High-frequency QRS in biometrics
Through the extraction of high-frequency QRS signals and deep learning technology, bioassay signatures are generated, which solves the problems of easy forgery of existing biometric technologies and limited application of ECG signals, and achieves more reliable and secure identity recognition.
Patent Information
- Application Number
- CN201980064545.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2018-10-01
- Filing Date
- 2019-09-27
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2039-09-27
AI Technical Summary
The existing biometric technology is prone to forgery, especially based on visible or audible features, and the widespread application of ECG signals is limited by the inconvenience and insufficient reliability of 12-lead detection, which leads to its application in identity recognition being hindered. .
Using high-frequency QRS (HF-QRS) signals as biological characteristics, stable and differentiated features are extracted from ECG signals through signal acquisition, filtering, digital signal processing and deep learning technologies, and a biometric signature is generated, and a database is used to perform the Verify and identify.
It improves the reliability and security of biometrics, reduces the risk of forgery, and achieves more accurate identity recognition. In addition, the number of features and data volumes of HF-QRS signals are much higher than that of traditional ECG signals, enhancing the accuracy of identification.
Smart Images

Figure CN112804937B_ABST
Abstract
Description
Background Art
[0001] Most biometric technologies used today rely on individual characteristics, including voice recognition, fingerprints, retina and facial recognition, and are based on identifiable visible or audible features that are therefore easy to forge because the technology that is already widely available and commercially available allows accurate reproduction and modification of these features, namely CCD cameras, 3D printers and digital recording. However, the instruments used to reproduce or modify modified electrical signals are not widely available. Generating an electrocardiograph (or ECG) requires the presence of a living and biologically present individual at the identification location, which makes it difficult to forge. In addition, the ECG signal has all the properties of a secure biometric: universality, measurability, uniqueness and permanence.
[0002] ECG signals can be transmitted for recording using conventional surface electrodes that are typically mounted on the subject's chest. The ECG signal consists of several components that represent different functional phases during each heartbeat and are projected according to the electrical orientation of the generating tissue. Due to variations in cardiac tissue structure and orientation between individuals, there are a variety of subject-specific details in the ECG signal.
[0003] The ECG has been extensively studied as a potential biometric, but the inconvenience and lack of reliability of the 12-lead detection required have prevented widespread use of the ECG. Although it is possible to use fewer leads and look for unique portions of the ECG signal or a derivative of the ECG signal, it is difficult to use the ECG signal to find reliable markers because the subject's pulse rate varies (this variation occurs on a continuous basis).
[0004] The electrical signal consists of a series of PQRST complexes, and most PQRST sequences are inconsistent. The time interval between two consecutive R signal peaks, called the RR interval, corresponds to the heart pulse, which usually has a rate in the range of 60 to 90 beats per minute (bpm). The P signal corresponds to atrial depolarization (right side, first, then left side). The larger QRS complex corresponds to ventricular depolarization and atrial depolarization. The T signal corresponds to ventricular repolarization. Sometimes a weaker U signal appears.
[0005] A "wave" includes a curve covering at least one complete segment (P, Q, R, S and / or T). The amplitude of a straight line extending between two consecutive signals within a time increment, such as the amplitude of a straight line from the end of an S wave to the beginning of the immediately following T wave, is called a "segment". In a graph that is at least partially curved, a time increment that includes at least one wave, such as from the beginning of a Q wave to the end of an S wave, is called an "interval".
[0006] The QRS interval, which is typically 50 to 100 milliseconds (msec) in length, represents the conduction time from the start of ventricular depolarization until the end of ventricular depolarization and includes the propagation of the electrical impulse through the ventricular muscle. The P wave signal is typically gradually rounded and has a length of approximately 50 to 110 msec. A QRS interval greater than approximately 120 msec typically indicates a ventricular arrhythmia or a blockage in one of the bundles.
