Machine learning, computer modelling and signal processing for biomedical data: a contribution to reperfusion arrhythmias understanding
Lead Research Organisation:
University of Oxford
Department Name: Computer Science
Abstract
This project falls within the Artificial Intelligence research area from the Healthcare Technologies theme of the Engineering and Physical Sciences Research Council (EPSRC). More precisely, this research will contribute to new approaches for overcoming some of the limitations of deep learning applications in biomedical data analysis, such as dealing with signal and imbalanced datasets, while providing new strategies for reperfusion arrhythmias prediction in patients receiving percutaneous transluminal coronary angiography.
Nowadays, the performance of health care professionals is constrained by their resources' limitations. In fact, several clinicians have to analyse large amounts of biomedical data each time they perform a disease diagnosis or any other decision-making task. Thus, the number of patients that one physician can daily attend could be incremented by providing them support systems, which can automatically show them all relevant biomarkers from an incoming patient's data
Moreover, some biomedical analysis related tasks can already be dealt by automatic systems based on shallow machine learning (ML) models. However, implementing shallow ML approaches implies designing a handcrafted feature extraction strategy, which within the biomedical research field can be overly complicated due to the complexity of the data. Namely, medical datasets are usually multimodal and imbalanced. In addition, these also have much more measures than subjects in comparison to other types of data. Therefore, most of the existing techniques achieve insufficient performance levels for the clinicians to daily employ them (e.g. FOG real-time detection in PD patients state-of-the-art results are lower than 85% for both, sensitivity and specificity) On the other hand, deep learning techniques have proven successful on several of these complex tasks; however, they have difficulties when dealing with small, highly imbalanced and time series datasets.
Therefore, within this study, novel theoretical deep learning models and signal processing strategies for adequately addressing the mentioned difficulties related to biomedical data will be developed, while improving the state-of-the-art results for reperfusion arrhythmia's prediction. More precisely, this work will focus on exploring and expanding the following methodologies and approaches: transfer learning techniques; special deep learning architectures simulating and generative models representation methods; and event prediction models. Furthermore, all our approaches will be evaluated using medical data from patients receiving elective prolonged percutaneous transluminal coronary angiography (PTCA).
The proposed study will be conducted by Julia Camps as his 3-year DPhil project in collaboration with the Computational Cardiovascular Science research group at the University of Oxford. This research group disposes from numerous sources of relevant data to tackle the reperfusion arrhythmias prediction problem. For example, they have a dataset acquired from 108 patients while being administrated elective PTCA. This data is composed of 12-lead ECG signals and images obtained by injecting Technetium Tc99m Sestamibi to facilitate the detection of coronary artery disease by localising myocardial ischemia. Moreover, these group includes scientists specialising in ML and deep learning techniques and computer modelling; besides, it has well-established links with clinical collaborators from the John Radcliffe Hospital through the Oxford Acute Myocardial Infarction (OXAMI) project. The OXAMI is an ongoing project, which aims at finding new biomarkers to predict risk and recovery features in patients with acute myocardial infarction. Therefore, the project will have access to substantial resources regarding both academic and data.
Nowadays, the performance of health care professionals is constrained by their resources' limitations. In fact, several clinicians have to analyse large amounts of biomedical data each time they perform a disease diagnosis or any other decision-making task. Thus, the number of patients that one physician can daily attend could be incremented by providing them support systems, which can automatically show them all relevant biomarkers from an incoming patient's data
Moreover, some biomedical analysis related tasks can already be dealt by automatic systems based on shallow machine learning (ML) models. However, implementing shallow ML approaches implies designing a handcrafted feature extraction strategy, which within the biomedical research field can be overly complicated due to the complexity of the data. Namely, medical datasets are usually multimodal and imbalanced. In addition, these also have much more measures than subjects in comparison to other types of data. Therefore, most of the existing techniques achieve insufficient performance levels for the clinicians to daily employ them (e.g. FOG real-time detection in PD patients state-of-the-art results are lower than 85% for both, sensitivity and specificity) On the other hand, deep learning techniques have proven successful on several of these complex tasks; however, they have difficulties when dealing with small, highly imbalanced and time series datasets.
Therefore, within this study, novel theoretical deep learning models and signal processing strategies for adequately addressing the mentioned difficulties related to biomedical data will be developed, while improving the state-of-the-art results for reperfusion arrhythmia's prediction. More precisely, this work will focus on exploring and expanding the following methodologies and approaches: transfer learning techniques; special deep learning architectures simulating and generative models representation methods; and event prediction models. Furthermore, all our approaches will be evaluated using medical data from patients receiving elective prolonged percutaneous transluminal coronary angiography (PTCA).
The proposed study will be conducted by Julia Camps as his 3-year DPhil project in collaboration with the Computational Cardiovascular Science research group at the University of Oxford. This research group disposes from numerous sources of relevant data to tackle the reperfusion arrhythmias prediction problem. For example, they have a dataset acquired from 108 patients while being administrated elective PTCA. This data is composed of 12-lead ECG signals and images obtained by injecting Technetium Tc99m Sestamibi to facilitate the detection of coronary artery disease by localising myocardial ischemia. Moreover, these group includes scientists specialising in ML and deep learning techniques and computer modelling; besides, it has well-established links with clinical collaborators from the John Radcliffe Hospital through the Oxford Acute Myocardial Infarction (OXAMI) project. The OXAMI is an ongoing project, which aims at finding new biomarkers to predict risk and recovery features in patients with acute myocardial infarction. Therefore, the project will have access to substantial resources regarding both academic and data.
Organisations
People |
ORCID iD |
| Blanca Rodriguez (Primary Supervisor) | |
| Julia Camps (Student) |
Studentship Projects
| Project Reference | Relationship | Related To | Start | End | Student Name |
|---|---|---|---|---|---|
| EP/N509711/1 | 01/10/2016 | 30/09/2021 | |||
| 1894700 | Studentship | EP/N509711/1 | 01/10/2017 | 30/09/2020 | Julia Camps |
| Description | I designed an electrocardiogram signal delineation system based on machine learning. This approach would allow analysing records from different databases automatically without introducing unknown variability in the delineation process. Although I published this approach as a conference paper, I still have to validate it properly on data that it's being currently being labelled through a collaboration with a medical centre in Poland. I also designed a bayesian-inference based pipeline that allows recovering the parameters in a cardiac model that would produce a specific electrocardiogram signal. This inference approach can be key for understanding the role of the electrical network structures in the inner-cardiac layers in the morphology of the signals recorded at the clinic. Both approaches are yet to be properly published and combined to understand the link between specific features in the electrocardiogram with anatomical properties of the heart in healthy and diseased pathologies. |
| Exploitation Route | The delineation approach, when properly published, could become a standard for electrocardiogram signal delineation, which would ease reproducibility in studies that analyse these recordings. On the other hand, the inference pipeline can be expanded to address personalisation on specific patients data. As a matter of fact, this approach is computationally inexpensive and can find a suitable cardiac-model calibration for a given combination of electrocardiogram and magneto resonance imaging data in a matter of hours. |
| Sectors | Healthcare,Pharmaceuticals and Medical Biotechnology |