Quantifying the impact of Signal Simplification, Data Quantity, and Task Difficulty on Vision Transformer Performance for Electrocardiogram Rhythm Classification

Jarod Hartley, W. Joseph MacInnes

Keywords: Vision Transformer, ViT, SwinV2, Signal Simplification, Data Quantity, Task Difficulty, Electrocardiogram, Electrocardiography, ECG, Machine Learning, Rhythm, Classification

TL;DR: Investigation into how different data conditions may affect vision transformer performance in ECG rhythm classification.

Abstract: This paper investigates the interplay between signal simplification, data quantity and task difficulty in the context of electrocardiography rhythm classification using vision transformer models. Recognising the complexity and variability inherent in ECG signals, we examine how applying a simple pre-processing technique aimed at enhancing human readability, through noise reduction, affects the accuracy and robustness of machine learning models. We also consider the broader implications for model development, particularly in relation to the challenges posed by real-world ECG data. Given the inherent complexity and diversity of ECG diagnoses, it is often impractical to gather substantial amounts of data for every potential diagnosis. By assessing how different dataset sizes affect performance, we seek to understand the extent to which vision transformers rely on data quantity. This is particularly important given the limitations of real-world datasets and its potential impact on automated diagnostic systems. This paper further examines the scalability of ECG classification by employing a dataset encompassing ten distinct conditions. While previous research has demonstrated success in scenarios involving a limited number of conditions, such controlled environments are rarely representative of real-world practice. Therefore, it is crucial to understand how vision transformer models perform when faced with more complex and varied classification tasks and to evaluate their capacity to manage increased diagnostic diversity. Our findings provide insight into optimising ECG classification pipelines with regards to balancing the need for data clarity, quantity, and diagnostic breadth to enable reliable and scalable AI-driven cardiac assessment.

Primary Subject Area: Application: Cardiology

Secondary Subject Area: Safe and Trustworthy Learning-assisted Solutions for Medical Imaging

Registration Requirement: Yes

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 330