Machine learning techniques to predict restenosis development in superficial femoral arteries using demographic, clinical and hemodynamic descriptors

点击翻译按钮获取中文摘要

Brief description of the context of the research including potential impactPeripheral arterial disease (PAD) is the third cause of cardiovascular morbidity and is mainly caused by atherosclerosis. If the disease is detected at an early stage, endovascular approaches, such as percutaneous balloon angioplasty (PTA), may prove to be effective as a minimally-invasive treatment for diseased superficial femoral arteries (SFAs). A common negative outcome of this procedure is restenosis, occurring in more than 60% of cases at 1- year follow-up. In addition to clinical risk factors, biomechanical factors play an important role on the lumen remodelling, mainly because of the altered hemodynamics occurring over time after the intervention. There is evidence that the combination of low time-averaged wall shear stress (< 1 Pa) and high oscillatory shear index (> 0.20) favours atherogenesis and restenosis. The distribution of the hemodynamic indexes, along the SFA wall, can be accessed by through patient-specific vessel geometry 3D reconstructions and setting of personalised computational fluid dynamics (CFD) simulations. Once clinical and biomechanical variables having the strongest relationship with restenosis development are identified, the use of artificial intelligence techniques, such as machine learning (ML), may help to classify patients according to their risk of developing restenosis in a prescribed range of time. This would help clinicians to predict whether treated areas are likely to undergo restenosis during a defined time interval, hence tailoring surveillance and prevention programmes for PAD. Aims and ObjectivesThe project aim is to fuse demographic, clinical and hemodynamic descriptors information for a subset of PAD patients to build a ML driven model able to predict the risk of developing restenosis in SFAs, after PTA intervention, in a prescribed time interval. The main steps to accomplish are:i) Perform a critical analysis on the current clinical pathway, in terms of imaging modalities, to identify the information they provide and how they can be feasibly improved to match the needs of the research. Standardised data would make the ML model replicable both in other projects and, potentially, in clinical use.ii) Data collection and preparation, allowing the computation of the hemodynamic indexes, along the SFA wall, by means of personalised CFD analyses and their combination with demographic and clinical information by ML algorithms. A dataset of at least 100 patients will be needed.iii) Development of the ML model best performing for the problem under investigation. Even if supervised ML techniques are generally used to develop models predicting future clinical events, unsupervised ML techniques will not be discarded a priori.iv) Validation and assessment of model performance, by comparing the obtained results with the ones obtained from classical statistical analyses, i.e. standard predictive multivariate logistic regression models, to determine if ML techniques perform better.Novelty of Research MethodologyTo the best of author's knowledge, this is the first study fusing demographic, clinical and hemodynamic information when using ML approaches to predict restenosis risk in the SFA. From literature review, ML techniques generally perform their prediction only using variables obtained in routinely clinical practice and are normally applied to coronary arteries.Alignment to EPSRC's strategies and research areasThe project aligns with EPSRC's strategies since its main aim is to optimise treatment by tailoring it according to patient's needs, to provide clinicians tools to make more efficient surveillance programmes and use available resources at their best.Any companies or collaborators involvedProf. Janice Tsui - Royal Free Hospital, London - Clinical supervisor providing clinical data and putting towards relevant clinical research questions