Artificial Intelligence Augmentation of the Paper ECG

Background

An electrocardiogram (ECG) is a non-invasive and widely available investigative modality which provides clinicians with important temporal and anatomic data about the heart’s electrical activity. ECGs are typically printed as paper documents using thermal printers. These are subsequently viewed and analysed by clinicians leaving their interpretation prone to inter-observer variability [1-2]. At present, very little of the data an ECG provides finds its way into electronic health records (EHRs). Where attempts to integrate ECGs into EHRs are made, the paper documents are typically scanned in a format where the signal data stored within ECGs cannot easily be retrieved.

Recent studies have demonstrated the potential for deep learning (DL) to enable rapid and reliable digital ECG interpretation and to detect subtle ECG signals with important clinical implications [3-4]. For example, DL-based ECG interpretation has been demonstrated to be capable of predicting the likelihood of a patient developing atrial fibrillation from a sinus rhythm ECG [5] or to accurately predict a patient’s serum potassium level from an ECG tracing. Such capabilities demonstrate the huge potential that DL-based ECG interpretation offers.

Despite this potential, translation of AI-ECG into widespread clinical practice is not possible owing to the absence of any technology to apply AI-ECG algorithms to paper ECGs. This challenge represents the subject of the proposed PhD.

Aims

The core aim of this project is to develop a software system capable of addressing this shortcoming with the following components:

a) An app running on a handheld device capable of taking pictures of paper ECG and an algorithm that leverages a supervised machine learning approach, such as a convolutional neural network, to digitise the ECG signals plotted in paper ECGs in a format suitable for input into AI-ECG algorithms (e.g. a numerical matrix with a row for each ECG lead and columns representing time). This can be seen as a regression problem where the input is an image of the paper ECG and the output is the ECG matrix. We have access to a large database of ECGs, covering a broad range of clinical findings in digital format (i.e. already in matrix format), together with synthetic images of printed ECGs that can be used for training/validation purposes.

b) A cloud environment that: i) interacts with the app to offer a number of validated AI-ECG algorithms to be applied on the digitised ECG to obtain a diagnosis or ECG characteristic, ii) provides an API for novel AI-ECG algorithms to be submitted to the system by developers and for these to be validated against private benchmark data curated for the range of conditions of interest, iii) a set of synthetic data to be freely downloaded for algorithm development, and iv) a set of workflows and policies for validating submitted algorithms and making them available to the app under a set of quality assurance rules.

Training outcomes

The student will receive state-of-the-art training in the core disciplines of image analysis, computational modelling, and data science while gaining expert knowledge in the context of ECG interpretation and cardiology more generally. This highly interdisciplinary approach is well aligned with the “T-shaped researcher” training requirements identified as key in the DTP.