Criticality-aware Deconfounded Classification of Long-tailed Multi-label 12-lead Electrocardiogram

Trisrota Deb, Ishan Sahu, Arijit Ukil, Arpan Pal, Sundeep Khandelwal, Utpal Garain

Abstract: We often observe long-tailed distribution in real-world classification problems and consequently, maintaining balanced predictive performance across all the classes is a research challenge. Further, we find, particularly in time series classification tasks like prediction of clinical diseases from physiological signals like Electrocardiogram (ECG), the existence of critically important rare classes and the cost of low sensitivity towards such rare yet critical classes are extremely high not only with higher treatment expenses, but also with higher chances of mortality. We focus on the practical challenge of maximizing the predictive sensitivity of rare yet critically important classes in a long-tailed time series classification with a favourable trade-off towards the aggregate classification performance of the remaining classes. We develop a class criticality-aware inference algorithm by customizing the total direct effect (TDE) in the deconfounded training by incorporating the domain knowledge into the degree of TDE that impacts the decision probabilities by minimizing the confounding variable (momentum in Stochastic Gradient Descent optimizer), which is responsible for the occurrences of high valued classification logits of the head classes. We demonstrate the real-world efficacy through empirical study on practically important Physionet challenge ECG 2020 dataset, which is a multi-label dataset with 27 different cardio-vascular disease classes from 12-lead ECG recordings. From the obtained experimental results with ablation study and state-of-the-art comparison investigation, we clearly observe that our proposed method outperforms the current benchmark algorithm and importantly, it is able to consistently improve the sensitivity metrics while predicting the clinically important yet rare classes.