[0007] When the ECG was first developed for clinical use, low-pass filters were used to remove high-frequency "noise" from the signal, making it easier to evaluate the shapes of the various waves and the lengths of the intervals in the displayed output. The high-frequency component of the QRS complex (HF-QRS) in the ECG signal - a unique part of the ECG signal and distinct from the ECG signal itself - was largely forgotten until NASA researchers demonstrated that it could be used to more accurately determine the health of heart tissue, especially as it relates to the electrical conduction system, because it more accurately represents the unique electrical conduction system of each individual heart. See U.S. Patents Nos. 8,924,736 and 7,539,535, both of which are incorporated by reference. The amount of data present in the HF-QRS signal significantly exceeds that present in the standard low-frequency ECG used today, and several patterns identified by NASA in the data were shown to be relatively stable from month to month (T. T. Schlegel et al., “Real-time 12-lead high-frequency QRS electrocardiography for enhanced detection of myocardial ischemia and coronary artery disease,” Mayo Clin Proc, March 2004, vol. 79, pp. 339-50).
[0008] Compared to other currently monitored biometric identification methods, such as fingerprints, faces, conventional ECG readings, or retinal vascular patterns, the number of features in the HF-QRS signal, and the amount of data that can be extracted from the HF-QRS signal, is much greater. Therefore, using the HF-QRS signal as a biometric feature can allow identification to be made more reliable. Summary of the invention
[0009] Advances in signal acquisition allow for the generation of standard ECG signals from one or more sensors, preferably electrodes (U.S. Patent No. 9,833,158 is incorporated by reference), including embedding sensors into smartphones or other devices (U.S. Publication No. 2018 / 0116538, U.S. Patent No. 8,615,290 are both incorporated by reference). One or more sensors can be on the surface of a smartphone or device to establish a connection using only minimal contact with the user, such as electrodes contacting the user through a fingertip, rather than electrodes connected by an adhesive or electrodes that require penetration of the skin. The sensor can also be remote and not in contact with the user - provided that the sensor can be electrically connected to the subject. The signal can also be collected and stored by a cardiac-related implantable device such as a pacemaker, atrial fibrillation monitor, or defibrillator; or obtained from any portable or wearable device such as a watch, smartphone, or glasses. See U.S. Patent No. 8,019,321 (incorporated by reference). The sensor can also be embedded in a pacemaker, atrial fibrillation monitor, or defibrillator.
[0010] Additional optional security measures include verifying that the real-time HF-QRS signal (or derived data) from an individual is not identical to the individual's stored biometric signature, as identity or close homology would indicate duplication of existing data, as well as network penetration or other fraud. Other fraud detections may also be included, such as measuring bioimpedance plethysmography and / or body heat to verify that a human subject is providing the sample.
[0011] The HF-QRS signal may be derived by well-known methods, including bandpass filtering the ECG signal, preferably within the range of about 80 Hz to 300 Hz, more preferably within the range of about 150 Hz to 250 Hz, or even within a wider range to allow measurement of the HF-QRS. The filtered signal may be further amplified and filtered to optimize the signal-to-noise ratio, and may be digitized and / or analyzed using a unique digital signal processing algorithm (described below) to retrieve information necessary to identify the identity of the individual in question. The extremely dense HF-QRS signal may be further defined and limited by functions, including determining the root mean square (RMS) voltage within a specified period, or other derived functions, including HFQE (described further below) to quantify HF-QRS variations; or HFAV (described further below) to reduce sensitivity to small changes in signal voltage.
[0012] In the present invention, HF-QRS signals from multiple subjects (in some embodiments, one of the values derived from the above, such as RMS voltage, HFQE or HFAV) are sampled multiple times to derive the following features or values: (i) the feature or value is similar enough in all samples for the same subject, but (ii) the feature or value has sufficient differences between different subjects to allow the above features or values to be identified. The set of these features or values is a value derived using, for example, a deep learning convolutional neural network and a HF-QRS signal or a function of a HF-QRS signal. Suitable features may be present in a specific HF-QRS signal or region thereof, or in particular a value derived from a HF-QRS signal. In one case, a feature or value that meets condition (i) also meets condition (ii) above; in other cases, the features and values that meet conditions (i) and (ii) are different.
[0013] The search for such a feature or value can be performed by first establishing a training set of such features or values using the features or values of a group of subjects, and sampling the HF-QRS representation of each subject in the group multiple times. The training set can then be compared to a validation set (and / or cross-validated) to identify overfitting and perform additional validation of appropriate features or values. One then searches for such features or values, and assuming that the searched features or values are appropriately derived and therefore sufficient to satisfy criterion (i), but insufficient to satisfy criterion (ii), then such features or values establish a signature for the individual.
