WO2023055846A1

WO2023055846A1 - Cardiac disease diagnosis using spatial feature extraction from vectorcardiography signals

Info

Publication number: WO2023055846A1
Application number: PCT/US2022/045104
Authority: WO
Inventors: Yasuyuki Kataoka; Hitonobu Tomoike
Original assignee: Ntt Research, Inc.
Priority date: 2021-09-30
Filing date: 2022-09-28
Publication date: 2023-04-06

Abstract

Embodiments disclosed herein use spatial feature extraction from vector cardiography (VCG) signals using minimum volume ellipsoid enclosure. The features may be used to classify left ventricular hypertrophy, but may be more generally used to identify any cardiac disease.

Description

CARDIAC DISEASE DIAGNOSIS USING SPATIAL FEATURE EXTRACTION FROM VECTORCARDIOGRAPHY SIGNALS

CROSS REFERENCE

[0001] This application claims priority from US Provisional application 63/250,850 filed 30 September 2021, the entirety of which is hereby incorporated by reference.

FIELD

[0002] The disclosure includes a system and method that elucidate cardiac electrical properties using the extraction of spatial features from a vectorcardiogram (VCG) signal. While the case of left ventricular hypertrophy (LVH) is highlighted, the system and methods have greater utility in exploring various spatiotemporal features of any cardiac disease using VCG signals. Beyond the simple technique to diagnose LVH (+) or LVH (-), the present analysis provides the pathophysiological characteristics of the propagation of electricalmotive force, which enables the interpretability of the diagnostic result.

BACKGROUND

[0003] A heart attack is catastrophic especially to a patient with silent heart disease. There are various types of cardiac disease that it would be desirable to detect, diagnose and promptly treat before a catastrophic event occurs abruptly. One of the common traits of such cardiac diseases is the presence of cardiac hypertrophy which is the abnormal enlargement, or thickening, of the heart muscle, resulting from increases in cardiomyocyte size and changes in other heart muscle components, such as the extracellular matrix. Left ventricular hypertrophy (LVH) develops silently over several years without symptoms and can lead to severe problems such as intractable heart failure, potentially lethal arrhythmia or sudden cardiac death. The underlying states resulting in LVH include hypertensive heart disease, valvular heart disease, obstructive and nonobstructive hypertrophic cardiomyopathy.

[0004] LVH produces changes in the QRS complex, the ST segment, and the T wave of an electrocardiogram (ECG), which are formulated as several diagnostic criteria for LVH on the standard 12 lead ECG. Although the most characteristic finding is an increased amplitude of the QRS complex, LVH-related repolarization changes usually occur alone as well as in cases with QRS changes. Recent clinical studies have shown that a Vectorcardiogram (VCG) is advantageous over the standard 12 lead ECG signal analysis for detecting repolarization variability. These clinical studies were disclosed in Muhammad A Hasan and Derek Abbott, “A Review of Beat-to-Beat Vectorcardiographic (VCG) Parameters for Analyzing Repolarization Variability in ECG Signals” in Biomedizinische Technik, Biomedical engineering, 61 (1 ): 3—17 (Feb. 2016), L G Dahlin, C Ebeling-Barbier, E Nylander, H Rutberg, and R Svedj eholm, “Vectorcardiography is Superior to Conventional ECG for Detection of Myocardial Injury after Coronary Surgery”, Scandinavian Cardiovascular Journal, 35(2): 125- 128 (Mar. 2001) and Andres Ricardo Perez Riera, Augusto H Uchida, Celso Ferreira Filho, Adriano Meneghini, Celso Ferreira, Edgardo Schapacknik, Sergio Dubner, and Paulo Moffa, “Significance of Vectorcardiogram in the Cardiological Diagnosis of the 21st century”, Clinical Cardiology. 30(7)319-323 (Jul. 2007.)

[0005] With respect to the diagnostic accuracy of LVH, there has been a longstanding disagreement in the literature as to whether the ECG or VCG is more informative. Some literature finds that the VCG is better and more informative (discussed in Andres Ricardo Perez Riera, Augusto H Uchida, Celso Ferreira Filho, Adriano Meneghini, Celso Ferreira, Edgardo Schapacknik, Sergio Dubner, and Paulo Moffa, “Significance of Vectorcardiogram in the Cardiological Diagnosis of the 21st century”, Clinical cardiology. 30(7) 19-323 (Jul. 2007) and J. David Bristow, “A Study of the Normal Frank Vectorcardiogram. American Heart Journal. 61(2):242-249 (1961). Other literature finds ECG and VCG similar as discussed in Romhilt Donald W., Greenfield Joseph C., And Estes E Harvey, “Vectorcardiographic Diagnosis of Left Ventricular Hypertrophy”, Circulation, 37(1): 15-19 (Jan. 1968). Other literature found that VCG was poorer than ECG as discussed in E.Raymond Borun, John M Chapman, and Frank J Massey, “Electrocardiographic Data Recorded with Frank Leads: In Subjects Without Cardiac Disease and those with Left Ventricular Overload”, The American Journal of Cardiology, 18(5):656-663 (1966).

