Application of a Tensor-Based Classification Method with Electrocardiogram Data
Date:
Risk stratification for sepsis-induced organ dysfunction is difficult because of its heterogeneity. Moreover, the information needed to compute the Sequential Organ Failure Assessment (SOFA) defined by Sepsis-3 may not be immediately available for patients that require care. Continuous electrocardiogram (ECG) recording is often available for patients in intensive care, and may be useful to identify those at risk of decompensating to septic shock. Additionally, structuring multi-lead ECG data as a higher-order tensor can preserve the data’s structure and prevent loss of information that can occur in traditional machine learning, which structures data as 1-dimensional vectors. The project discussed here is a first step in this research: the development of a tensorial data classification method, and its testing with publicly available ECG data. Methods: We developed a tensor classification method based on multilinear discriminant analysis via Kempf-Ness Theory (KNMDA). We compare performance of this method to existing tensor methods and traditional machine learning, including support vector machine (SVM) and linear discriminant analysis (LDA), at different noise levels. We constructed 3rd order tensors using 12-lead ECG data from the Physionet PTB database, which contains samples of both normal control and unhealthy subjects. Results: Over thirty iterations, our KNMDA method obtained average area under the receiver operating characteristic curve (AUC) of 0.94 (SD 0.03), an improvement compared to linear SVM with 0.91 (0.03) and LDA with 0.68 (0.07) with no additional noise. When signal-to-noise ratio of the ECG signal was changed to -10, KNMDA achieved average AUC of 0.92 (0.03), compared to SVM with 0.85 (0.06) and LDA with 0.59 (0.07). Impact: The results show that our proposed method can outperform existing ones when using real ECG data. We can therefore apply this method to our larger goal of improving risk-stratification for septic patients, to augment physician decision-making and improve patient care.