[0014] Each individual's biometric signature is stored in a database for later identification. For identification, the stored biometric signature is compared to features or values of a real-time live HF-QRS signal acquired from the individual to identify the individual or determine a mismatch with anyone in the database, and optionally, alert others in the event of a "match" and "mismatch" (or both). The real-time signal may also be temporarily or transiently stored for later use.
[0015] With recent advances in signal processing, big data analysis, and machine learning techniques, HF-QRS signals from multiple individuals can be analyzed to find features that meet criteria (i) and (ii) and used to generate a unique biometric signature. A preferred approach is to fuse two or more signals or signal derivatives, such as fusing a spectrogram with the original signal. Various signal derivative functions can also be used to determine values that differ sufficiently between individuals to generate a signature.
[0016] After finding features that satisfy conditions (i) and (ii) and using these results to generate a biometric signature, another step is to verify the stability of the signature over time. This can be done with a group of test subjects whose potential features (selected as above) are monitored over time. For any changes (instability) in such identifiers, one would quantify the changes over time and then determine a deviation threshold for a theoretical subject, where identifiers within the deviation threshold are accepted for an individual; identifiers outside the deviation threshold are rejected. Ultimately, all individuals whose identifiers are in the database should also be monitored regularly, and if there are significant changes (the individual's signature is placed outside the deviation threshold), the stored identifiers should be updated.
[0017] Each time an individual's HF-QRS is determined and sent for confirmation, there may be variations in the resulting data. Preferably, variability correction data is generated at the time of the individual's biometric signature and used during verification. The variability correction data may include ambiguity resolution data for reducing the sensitivity of values or features obtained from the individual and / or error correction data used by error detection and correction algorithms to correct errors in features or values (see also U.S. Patent No. 7,882,363, which is incorporated herein by reference).
[0018] Biometrics are already widely used in various industries, including healthcare, security, forensics, finance, and consumer products. Digital keys (using HF-QRS as described in this article) can be used for any Internet-related security application or any other type of security privilege or access.
[0019] The invention includes determining features or values of HF-QRS signals or features or values of HF-QRS functions or signal derivatives thereof, which features or values are individually or cumulatively characterized as follows: (i) being sufficiently similar in all samples for the same subject, but (ii) being sufficiently different for different subjects to allow identification; and (iii) being sufficiently stable over time. The invention also includes comparing the HF-QRS signal of an individual or his / her HF-QRS function or derived value with signals, functions or derived values sampled separately from the individual; and identifying the individual or determining no match with anyone in a database, and optionally, signaling to others or storing one or both of the match and no match in a database. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 is a flow chart showing how to use HF-QRS signals or derived features or values to find an individual's biometric signature through machine learning.
[0021] Figure 2A This is the first part of a flow chart for using the biometric signature of a HF-QRS signal or derived features or values for security related applications, such as providing access rights or privileges to an individual.
[0022] Figure 2B yes Figure 2A The second part of the flowchart.
[0023] Figure 3A The testing phase is outlined in a flowchart showing how to find and verify the Figure 1 the stability of the selected biometric signature over time; and establishing a deviation threshold for the signature.
[0024] Figure 3B The use of the selected identifier in identifying the subject; and the continued monitoring of the subject's signature are outlined in a flow chart.
[0025] Figure 4 Depicted are steps on how a specific unique feature can be extracted from the HF-QRS signal to be used as a biometric signature. DETAILED DESCRIPTION
[0026] Figure 1 is a flow chart showing how to use a neural network to find a biometric signature in a HF-QRS signal or a feature or value derived from a HF-QRS signal. The neural network allows one to find a signal, feature or value that is (i) sufficiently similar for samples collected at different times for the same individual; and (ii) sufficiently different between all individuals regardless of when the sample was taken to serve as a biometric signature. Methods for finding data sets with such similarities and differences are known and can be accomplished using any of a variety of neural networks capable of performing deep learning. Similarly, it is known to build a classifier and use regularization to reduce overfitting. The classifier can be a single-class SVM classifier, or can be any of naive Bayes, logistic regression, support vector machine, Bayesian network, multilayer perceptron, deep neural network, random forest, or Adaboost.