[0006] A VCG depicts the orientation and strength of an integrated cardiac electromotive force representing a trajectory of instantaneous overall cardiac activity throughout the cardiac cycle in orthogonal three-dimensional space. Bonomini et al. (Maria Paula Bonomini, Fernando Juan Ingallina, Valeria Barone, Ricardo Antonucci, Max Valentinuzzi, and Pedro David Arini, “Left Ventricular Hypertrophy Index Based on a Combination of Frontal and Transverse Planes in the ECG and VCG: Diagnostic Utility of Cardiac Vectors”, Journal of Physics: Conference Series, 705:12031 (2016) studied these VCG features in the transverse and frontal planes to construct LVH indexes and compared their diagnostic performance in the ECG and VCG. Hasan et al. (Muhammad A Hasan and Derek Abbott, “A Review of Beat- to-Beat Vectorcardiographic (VCG) Parameters for Analyzing Repolarization Variability in ECG Signals”, Biomedizinische Technik, Biomedical engineering, 61(1):3- 17 (Feb. 2016) surveyed VCG features such as total cosine R to T, loop area, or azimuth and elevation. Since these approaches were partially successful in improving the diagnostic accuracy of heart diseases, further focusing on spatiotemporal features of VCG which are not directly accessible in the standard 12 lead ECG will be expected.

[0007] Moreover, VCG also allows intuitive interpretation. Specifically, visualizing the VCG in a familiar three-dimensional space helps physicians and users to understand the VCG more intuitively than 12 lead ECG. It can be said that this is a useful feature in the medical field where the explanation of the diagnosis is required. Furthermore, since the number of channels required for the measurement is smaller than that of the 12 lead electrocardiogram, the user load is low and the measurement can be easily performed. As a result, it is expected that the amount of machine learning data can be secured more easily for VCG than for ECG 12 lead systems.

[0008] It is thus desirable to use VCG to diagnose/classify cardiac disease using machine learning and it is to this end that the disclosure is directed.

SUMMARY

[0009] In an embodiment, method of predicting cardiac disease using a vectorcardiography (VCG) signal is provided. The method may include segmenting a vector cardiography (VCG) signal into a plurality of segments; extracting, using ellipsoidal fitting, a plurality of features from each segment wherein the extracted features indicate three-dimensional spatial characteristics of the VCG signal; deploying a machine learning model on the extracted features to predict a cardiac disease associated with the VCG signal; and determining that a subset of the extracted features is strongly correlated with the cardiac disease, the subset of the extracted features configured to be used for a clinical diagnosis of the cardiac disease.

[0010] In another embodiment, a system is provided. The system may include anon-transitory storage medium storing computer program instructions; and one or more processors configured to execute the computer program instructions to cause operations comprising: segmenting a vector cardiography (VCG) signal into a plurality of segments; extracting, using ellipsoidal fitting, a plurality of features from each segment wherein the extracted features indicate three-dimensional spatial characteristics of the VCG signal; deploying a machine leaming model on the extracted features to predict a cardiac disease associated with the VCG signal; and determining that a subset of the extracted features is strongly correlated with the cardiac disease, the subset of the extracted features configured to be used for a clinical diagnosis of the cardiac disease.

BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 illustrates a cardiac disease diagnosis system that uses vectorcardiography (VCG) feature extraction to classify and diagnose cardiac disease;

[0012] Figure 2 illustrates more details of the VCG feature extracting element of the system in Figure 1;

[0013] Figures 3A and 3B illustrate an example of the dataset used for LVH diagnosis and the segments of the VCG signal used;

[0014] Figure 4 illustrates more details of the VCG analytics engine of the system in Figure i;

[0015] Figures 5 A and 5B illustrate a method for cardiac disease classification using spatial VCG features;

[0016] Figures 6A and 6B each illustrate an example of segmentation for ECG and VCG data;

[0017] Figures 7A and 7B each illustrate data from the minimum volume ellipsoid enclosure process of the method in Figure 4; and

[0018] Figure 8 illustrates an example of the minimum volume ellipsoid enclosure process for the different segments and the spatial features of VCG identified and the relative Euler angles among the segments;

[0019] Figures 9A-9D compares the classification performance of VCG versus 12 lead ECG for classifying cardiac disease;

[0020] Figures 10A and 10B show the proposed VCG spatial features and the standard 12 lead ECG criteria for LVH diagnosis; and

[0021] Figures 11A and 1 IB show the distribution of features in LVH and normal labels. DETAILED DESCRIPTION OF ONE OR MORE EMBODIMENTS

[0022] The disclosure is particularly applicable to assessing left ventricular hypertrophy (LVH) using a spatiotemporal feature, extractable from a vectorcardiogram (VCG) but not ECG, and it is in this context that the disclosure will be described and illustrated. However, the present technique is not simply to diagnose LVH (+) or LVH (-) and it will be appreciated that the system and methods have greater utility in exploring various spatiotemporal feature of any cardiac disease using VCG. The present analysis provides the pathophysiological characteristics of the propagation of electrical-motive force.

