Anomaly Detection and Disease Classification in Electrocardiogram Signals using Advanced Signal Processing and 3D Image Classification

For provision of correct medicine, treatment and correct therapy, an accurate diagnosis is required. There are situations in which time-critical decisions need to be made to provide the right treatment to avoid long term morbidity or in some cases even death. In England alone, cardiovascular diseases’ (CVD) related healthcare costs amount to an estimated £7.4 billion per year, and annual costs to the wider economy being an estimated £15.8 billion. The UK government is focused on telemedicine, and in next 5 years with the help of NHS, they plan to save at least 150,000 lives each year by avoiding/managing heart attacks.

The procedure of placing the electrodes on the human body and collecting the electric signals from a cardiac muscle during each cardiac cycle is known as Electrocardiography (ECG) and the obtained signal is known as ECG signal. The health of the cardiac muscle is commonly measured using ECG signals, which are used to measure; the rhythm and rate of each cardiac cycle, damage of cardiac muscle or conduction system, the function of inserted pacemakers in the heart, the effects of cardiac-related medications and the dimensions of the cardiac chambers.

To detect and classify CVDs, the time-sequence of ECG signals is carefully examined by a cardiologist or electrophysiologist. The detection of abnormalities in the ECG signals is stated as anomaly detection. The ECG time-series signals are composed of massive data comprising of fast cardiac rhythms and these vary from patient to patient, which makes anomaly detection a challenge. Computer aided anomaly detection in ECG signals and its correct classification is significantly important to identify dissimilar cardiac beats, which further results in detecting related CVDs through manual diagnosis performed by a cardiologist or electrophysiologist.

The existing anomaly detection approaches in ECG signals undergo a high false-positive (FP) and true-negative (TN) rate. The existing methods are not intelligent enough to distinguish between normal ECG signals and ECG artefacts. In terms of frequency or rhythms, the ECG artefacts which are quite similar with normal ECG signals are indistinguishable. Current research, which is mostly based on powerful deep learning networks i.e., long short-term memory units, generate high FP and TN rates because they focus only on clean signals without concerning noise or artefacts in signals. Ignoring the noise or artefacts in ECG signals is impractical and causes the outcomes to inappropriately ensure the normal cardiac cycle or its morphology, which is critical for further analysis of disease classification. The FP and TN rate of anomaly detection in ECG signals tends to rise substantially even in powerful deep learning networks, which remains a challenge for the reliable real-time implementation.

This proposal focuses on introducing smart ECG machines which can detect anomalies in ECG signals in real-time. Based on an advanced artificial intelligence (AI) aided signal processing models, these machines will be able to memorise the sequences of ECG signals. In order to classify the CVDs, the acquired abnormal signals will be further converted to spectrogram images for input to a 3D image classification model. One of the biggest hurdles in training an accurate AI-based disease classification system is the training data. There are several open-source ECG databases available all across the globe which are free to use and contain tens of thousands of real-time ECG samples.

This project will mainly focus on achieving the following goals:

• Detection of anomalies in ECG signals using AI-aided advanced signal processing, including localisation of abnormal signals.
• CVD classification of localised abnormal signals using 3D image classification modelling.
• Achieving generalisation in detection and classification of CVDs in all patients, which will essentially lead to a solution that will be able to detect anomalies and classify CVDs in all kinds of patients.
• Achieve ECG leads generalisation.
• Reducing the FP and TN rate, making it a time and resource efficient solution for the physicians.
• Real-time implementation of Smart ECG machines for clinical experimentation and testing. MoUs will be established to work with NHS trusts for successful implementation of the project.