[0027] Figure 2A and Figure 2BFlow chart of the following steps: obtaining HF-QRS signal and deriving from the HF-QRS signal, storing individual's biometric signature, then comparing the biometric signature with the real-time signal for identification, and allowing certain actions by the correctly identified individual or issuing an alarm to the monitor (server or individual) if there is a view to violate security. The first step is to use the HF-QRS signal of the individual to be identified or the derived features or values to establish a data set representing the individual's biometric signature. The signal or the derived features or values can be filtered and fused (as described below). When performing identity verification, the individual's real-time signal, feature or value is compared with the stored biometric signature, and if the individual's real-time signal, feature or value passes the similarity threshold, the individual will be identified and the identified individual will be allowed to obtain access rights or privileges. On the contrary, if there is no match, an alarm can be issued to the monitor (server or individual).
[0028] Figure 3A A first step, the testing phase, is shown, in which signatures are selected that remain stable over time. A deviation threshold is also determined in the first step. The deviation threshold accepts signatures that vary within the deviation threshold range and rejects signatures that vary beyond the deviation threshold range; thus, it is conceivable that the signatures vary to some extent over time.
[0029] Figure 3B A second step is shown in which the selected signature determined to be stable over time is used. If within the deviation threshold, the subject's signature is accepted and the subject is granted access or privileges.
[0030] Methods for determining similar image features are described in US Pat. No. 7,729,545 (incorporated by reference), and may be applied to the signals or fused signals, or other conventional similarity measures may be applied to the derived functions or values.
[0031] As an optional step, bioimpedance plethysmography (see U.S. Pat. No. 5,732,710, which is incorporated by reference) or thermal sensors, which measure electrical impedance in tissue, can be used to confirm that the transmitted signal, characteristic, or value is of human origin. Bioimpedance can be measured by having the subject touch electrodes with different fingers (or other parts of the body), then sending a weak electrical signal from one electrode and measuring the impedance of the other electrode, and then determining whether the impedance corresponds to the characteristic impedance of the human body. Thermal sensors can additionally or alternatively verify that the subject has human body temperature characteristics. The electrodes or sensors can be located in a fixed location, such as a door entry panel, or can be included in an individual's personal device, including in a mobile phone.
[0032] As another optional step, the degree of homology between the stored features or values and the real-time features or values may indicate a match, mismatch, or attempted fraud. It is contemplated that the HF-QRS signal stored as a biometric signature is not identical to a HF-QRS signal obtained in another sampling.
[0033] Identity would indicate duplication of an existing signal, function or value set as well as fraud. Homology thresholds can be established to distinguish between matches, mismatches and identities (fraud).
[0034] Another option is to examine the high frequency portion of the ECG, such as the data rich PR and ST segments of the HF-QRS, so variations on different samples can always be assumed. Homology thresholds for these segments can also be established to distinguish between matches, mismatches and identities (frauds).
[0035] Verification data may be stored on the user's device, in a local area network, or on the Internet. Verified user data may act as a digital key that provides access, privileges, or other security-related operations, however, the identity may alert monitors to fraud or attempted intrusion. A digital key may be provided on the user's device so that a verified individual may use a digitally encrypted device to provide access or privileges to any of a number of items.
[0036] Derived Features and Values
[0037] Figure 4 Depicted are steps showing how to extract one specific such unique feature from the HF-QRS signal for use as a biometric signature. Figure 4 In step (a), an averaged or unaveraged ECG signal is first determined. The ECG signal is determined to be within a 2mV range. From the signal, a bandpass filter (most preferably, within a range of about 150Hz to 250Hz) is used to separate the HF-QRS signal, in which case the signal is within a 10μV range within 90 milliseconds. From the derived HF-QRS signal, the profile of the original HF-QRS signal can be used as a biometric signature of the individual; alternatively, the amplitude reduction region within the signal (see U.S. Patent No. 7,539,535, U.S. Publication No. 2018 / 0116538, both of which are incorporated by reference, which describe monitoring HF-QRS signals and finding amplitude reduction regions) can be separated. The profile of the original HF-QRS signal and the signal amplitude reduction portion can be fully distinguished between individuals, but for the same individual, the profile of the original HF-QRS signal and the signal amplitude reduction portion are similar in multiple samplings, thereby allowing the signature to be extracted from the original HF-QRS signal; as described herein.
[0038] In looking for unique features, it may be useful to enhance the HF-QRS signal by changing the scale or changing the attenuation or gain of a particular wave. This may also be applied to the PR and ST segments of the ECG in cases where they are monitored for increased safety as noted above.