[0023] Figure 1 illustrates a cardiac disease diagnosis system 10 that uses vectorcardiography (VCG) feature extraction to classify/assess and diagnose cardiac disease. The system 10 may be implemented with cloud computing resources or in a client server architecture and the like. One or more computing devices 12 may capture and/or store VCG data that is then communicated over a communications path 14 to a VCG feature system 16 that outputs a VCG signal based heart analysis result that may be used to classify a particular type of cardiac disease or diagnose a particular type of cardiac disease, including LVH. Each computing device 12 may be a processor based device with memory, a display, connectivity circuits, etc. that can capture VCG data and/or store and communicate the VCG data (that is typically raw VCG data). The computing device 12 may be a smartphone or cellular phone device 12A, a tablet computer device 12B, a wearable device that captures the VCG data or a VCG monitoring device 12N that captures the VCG data from a patient. The communications path 14 may be a wired or wireless communication path so that each computing device 12 communicates with the VCG feature system 16. For example, the communication path 14 may be one or more of Ethernet, the Internet, digital cellular network, cellular digital data network, computer network, digital computer network and the like that uses typical data and communication protocols to connect and exchange data.

[0024] The VCG feature system 16 may be implemented using one or more computing resources, such as cloud computing resources, server computers, blade servers, a personal computer, etc. that perform the operations of the classification/diagnosis of the cardiac disease as detailed in Figures 4-5 discussed below. The VCG feature system 16 may receive the VCG signals, segment the VCG signals into one or more segments such as the example segments in Figure 3B, perform minimum volume ellipsoid enclosure and perform machine learning to classify/diagnose a particular cardiac disease or heart condition. As shown in Figure 2, the VCG feature system 16 may further include storage 200, memory 202 and a processor 204 and a VCG analytics engine 206. In one embodiment, the VCG analytics engine 206 may be implemented as a plurality of lines of computer code/instructions that are stored in the storage 200 or memory 202 and executed by the processor 204 so that the processor and the computer based VCG system 16 are configured to perform the VCG analysis processes shown in Figures 4-5 that are discussed below.

[0025] Figure 4 illustrates more details of the VCG analytics engine 206 of the system in Figure 1. Each element of the VCG analytics engine 206 may also be implemented as a plurality of lines of computer code/instructions executed by the processor so that the computer/processor is configured to perform the operations of the VCG analysis/classification/diagnosis process. The VCG analytics engine 206 may include a model training engine 400 and a segmentation engine 402 that receives the raw VCG data. The VCG analytics engine 206 also may include a VCG feature extraction engine 404 that receives the data from the segmentation engine 402 and a machine learning engine 406 that, based on the data of the extracted VCG features and a cardiac disease labeled data set (for a particular cardiac disease), performs a classification/diagnosis process for the particular cardiac disease. The VCG analytics engine 206 also may have a user interface engine 408 that generates a visual representation/data of the VCG heart analysis result that is returned to a user. If there is a labeled cardiac disease data set for a plurality of cardiac disease, then the machine learning engine 406 can perform classification for each different cardiac disease using the extracted VCG feature data. Alternatively, the system may have a plurality of machine learning models that are each trained to classify a different cardiac disease and the system can perform classification for each type of cardiac disease.

[0026] The model training engine 400 may receive other cardiac disease labeled data or raw VCG data and use that data to train the machine learning model is a known manner to classify/diagnose a particular cardiac disease based on the raw VCG data received. The segmentation engine 402 may segment the raw VCG data into one or more segments, such as the six segments shown in Figure 3B and described in more detail below along with the segmentation method that is discussed below with reference to Figures 5A and 5B. The VCG feature extraction engine 404 may extract one or more features from the raw VCG data wherein the feature extraction method is discussed below with reference to Figures 5A and 5B. The machine learning engine 406 may perform the machine learning process using the cardiac disease labeled data for a particular cardiac disease and the raw VCG data and extracted features to classify data that shows that the particular cardiac disease may be present in the patient. In one embodiment, a known tree-based classifier, such as Random Forest, may be used for the machine learning although it is understood that the system and method may be implemented with other classifiers or other machine learning processes (understanding that the labeled data and training may be different if different machine learning processes are used. The user interface engine 408 may receive the output from the machine learning and generate a visual representation or textual summary of the resulting cardiac disease diagnosis or classification that may be provided to the user who submitted the raw VCG data. Thus, the system and method generate a tangible result (the cardiac disease diagnosis that is presented in human readable form based on the machine learning results) that is presented to a user.

[0027] Figure 5 A illustrates a method 500 for cardiac disease classification using spatial VCG features in which raw VCG data for a patient is input into the method for cardiac disease classification. The method 500 shown in Figure 5A may be performed on each piece of set of raw VCG data submitted. Furthermore, the method 500 may be tuned to classify/diagnose any particular cardiac disease or multiple different cardiac diseases at the same time as described above. The processes shown in Figure 5 A and 5B are the functions/ operations that quantify spatial characteristics of Vectorcardiography (VCG) for a patient in a clinically interpretable manner. The method shown in Figures 5A and 5B may be performed using the system and its features described above, but it may also be performed using other systems and hardware.

