Novel Imaging Revealing Inner Dynamics for Cardiovascular Waveform Analysis via Unsupervised Manifold Learning

Abstract

Background: Cardiovascular waveforms contain information for clinical diagnosis. By "learning" and organizing the subtle change of waveform morphology from large amounts of raw waveform data, unsupervised manifold learning helps delineate a high-dimensional structure and display it as a novel three-dimensional (3D) image. We hypothesize that the shape of this structure conveys clinically relevant inner dynamics information. Methods:To validate this hypothesis, we investigate the electrocardiography (ECG) waveform for ischemic heart disease and arterial blood pressure (ABP) waveform in dynamic vasoactive episodes. We model each beat or pulse to be a point lying on a manifold, like a surface, and use the diffusion map (DMap) to establish the relationship among those pulses. The output of the DMap is converted to a 3D image for visualization. For ECG datasets, first we analyzed the non-ST-elevation ECG waveform distribution from unstable angina to healthy control in the 3D image, and we investigated intraoperative ST-elevation ECG waveforms to show the dynamic ECG waveform changes. For ABP datasets, we analyzed waveforms collected under endotracheal intubation and administration of vasodilator. To quantify the dynamic separation, we applied the support vector machine (SVM) analysis and reported the total accuracy and macro F1 score. We further carried out the trajectory analysis and derived the moving direction of successive beats (or pulses) as vectors in the high-dimensional space.Results: For the non-ST-elevation ECG, a hierarchical tree structure comprising consecutive ECG waveforms spanning from unstable angina to healthy control is presented in the 3D image (accuracy = 97.6%, macro-F1=96.1%). The DMap helps quantify and visualize the evolving direction of intraoperative ST-elevation myocardial episode in a 1-hour period (accuracy=97.58%, macro-F1=96.06%). The ABP waveform analysis of Nicardipine administration shows inter-individual difference (accuracy=95.01%, macro-F1=96.9%) and their common directions from intra-individual moving trajectories. The dynamic change of the ABP waveform during endotracheal intubation shows a loop-like trajectory structure, which can be further divided using the manifold learning knowledge obtained from Nicardipine. Conclusions:The DMap and the generated 3D image of ECG or ABP waveforms provides clinically relevant inner dynamics information. It provides clues of acute coronary syndrome diagnosis, shows clinical course in myocardial ischemic episode and reveals underneath physiological mechanism under stress or vasodilators.Keywords: manifold learning, diffusion map, electrocardiography, myocardial ischemia, arterial 10,583 ECG beats, including clinical unstable angina (red), ischemic ECG pattern (green), and healthy control (blue). Video F3CCorresponding to Figure 3 panel C, it shows the 3D animation of 1-min ECG beats during the intra-operative ST-elevation event. Video F3DCorresponding to Figure 3 panel D, it shows the 3D animation of 4,299 consecutive ECG 31 beats from the IS...

Robust and scalable manifold learning via landmark diffusion for long-term medical signal processing

“…One solution to depict intrinsic dynamics directly from the original waveform is obtaining as many features as possible, and selecting suitable parameters for the learning purpose [19]. Another solution is applying manifold learning algorithms to the original physiological waveforms [28,48]. The basic idea in [28,48] is truncating the physiological waveform into pieces according to some rules, and then apply the spectral embedding algorithm, like the diffusion maps (DM) [9], to embed those pieces into a finite dimensional Euclidean space, which represents the intrinsic dynamics.…”

“…Another solution is applying manifold learning algorithms to the original physiological waveforms [28,48]. The basic idea in [28,48] is truncating the physiological waveform into pieces according to some rules, and then apply the spectral embedding algorithm, like the diffusion maps (DM) [9], to embed those pieces into a finite dimensional Euclidean space, which represents the intrinsic dynamics. If the physiological waveform is embedded into the three dimensional Euclidean space, the physiological waveform is converted into a three dimensional image so that users can visualize the waveform from a different perspective.…”

“…It is commonly believed that more data lead to more comprehensive "knowledge" in the learning procedure, and hence more informative features and better visualization of the intrinsic dynamics. We thus hypothesize that if the technique shown in [28,48] could be applied to analyze long-term physiological waveforms of length on the order of days or weeks, it will be beneficial to supporting the short-term memory of the human brain when handling long-term physiological waveforms. Specifically, due to the short-term memory limitation, it is easy to overlook information hidden in the long-term waveform; for example, what is the relationship between the waveform in the first hour of the operator and the 7th hour?…”

“…Will the surgeon remember the useful information in the first hour after 5 hours? While this idea is natural, unfortunately, it is prohibited by the computational complexity inherited in most spectral decomposition based machine learning algorithm, including the DM applied in [48].…”

“…In practice, it is natural to consider subsampling the dataset, however, we may loss information. As a result, although the DM works well and provides alternative clinical information [48], it is limited to datasets of length about one hour.…”