[0039] Another value that can be derived from the HF-QRS signal and used for the biometric signature is the RMS voltage value. The RMS voltage value is derived from the filtered HF-QRS signal by the following steps (see Xue et al., page 240, column 2): using the spatial velocity of the three leads to determine the QRS onset and offset (reference point) of the average unfiltered waveform (uqon, uqoff), and measuring the unfiltered QRS duration (UQRSD = uqoff-uqon). Preferably, the data of each lead is filtered using a spectral filter (150Hz to 250Hz). With VM = (fX 2 +fY 2 +fZ 2 ) 1 / 2 The vector magnitude (VM) of the three filtered signals (fX, fY, and fZ) is calculated. The average noise level (AVNL) can be measured in the ST segment in a 40ms interval 60ms offset from the unfiltered QRS. The filtered QRS reference points (fqon and fqoff) can be determined in the VM waveform using a threshold equal to twice the measured AVNL, and the filtered QRS duration (FQRSD = fqon-fqoff) can be measured. RMSV can be calculated for all samples in the VM waveform, VM1, over the duration of the FQRSD as follows:
[0040]
[0041] Xue et al. also determined a new value HFQE to quantify the change in HF-QRS. First, the RMSV is calculated using the unfiltered QRS reference point and the unfiltered QRS duration (uqon, uqoff, and UQRSD) instead of the filtered reference point and the filtered QRS duration as in Equation 1. The HFQE reference point is calculated based on the filtered signal using the unfiltered QRS plus sample padding corresponding to the additional duration of 10ms before uqon and 10ms after uqoff:
[0042]
[0043] Xue et al. also determined a third value, HFAV, to reduce sensitivity to small changes in signal voltage using the following function:
[0044]
[0045] Xue et al. also noted that the unfiltered QRS duration was more stable than the filtered QRS duration for all groups they tested.
[0046] Any of RMSV, HFQE, and HFAV, or other suitable derived features or values, may be used in the biometric signatures described herein.
[0047] Aversano et al. describe data processing steps, wherein the RMS voltage of a very dense HF-QRS signal can optionally be further qualified and limited by: (i) Fourier transforming the data and attenuating low frequency harmonics, then performing an inverse Fourier transform to obtain a filtered signal, thereby determining the onset and offset of the QRS, and then determining the RMS voltage of the entire QRS complex; (ii) as a further optional step, determining the local peaks and valleys of the filtered waveform, and finding the envelope of the high frequency ECG by tracing the connections between the peaks and valleys; (iii) also optionally, determining the coefficient of variation and the average standard deviation for a group of subjects in order to determine appropriately high confidence limits (e.g., 95%) for the variability of the HF-QRS root mean square (RMS) voltage, so that RMS voltage results outside the confidence limits can be disregarded or ignored. Such RMS voltages can be used for the biometric signatures described herein.
[0048] The initial signal used to derive the RMS voltage value can be any representation of the HF-QRS, including unique features that can be preferentially determined by deriving the spectrogram of the HF-QRS signal (I. Odinaka et al., "ECG Biometrics: a robust short-time frequency analysis," WIFS'2010, December 12-15, 2010; J. da Silva Luz et al., "Learning Deep Off-the-Person Heart Biometrics Representations," IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 13, pp. 1258-70 (2018)). The spectrogram is a visual representation of the signal energy as a function of frequency and time. If it is not desired to derive the RMS voltage value and related functions, but to use the spectrogram itself, it is preferred in one example to fuse the spectrogram with the low-frequency original signal (see J. da Silva Luz et al., supra) so that a single heartbeat is represented by two separate views, thereby reducing system errors. Another option is to manually derive the RMS voltage and related values (HFQE and HFAV) or other values or features from the time series or spectrogram, or both, or to provide unsupervised derivation (in a sense "automatic") for the time series, spectrogram, or both. Furthermore, further analysis can even be performed on the spectrogram itself or the original signal itself to extract unique features rather than extracting features from the fused result.