[0028] The method may first segment the VCG data (502) into one or more segments. In one embodiment, the well-known P, QRS, and T waves of the heart signals, usually referred to in ECG, may be divided into the following six parts by conventional techniques using wavelet transforms: 1) P onset to P peak (seg pi); 2) P peak to P offset (seg P2); 3) Q or QRS complex onset to R peak (seg n); 4) R peak to S or QRS complex offset (seg ); 5) S offset (T wave onset) to T peak (seg ti); and 6) T peak to T offset (seg t2). Figure 6A shows these six segments mapped onto the typical heart signal for ECG while Figure 6B shows the same six segments on a 3D VCG signal. In other words, each wave composes a loop in three-dimensional space, meaning the onset point should be close to the offset point in the three-dimensional space. Thus, the onset value should be determined by the least Euclidian distance to the offset value. The least Euclidian distance is searched in the range [a, P] where = (the time t at the peak value), a = (P - 2A), A = (the time t of offset - P ). An example of this segmentation method is shown in Figures 6 A and 6B. [0029] The P, QRS, and T waves of the ECG signals are commonly used in clinical practice because they are directly related with the electrical propagations characteristic of heart. Thus, the algorithm to automatically divide these segments from raw ECG signals has been studied and is well known to achieve highly accurate performance. Beyond this well-known division, the disclosed system and method divides these three waves further into six divisions. Based on the observation of each loop, the spatial morphologies are asymmetrical. This is because the first half and the second half of each loop are derived from different heart activities. For an example of the P wave, the electrical signal begins in the sinoatrial node which is located in the right atrium and, travels to the right and left atria, causing them to contract and pump blood into the ventricles. Because of this asymmetric phenomenon, the loop of the P wave can become asymmetric by its nature.

[0030] Using the segments generated by segmentation 502, the method performs feature extraction (504) on those segments as it detailed in Figure 5B and discussed below. Once the feature extraction is completed on a segment, the method determines if there are more segments (506) and loops back to the feature extraction process 504 if there are segments from which features need to be extracted. If there are no more segments, then the extracted VCG segment features are fed into a classification process 508 that uses a cardiac disease labeled data set for a particular cardiac disease and the extracted VCG segment features to determine if there are signs of the cardiac disease in the patient based on the raw VCG data of the patient.

[0031] Figure 5B illustrates further details of the method 500 that receives the raw VCG data 512 from a patient in order to classify cardiac disease for the patient based on the VCG data that has the benefits noted above. The raw VCG data may be pre-processed (514) that may include noise and baseline removal by bandpass filter and standardization and normalization of the VCG signals. The pre-processed VCG signals may then be segmented (516) such as into the six segments noted above. In more detail, segmentation involves Peak, Onset and/or Offset point estimations in the VCG signals using known wavelet methods. For example, the segmentation (516) may use a method described in an IEEE paper by J.P. Martinez et al., “A Wavelet-Based ECG Delineator: Evaluation on Standard Databases”, IEEE Transactions on Biomedical Engineering, Vol. 51, Issue 4 (Apr. 2004) that is incorporated herein by reference. The segmentation (516) may also use a Deep Learning based approach for the segmentation task whose code may be found at github.com/jergusadamec/ecg-deep-segmentation that is also incorporating herein by reference. [0032] During the segmentation process (516), the one or more segments (six in one embodiment) are determined from a heart beat signal of the patient. The first segment is the P onset to P peak segment (seg Pi) discussed above. This is the electrical signal that begins in the sinoatrial node which is located in the right atrium and travels to the right and left atria, causing them to contract and pump blood into the ventricles. This electrical signal is recorded as the P wave on the ECG. The PR Interval is the time, in seconds, from the beginning of the P wave to the beginning of the QRS complex.

[0033] The second segment is the P peak to P offset segment (seg P2) that is the electrical signal that passes from the atria to the ventricles through the atrioventricular (AV) node. The signal slows down as it passes through this node, allowing the ventricles to fill with blood. This slowing signal appears as a flat line on the ECG between the end of the P wave and the beginning of the Q wave. The PR segment represents the electrical conduction through the atria and the delay of the electrical impulse in the atrioventricular node.

[0034] The third segment is the Q onset to R peak (seg ri) and the fourth segment is R peak to S offset (seg ) in which, after the signal leaves the AV node it travels along a pathway called the bundle of His (3) and into the right and left bundle branches (4, 5), the signal travels across the heart’s ventricles causing them to contract, pumping blood to the lungs and the body. This signal is recorded as the QRS waves on the ECG. Because these waves occur in rapid succession they are usually considered together as the QRS complex. The fifth segment is the S offset to T peak (seg ti) and the sixth segment is T peak to T offset (seg t2) which is when the ventricles then recover to their normal electrical state, shown as the T wave. The muscles relax and stop contracting, allowing the atria to fill with blood.

[0035] Once the VCG signals are segmented (an example of which is shown in Figure 6B), the feature extraction process 504 is performed with Figure 5B showing the details subprocesses of the feature extraction. As noted above, the method performs a loop for each segment (518-530) until all of the segments have been processed to extract the features.

