AI Applied To THe Electrical Activation Of The Atria During Atrial Fibrillation
Atrial fibrillation (AF) is one of the most prevalent heart rhythm disorders at present. It has been a challenge for a century and is becoming a significant issue in the future without substantial action. Experts in most Western countries advise that this problem will cause even greater challenges without stronger interventions. Despite extensive background knowledge, managing AF is not straightforward, and the detection process can be labor-intensive. The decision-making process is influenced by factors such as the patient's symptoms, preferences, and overall health condition. The electrocardiogram (ECG) required to confirm the diagnosis necessitates substantial expertise, and multiple recordings are needed to confirm the diagnosis, making the analysis a laborious task. New perspectives and developments are welcome to alleviate this situation, and a transformative solution is desired.
The analysis of multiple ECG recordings requires significant effort from cardiologist professionals. Furthermore, the healthcare environment can be stressful, exacerbating the situation and prolonging the AF detection process, leading to increased expenses. Assistive tools are needed to ease this challenge. Moreover, early diagnosis is crucial for effective treatment, as current AF treatments have limitations in terms of efficacy and adverse effects, posing challenges in managing the condition effectively. This situation is particularly challenging for asymptomatic or paucisymptomatic AF patients because home monitoring would be beneficial before the condition progresses to more severe symptoms, after which the treatment's success rate usually declines. These multifaceted challenges in diagnosing and managing AF emphasize the need for ongoing research and tailored approaches to improve detection and treatment outcomes.
Wearable devices are one of the modern technologies that can address the challenge of AF management, particularly in early detection. Smart devices, such as smartwatches and small heart monitors, can continuously track the heart's rhythm and send alerts when arrhythmia, including AF, is detected. Furthermore, the latest advancements in artificial intelligence (AI), particularly deep learning (DL), have shown promising potential in many challenging applications and are becoming a strong contributor to the future. However, the AI integrated into these devices faces computational limitations. In this work, publicly available databases included 23 AF patient ECGs, 18 sinus rhythm patient ECGs, 48 different arrhythmia patients ECGs, 8528 different ECG recordings (AF, sinus rhythm, other arrhythmias, and noisy recordings), and 35 patient PPGs (19 AF, 16 not AF) data were used to train and test the proposed model. The uncompressed baseline model is a convolutional neural network. The compressed baseline model developed in this work had 1600-1700 parameters, which is 91% less compared to the uncompressed model (17900). The compressed model file size is 46-48 kB, which is a 56% reduction compared to the uncompressed one (104 kB), a significant factor for low-computational devices. The compressed DL model achieved 93.27%, 97.88%, and 95.62% accuracy in three different test sets. Furthermore, a 5-fold cross-validation accuracy of 97.26% demonstrates its potential to address the challenge. Nevertheless, the models need to be updated periodically due to the limited training data, and low-computational devices have limitations for the training process. Therefore, partial training of the model was also studied, where the DL model layers were set as modules, and their retraining results were compared to the whole model training, with a difference of only 0.2 percentage points at the lowest. This aspect provides a feasible setup for the fully built-in algorithm, although in that case, the data would be acquired from a single patient. To address this, a setup where data from an individual patient was used for training and testing was implemented, improving the accuracy from 0.52 to 50.94 percentage points for different patients using modules.
History
Supervisor(s)
Timothy C. Pearce; G. Andre NgDate of award
2025-03-07Author affiliation
School of EngineeringAwarding institution
University of LeicesterQualification level
- Doctoral
Qualification name
- PhD