Selective structured state-space models for ECG classification
Abstract
Electrocardiograms (ECGs) are a standard, life-saving test, yet their interpretation and classification still rely on clinicians’ expertise. End-to-end deep learning has been shown to enable an automated ECG classification pipeline that outperforms medical students (https://www.nature.com/articles/s41467-020-15432-4). However, many challenges still need to be addressed.
We focus on two main challenges. The first of these challenges concerns the applicability of these deep learning-based automatic classification pipelines to other types of exams, like Holter monitoring, where the length of the traces poses significant challenges—in terms of computational scalability and feasibility of real-time classification—to traditional deep learning architectures. We intend to address this challenge by developing end-to-end architectures based on Selective State-Space Models, with the aims of (i) scalability to extremely long sequences, (ii) real-time exam classification, and (iii) applicability to ECGs with different number of leads.
Another relevant aspect of automated ECG classification is transferability, i.e., whether a model trained on patient data from a certain medical cohort generalizes to other cohorts or not. Ideally, we would want models that are general and robust enough to withstand a change in cohort without losing performance. The idea of training adversarially, or to train on samples perturbed by an ill-intended adversary, makes for more robust models that can withstand these kinds of input changes. The second part of the project will therefore be devoted to investigating if adversarial training can increase the transferability of a model, in comparison with its traditionally trained counterpart.