[0036] For each segment, processes 520-528 are performed to extract the features for each segment. During the feature extraction process, the method may perform an origin symmetric complement process 520 for each segment because parametrizing the asymmetric shape of each segment (half of torus) may not nicely fit into an ellipsoid enclosure. This process performs a data complementation for symmetric ellipsoid enclosure in which the process 520 determines an origin as the center of the start point and the end point and virtually calculates the position of a point to be origin symmetry with respect to an original point group in a segment. In more detail, the half loop of each segment from the VCG signal is virtually complemented to make a complete loop. Because the original trajectory of each segment is a half loop, applying an ellipsoid enclosure to the original data points causes misfitting. Thus, the preprocessing is applied by adding virtual points that are origin symmetric points to the original data points in the segment. An example of the virtual points for each segment are shown by the dots in Figure 8 while Figures 7A and 7B show the segments before this process (Figure 7A) and after the process (Figure 7B). Figure 8 shows this process for the P wave loop (seg pi and seg P2), for the QRS complex (seg n and seg n) and for the T wave loop (seg ti and seg t2).

[0037] When the full loop for each segment has been generated by process 520, a minimum volume ellipsoid enclosure process (522) is performed for each segment that extract features for each segment to thus extract interpretable features from each segment of VCG in three- dimensional space. In this process, the parametrization of the spatial torus signal by ellipsoid enclosure is performed by first applying the minimum volume ellipsoid enclosure (522). During this process, the method computes the minimum volume covering ellipsoid that encloses N points in a three-dimensional space (an example of which is disclosed by Nima Moshtagh (2021), “Minimum Volume Enclosing Ellipsoid” (www.mathworks.com/matlabcentral/fileexchange/9542-minimum-volume-enclosing- ellipsoid), MATLAB Central File Exchange. Retrieved March 29, 2021, all of which are incorporated herein by reference.

[0038] Once the minimum volume covering ellipsoid is determined, the method may extract rigid transformation features from the minimum volume covering ellipsoid enclosure (MVEE) (up to six features) (524). This process is used because body posture changes the gravity force to the heart, the heart mechanically tilts and the VCG’s spatial information will be altered. The minimum volume ellipsoid enclosure computes the optimal rotation matrix and the center coordinate. This rigid transformation information can determine positional and postural transformations. The positional feature is given by the center coordinates (x, y, z) G R(3) of the estimated ellipsoid. This represents the average voltage within the segment. The postural information is given by the rotation matrix of the ellipsoid. The postural feature, (roll, pitch, yaw) G S(3) angles, can be computed from the rotation matrix. The ellipsoid shape is the components of ellipsoid such as the three axes, the maximum area, and the volume of the estimated ellipsoid. When combined together, 11 features are computed for each segment. Finally, the relative angles between the longest axes of the estimated ellipsoids are computed. The six ellipsoids provide 15 relatives angles by taking the combinations from them. In one embodiment, the features may include the Euler Angle —> 3 features and a transition vector —> 3 features.

[0039] The method may also extract spatial morphological features from the MVEE (526) because each cardiac wave is a circular shape in 3D space and although the shape is asymmetrical, the impact can be mitigated by the segmentation. The minimum volume ellipsoid enclosure computes the morphological feature of the circular shape. In one embodiment, this process generates 5 features that may include maximum volume, maximum space and the three axes of the ellipsoid. Once these processes (524, 526) are completed, the extracted features (11 features extracted from one segment) are aggregated (528).

[0040] Once all of the segments are processed and features extracted and aggregated, the method aggregates a total of 66 features (11 extracted features for each segment for each of the six segments) from one segment (532). The method then performs machine learning (534) using the cardiac disease labeled data set to classify any cardiac disease indicators from the VCG data. In one embodiment, the machine learning model was trained using the dataset and the proposed features for LVH classification problems shown in Figure 3A and a tree classifier, such as random forest classifier, was used. Tree-based machine learning methods are based on decision trees built by recursively splitting a training sample, using different features from a dataset at each node that splits the data most effectively. These unique characteristics of the tree-based method enable model interpretation via feature importance which is calculated as the decrease in node impurity, how well the trees split the data, and by weighting by the probability of reaching that node. The node probability or feature importance can be calculated by the number of samples that reach the node, divided by the total number of samples. Figure 8 illustrates an example of the minimum volume ellipsoid enclosure process for the different segments for training data and the spatial features of VCG identified for each segment and the relative Euler angles among the segments.

[0041] The above-described system and method can be useful for cardiac state prediction. By applying machine learning to heart disease prediction problems using this proposed method, it becomes possible to classify diseases by explicitly considering spatial characteristics of VCG. The system and method may also be used for clinical diagnosis criterion discovery in that, by employing a decision tree-based model as a machine learning model, it is possible to quantify what spatial features of which VCG segments are significantly problematic (e.g., having a strong correlation with) in a particular cardiac disease. It is noted that the handcrafted features disclosed herein are superior to self-learning features such as deep learning and has the advantage of explainability to the patients in the clinical field.

