information. Crucially, although the clinical guidelines recommend a standard treatment regime, the database contains data from patients treated with a range of combinations of the available treatments. This can occur due, for example, to allergies/intolerance for drugs and/or patient compliance. This variation in treatment regime should offer sufficient scope to allow the machine learning models to learn which treatment combinations are most effective for different patient subgroups. Our primary hypothesis is that these learnt personalisation models will lead to improved outcomes for heart failure patients.
We will make use the GSTT clinical imaging database in this research. This contains over 25,000 datasets from patients scanned at Guy’s and St. Thomas’ Hospitals with heart disease (of which approximately 6,000 are heart failure patients). This number will increase by approximately 5,000 per year (1,000 with heart failure). Opt-out ethical approval is in place for research use of all retrospective and future data within the database. Through the London Medical Imaging & Artificial Intelligence Centre for Value-Based Healthcare technological infrastructure has been developed to enable easy access to deidentified imaging data from this database. Linked patient EHR data is held on a secure GSTT server with the capability to upload and train/evaluate AI models. Each dataset will contain MRI and echo data, as well as ECG, treatment and outcome data stored in the EHR.
AI, deep learning, cardiac, knowledge representation, disentanglement