[0049] Deriving and using characteristics or values
[0050] Unsupervised deep learning algorithms or manually identifying features can be used to extract features or values that allow individuals to be distinguished. A deep learning algorithm can be used to find one or more classifiers by fitting into a training set using a spectrogram, a low-frequency original signal or a fusion result (a spectrogram with a low-frequency original signal) or RMS voltage and related values (HFQE and HFAV) or other values or features. See V. Jindal et al., "An Adaptive Deep Learning Approach for PPG-Based Identification", IEEE pp. 6401-04 (2016). Alternatively, such common features can be manually identified and extracted. A preferred deep learning algorithm is a convolutional neural network (CNN), which can learn depth (several layers) and distinguish representations and classifiers directly from the data itself given a large amount of data. Other suitable deep learning algorithms include deep neural networks ("DNN"), long short-term memory recurrent neural networks ("LSTM-RNN"), or convolutional long short-term memory deep neural networks ("CL-DNN").
[0051] The same features or values in the training set collected at different times or different subjects or collected at different times and in different subjects are used as validation sets or test sets, i.e., one of the following: the fusion result (spectrogram with low-frequency original signal) or RMS voltage and related values (HFQE and HFAV) or other features or values. The validation set is used to tune the hyperparameters of the classifier and avoid overfitting, for example, by using the validation set for regularization. In a simple example, regularization is by stopping training early when the error on the validation dataset increases, because this is a sign of overfitting to the training dataset. Overfitting can also be reduced by using dropout, which regularizes the objective function by randomly setting a certain proportion of nodes in the neural network to zero (ignoring them) during training.
[0052] Other modifications used to enhance the results include data augmentation to force the network to recognize the heartbeat even when some parts are impaired (pulse or physiological changes of the subject). This modification consists in duplicating the heartbeats of the training dataset in other versions of the heartbeat, where the new versions are generated by changing the scale or by varying the attenuation or gain of a specific wave by a specified percentage. J. da Silva Luz et al., supra.
[0053] Then, as in Choi et al., “Biometric Authentication Using Noisy Electrocardiograms Acquired by Mobile Sensors”, IEEE pp. 1266-73 (2016), the signal features or values are classified by an appropriate one-class classifier, where a one-class support vector machine (SVM) classifier is preferred, but one-class classifiers may also include: Naive Bayes, Logistic Regression, Support Vector Machine, Bayesian Network, Multilayer Perceptron, Deep Neural Network, RBF Network, Guided Aggregation Algorithm (Bagging), Random Forest, and Adaboost. Choi et al. designed a cascaded bandpass filter for noise elimination and proposed eight benchmark characteristics. For classification-based authentication, they used a radial basis function-based, kernel-based support vector machine, which showed the best performance among nine classifiers through experimental comparison. After classification, the selected one-class machine learning algorithm will determine whether the input function or value matches the input function or value in the database. Those input functions or values that do not match can be measured again.
[0054] The embodiments, parts, steps, features, purposes, benefits and advantages discussed are merely exemplary and not restrictive. All articles, patents, patent applications and other publications cited in this disclosure are incorporated herein by reference. Relational terms such as "first" and "second" can be used only to distinguish one entity or action from another entity or action, without requiring or implying any actual relationship or order between one entity or action and another entity or action. The terms "comprise", "including", "comprising" and any other variations thereof are intended to be used in conjunction with the element list in the specification or claims to indicate that the list is not exclusive and may include other elements. Similarly, in the absence of further restrictions, elements starting with "one" or "a kind of" do not exclude the presence of other elements of the same type. The present invention is limited only in the appended claims and includes all equivalents of the elements in the claims.
Claims
1. A method for identifying / authenticating an individual, the method comprising: Sampling HF-QRS signals or derived features or values from a group of subjects, wherein each sampling is performed at a plurality of different times for each subject; Using a deep learning neural network to determine features or values that (i) are sufficiently similar for samples collected at different times for the same individual; and (ii) are sufficiently different between all individuals regardless of when the samples were collected to serve as a biometric signature, wherein the features or values used as the biometric signature are determined by first establishing a training set of features or values from the group of subjects, wherein the HF-QRS of each subject is sampled multiple times; comparing the training set to a validation set for identifying overfitting of the classifier and for additional validation of appropriate features or values; Sampling the features or values from an individual and storing the features or values in a database as biometric signatures corresponding to respective features or values; and The feature or value from the unknown individual is later sampled and a determination is made as to whether the feature or value matches any of the stored biometric signatures, wherein if the feature or value from the unknown individual and the stored feature or value satisfy a similarity threshold, then the feature or value from the unknown individual is determined to match the stored biometric signature; and wherein variability correction stored data is incorporated in the sampling of the feature or value from the unknown individual.