TESTING AND RESULTS

[0042] For testing the above method, a target dataset was collected from the known PTB-XL in PhysioNet as discussed in Patrick Wagner, Nils Strodthoff, Ralf-Dieter Bousseljot, Dieter Kreiseler, Fatima I Lunze, Wojciech Samek, and Tobias Schaeffler, “PTB-XL, a Large Publicly Available Electrocardiography Dataset”, Scientific Data, 7(1): 154 (2020) and Ary L Goldberger, Luis AN Amaral, Leon Glass, Jeffrey M Hausdorff, Plamen Ch Ivanov, Roger G Mark, Joseph E Mietus, George B Moody, Chung-Kang Peng, and H Eugene Stanley, “Physiobank, Physiotoolkit, and Physionet: Components of a New Research Resource for Complex Physiologic Signals”, Circulation. 101(23):e215-e220 (2000).

[0043] The open dataset includes heart disease labels and standard 12 lead ECG for 10 seconds from a subject at 500[Hz], In this study, a single heartbeat was considered as a sample, resulting in approximately 10 data samples from a subject. First, VCG was derived from 12 lead ECG using Kors regression transformation which was reported as the most accurate VCG approximation among five different transformation methods as discussed in Rene Jaros, Radek Martinek, and Lukas Danys, “Comparison of Different Electrocardiography with Vectorcardiography Transformations”, Sensors. 19(14):3072 (2019). Next, the labeled data of LVH and normal were selected by filtering PTBXL as follows. The LVH labeled data, 100% LVH and 0% for the other types, were selected. Likewise, the normal labeled data were selected, which has only 100% normal for the label. Both labeled data were also limited to the data verified by at least one cardiologist. The above process filtered the original dataset into 61 subjects with the LVH label and 5874 subjects with the normal label. This class ratio of the obtained labeled data is highly imbalanced. Concerning the imbalanced class which badly affects the performance of machine learning, two undersampled datasets were derived for evaluation in this study as shown in Figure 3 A. The undersampled data were chosen randomly.

RESULTS

[0044] The results of LVH classification on the test dataset is shown in Figures 9A-9D. The outliers were determined if the result was outside of the range of 1.5 times the interquartile range above the upper quartile and below the lower quartile. In evaluating the LVH 122 data set with 95% confidence interval, the average accuracy and the macro average Fl score result in 0.904 (CI: 0.861-0.947) and 0.903 (CI: 0.860-0.946) for the proposed VCG spatial feature, and 0.867 (CI: 0.813-0.920) and 0.866 (CI: 0.812-0.919) for the 12 lead ECG feature as shown in Figures 9A and 9C. Likewise, in LVH 366 dataset, the average accuracy and the macro average Fl score are 0.932 (CI: 0.918-0.946) and 0.872 (CI: 0.838-0.906) for the proposed VCG spatial feature, and 0.935 (CI: 0.913-0.957) and 0.885 (CL0.826-0.943) for the 12 lead ECG feature as shown in Figures 9B and 9D. The statistical significance was not recognized due to the high variance.

Feature Importance for LVH Classification

[0045] The top-ranked 30 of feature importance of the trained Random Forest using LVH 366 data are shown in Figures 10A and 10B.

[0046] The foregoing description, for purpose of explanation, has been with reference to specific embodiments. However, the illustrative discussions above are not intended to be exhaustive or to limit the disclosure to the precise forms disclosed. Many modifications and variations are possible in view of the above teachings. The embodiments were chosen and described in order to best explain the principles of the disclosure and its practical applications, to thereby enable others skilled in the art to best utilize the disclosure and various embodiments with various modifications as are suited to the particular use contemplated.

[0047] The system and method disclosed herein may be implemented via one or more components, systems, servers, appliances, other subcomponents, or distributed between such elements. When implemented as a system, such systems may include and/or involve, inter aha, components such as software modules, general-purpose CPU, RAM, etc. found in general- purpose computers,. In implementations where the innovations reside on a server, such a server may include or involve components such as CPU, RAM, etc., such as those found in general-purpose computers.

[0048] Additionally, the system and method herein may be achieved via implementations with disparate or entirely different software, hardware and/or firmware components, beyond that set forth above. With regard to such other components (e.g., software, processing components, etc.) and/or computer-readable media associated with or embodying the present inventions, for example, aspects of the innovations herein may be implemented consistent with numerous general purpose or special purpose computing systems or configurations. Various exemplary computing systems, environments, and/or configurations that may be suitable for use with the innovations herein may include, but are not limited to: software or other components within or embodied on personal computers, servers or server computing devices such as routing/connectivity components, hand-held or laptop devices, multiprocessor systems, microprocessor-based systems, set top boxes, consumer electronic devices, network PCs, other existing computer platforms, distributed computing environments that include one or more of the above systems or devices, etc.

[0049] In some instances, aspects of the system and method may be achieved via or performed by logic and/or logic instructions including program modules, executed in association with such components or circuitry, for example. In general, program modules may include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular instructions herein. The inventions may also be practiced in the context of distributed software, computer, or circuit settings where circuitry is connected via communication buses, circuitry or links. In distributed settings, control/instructions may occur from both local and remote computer storage media including memory storage devices.

[0050] The software, circuitry and components herein may also include and/or utilize one or more type of computer readable media. Computer readable media can be any available media that is resident on, associable with, or can be accessed by such circuits and/or computing components. By way of example, and not limitation, computer readable media may comprise computer storage media and communication media. Computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and can accessed by computing component. Communication media may comprise computer readable instructions, data structures, program modules and/or other components. Further, communication media may include wired media such as a wired network or direct-wired connection, however no media of any such type herein includes transitory media. Combinations of the any of the above are also included within the scope of computer readable media.

[0051] In the present description, the terms component, module, device, etc. may refer to any type of logical or functional software elements, circuits, blocks and/or processes that may be implemented in a variety of ways. For example, the functions of various circuits and/or blocks can be combined with one another into any other number of modules. Each module may even be implemented as a software program stored on a tangible memory (e.g., random access memory, read only memory, CD-ROM memory, hard disk drive, etc.) to be read by a central processing unit to implement the functions of the innovations herein. Or, the modules can comprise programming instructions transmitted to a general-purpose computer or to processing/graphics hardware via a transmission carrier wave. Also, the modules can be implemented as hardware logic circuitry implementing the functions encompassed by the innovations herein. Finally, the modules can be implemented using special purpose instructions (SIMD instructions), field programmable logic arrays or any mix thereof which provides the desired level performance and cost.

[0052] As disclosed herein, features consistent with the disclosure may be implemented via computer-hardware, software, and/or firmware. For example, the systems and methods disclosed herein may be embodied in various forms including, for example, a data processor, such as a computer that also includes a database, digital electronic circuitry, firmware, software, or in combinations of them. Further, while some of the disclosed implementations describe specific hardware components, systems and methods consistent with the innovations herein may be implemented with any combination of hardware, software and/or firmware. Moreover, the above-noted features and other aspects and principles of the innovations herein may be implemented in various environments. Such environments and related applications may be specially constructed for performing the various routines, processes and/or operations according to the invention or they may include a general-purpose computer or computing platform selectively activated or reconfigured by code to provide the necessary functionality. The processes disclosed herein are not inherently related to any particular computer, network, architecture, environment, or other apparatus, and may be implemented by a suitable combination of hardware, software, and/or firmware. For example, various general-purpose machines may be used with programs written in accordance with teachings of the invention, or it may be more convenient to construct a specialized apparatus or system to perform the required methods and techniques.

[0053] Aspects of the method and system described herein, such as the logic, may also be implemented as functionality programmed into any of a variety of circuitry, including programmable logic devices ("PLDs"), such as field programmable gate arrays ("FPGAs"), programmable array logic ("PAL") devices, electrically programmable logic and memory devices and standard cell-based devices, as well as application specific integrated circuits. Some other possibilities for implementing aspects include: memory devices, microcontrollers with memory (such as EEPROM), embedded microprocessors, firmware, software, etc. Furthermore, aspects may be embodied in microprocessors having software-based circuit emulation, discrete logic (sequential and combinatorial), custom devices, fuzzy (neural) logic, quantum devices, and hybrids of any of the above device types. The underlying device technologies may be provided in a variety of component types, e.g., metal-oxide semiconductor field-effect transistor ("MOSFET") technologies like complementary metal- oxide semiconductor ("CMOS"), bipolar technologies like emitter-coupled logic ("ECL"), polymer technologies (e.g., silicon-conjugated polymer and metal-conjugated polymer-metal structures), mixed analog and digital, and so on.

[0054] It should also be noted that the various logic and/or functions disclosed herein may be enabled using any number of combinations of hardware, firmware, and/or as data and/or instructions embodied in various machine-readable or computer-readable media, in terms of their behavioral, register transfer, logic component, and/or other characteristics. Computer- readable media in which such formatted data and/or instructions may be embodied include, but are not limited to, non-volatile storage media in various forms (e.g., optical, magnetic or semiconductor storage media) though again does not include transitory media. Unless the context clearly requires otherwise, throughout the description, the words "comprise," "comprising," and the like are to be construed in an inclusive sense as opposed to an exclusive or exhaustive sense; that is to say, in a sense of "including, but not limited to." Words using the singular or plural number also include the plural or singular number respectively. Additionally, the words "herein," "hereunder," "above," "below," and words of similar import refer to this application as a whole and not to any particular portions of this application. When the word "or" is used in reference to a list of two or more items, that word covers all of the following interpretations of the word: any of the items in the list, all of the items in the list and any combination of the items in the list.

[0055] Although certain presently preferred implementations of the invention have been specifically described herein, it will be apparent to those skilled in the art to which the invention pertains that variations and modifications of the various implementations shown and described herein may be made without departing from the spirit and scope of the invention. Accordingly, it is intended that the invention be limited only to the extent required by the applicable rules of law. [0056] While the foregoing has been with reference to a particular embodiment of the disclosure, it will be appreciated by those skilled in the art that changes in this embodiment may be made without departing from the principles and spirit of the disclosure, the scope of which is defined by the appended claims.