2. The method according to claim 1, wherein: The method also includes allowing the unknown individual to gain access rights or privileges if the biometric signature of the unknown individual matches the stored biometric signature.
3. The method according to claim 1, wherein: The method also includes issuing an alert to a monitor if the biometric signature of the unknown individual does not match the stored biometric signature.
4. The method according to claim 1, wherein: The deep learning neural network is one or more of a convolutional neural network, a deep neural network, a long short-term memory recurrent neural network, and a convolutional long short-term memory deep neural network.
5. The method according to claim 1, wherein: The method also includes comparing the training set to a validation set for verifying the features or values used as a biometric signature.
6. The method according to claim 1, wherein: The method also includes reducing overfitting by regularization.
7. A method for finding a biometric signature using a HF-QRS signal, the biometric signature being stable over time, the method comprising: Sampling HF-QRS signals or derived features or values from a group of subjects, wherein each sampling is performed at a plurality of different times for each subject; A deep learning neural network is used to determine features or values that (i) are sufficiently similar for samples collected at different times for the same individual; and (ii) are sufficiently different between all individuals regardless of when the samples were collected to serve as a biometric signature, wherein the features or values used as the biometric signature are determined by first establishing a training set of a set of features or values from the group of subjects, wherein the HF-QRS of each subject is sampled multiple times; comparing the training set to a validation set for identifying overfitting of the classifier and for additional validation of appropriate features or values, and wherein the features or values used as the biometric signature are also determined by: applying bandpass filtering to the HF-QRS signal and / or deriving other functional representations of the HF-QRS or a region of the HF-QRS, the other functional representations comprising one or more of RMS voltage, HFQE, and HFAV; monitoring selected biometric signatures of a group of test subjects over time to look for stable one or more signatures that are stable over time; determining a deviation threshold based on the monitoring such that the test subject's stable signature falls within the deviation threshold over the measured time period; Building a database of stable signatures obtained from actual subjects; and monitoring the stable signature of the actual subject at least at the frequency of the time period; and A signature from an unknown subject is sampled, and if the signature is within the deviation threshold, it is determined that the signature matches one of the stored signatures, wherein the variability-corrected stored data is incorporated in the sampling of the signature from the unknown subject.
8. The method according to claim 7, wherein: The method also includes further defining and limiting the HF-QRS signal by: (i) performing a Fourier transform on the data and attenuating low frequency harmonics, then performing an inverse Fourier transform to obtain a filtered signal, determining the onset and offset of the HF-QRS from the filtered signal, and then determining the root mean square voltage of the entire HF-QRS complex wave; and / or (ii) determining local peaks and valleys of the filtered waveform and finding the envelope of the high frequency ECG by tracing the connections between the peaks and valleys; and / or (iii) determining the coefficient of variation and average standard deviation of the HF-QRS signals, features or values from a group of subjects to determine the variability in the HF-QRS signals, features or values from the group of subjects.
9. The method according to claim 7, wherein: Bandpass filtering is used for frequencies from 150 Hz to 250 Hz.
10. The method according to claim 7, wherein: The deep learning neural network is one or more of a convolutional neural network, a deep neural network, a long short-term memory recurrent neural network, and a convolutional long short-term memory deep neural network.
11. The method according to claim 7, wherein: The method also includes comparing the training set to a validation set for verification of features or values used as biometric signatures.
12. The method according to claim 7, wherein: The method also includes reducing overfitting through regularization.
13. The method according to claim 7, wherein: The classifier is a one-class SVM classifier.
14. The method according to claim 13, wherein: The classifier is one of Naive Bayes, Logistic Regression, Support Vector Machine, Bayesian Network, Multilayer Perceptron, Deep Neural Network, Random Forest, and Adaboost.
15. The method according to claim 7, wherein: The method also includes reducing overfitting by dropout.
Citation Information
Patent Citations
Method and apparatus for detecting electrocardiographic abnormalities based on monitored high frequency QRS potentials
US20180116538A1
Method and device for stable impedance plethysmography
US5732710A
Real-time, high frequency QRS electrocardiograph with reduced amplitude zone detection
US7539535B1
Signal processing method and method for determining image similarity
US7729545B2
Biometric authentication system
US7882363B2