Claims

What is claimed is:

1. A method of predicting cardiac disease using a vectorcardiography (VCG) signal, the method comprising: segmenting a vector cardiography (VCG) signal into a plurality of segments; extracting, using ellipsoidal fitting, a plurality of features from each segment wherein the extracted features indicate three-dimensional spatial characteristics of the VCG signal; deploying a machine learning model on the extracted features to predict a cardiac disease associated with the VCG signal; and determining that a subset of the extracted features is strongly correlated with the cardiac disease, the subset of the extracted features configured to be used for a clinical diagnosis of the cardiac disease.

2. The method of claim 1, further comprising: segmenting a plurality of training VCG signals into a corresponding plurality of training segments; extracting, using ellipsoidal fitting, a plurality of training features from each training segment wherein the extracted training features indicate three-dimensional spatial characteristics of the training VCG signal; and training the machine learning model on the extracted training features.

3. The method of claim 1, wherein the cardiac disease comprises left ventricular hypotrophy (LVH).

4. The method of claim 1, wherein the plurality of segments comprise: a first segment comprising P wave onset to P wave peak in the VCG signal; a second segment comprising P wave peak to P wave offset in the VCG signal; a third segment comprising QRS complex onset to R peak in the VCG signal; a fourth segment comprising R peak to QRS complex offset in the VCG signal; a fifth segment comprising T wave onset to T peak in the VCG signal; and a sixth segment comprising T peak to T wave offset in the VCG signal.

5. The method of claim 1, wherein the three-dimensional spatial characteristics comprise spatial morphological features comprising, for an estimated ellipsoid of the ellipsoidal fitting, maximum volume, maximum area, and three axes.

6. The method of claim 1, wherein the three-dimensional spatial characteristics comprise rigid transformation features comprising, for an estimated ellipsoid of the ellipsoidal fitting, three-dimensional center coordinates (x, y, z) and postural features (roll, pitch, yaw).

7. The method of claim 1, wherein the ellipsoid fitting comprises performing an origin symmetric complement process.

8. The method of claim 7, wherein the origin symmetric complement process comprises: determining an origin; and for each original data point, adding a virtual point that is symmetric about the origin to the original data point, such that half loop of a corresponding segment is virtually complemented to make a complete loop.

9. The method of claim 8, wherein the ellipsoid fitting further comprises applying a minimum volume ellipsoid enclosure around the complete loop.

10. The method of claim 1 , wherein the machine learning model comprises a tree classifier.

11. A system comprising: a non-transitory storage medium storing computer program instructions; and one or more processors configured to execute the computer program instructions to cause operations comprising: segmenting a vector cardiography (VCG) signal into a plurality of segments; extracting, using ellipsoidal fitting, a plurality of features from each segment wherein the extracted features indicate three-dimensional spatial characteristics of the VCG signal; deploying a machine learning model on the extracted features to predict a cardiac disease associated with the VCG signal; and determining that a subset of the extracted features is strongly correlated with the cardiac disease, the subset of the extracted features configured to be used for a clinical diagnosis of the cardiac disease.

12. The system of claim 11, wherein the operations further comprise: segmenting a plurality of training VCG signals into a corresponding plurality of training segments; extracting, using ellipsoidal fitting, a plurality of training features from each training segment wherein the extracted training features indicate three-dimensional spatial characteristics of the training VCG signal; and training the machine learning model on the extracted training features.

13. The system of claim 11, wherein the cardiac disease comprises left ventricular hypotrophy (LVH).

14. The system of claim 11, wherein the plurality of segments comprise: a first segment comprising P wave onset to P wave peak in the VCG signal; a second segment comprising P wave peak to P wave offset in the VCG signal; a third segment comprising QRS complex onset to R peak in the VCG signal; a fourth segment comprising R peak to QRS complex offset in the VCG signal; a fifth segment comprising T wave onset to T peak in the VCG signal; and a sixth segment comprising T peak to T wave offset in the VCG signal.

15. The system of claim 11 , wherein the three-dimensional spatial characteristics comprise spatial morphological features comprising, for an estimated ellipsoid of the ellipsoidal fitting, maximum volume, maximum area, and three axes.

16. The system of claim 11 , wherein the three-dimensional spatial characteristics comprise rigid transformation features comprising, for an estimated ellipsoid of the ellipsoidal fitting, three-dimensional center coordinates (x, y, z) and postural features (roll, pitch, yaw).

17. The system of claim 11, wherein the ellipsoid fitting comprises performing an origin symmetric complement process.

18. The system of claim 17, wherein the origin symmetric complement process comprises: determining an origin; and for each original data point, adding a virtual point that is symmetric about the origin to the original data point, such that half loop of a corresponding segment is virtually complemented to make a complete loop.

19. The system of claim 18, wherein the ellipsoid fitting further comprises applying a minimum volume ellipsoid enclosure around the complete loop.

20. The system of claim 11, wherein the machine learning model comprises a tree classifier.

- 21 -