JP5642210B2 - Method and apparatus for electronic biometric identification recognition - Google Patents

Description

  This application claims the benefit of PCT application WO 2004/012388, filed July 4, 2003, and US provisional application 60 / 39,883, filed July 29, 2002. The entire disclosures of which are incorporated herein by reference.

  Numerous including automated banking services, e-commerce, e-banking, e-investment, e-data protection, remote access to resources, e-transactions, work security, anti-theft devices, crime recognition, safety entry, and workplace entry registration Identification recognition plays an important role in the life identification phase.

In computerized systems, passwords and personal identification numbers (PINs) are often used for user identification. However, in order to maintain security, the password must be changed periodically, which places a burden on the user. Similarly, signature verification methods suffer from other shortcomings including counterfeiting and fraud. See, for example, US Pat.

  As a result, identification and recognition systems that utilize the measurement of individual biological phenomena—biometrics—have grown in recent years. Used alone or integrated with other technologies such as smart cards, cryptographic keys, digital signatures, biometrics are expected to reach almost every aspect of the business community and our daily lives.

  Several advanced techniques have been developed for biometric recognition including fingerprint recognition, retinal iris recognition, face recognition, and speech recognition. For example, U.S. Patent No. 6,053,086 by Shockley et al. Outlines the use of biometric data such as fingerprint recognition to authorize computer access for individuals. U.S. Pat. No. 6,053,098 by Scheidt et al. Describes identification and authentication using biometric data. U.S. Pat. No. 6,053,096 to Dulude et al. Describes the use of fingerprints, palm recognition, retinal iris scans, and speech patterns as part of a biometric authentication certificate. U.S. Pat. No. 6,057,031 to Murakami et al. Describes “physiological and histological markers” including infrared radiation for biometric authentication. However, these types of technologies only penetrate limited markets due to complex and unfriendly acquisition procedures, sensitivity to environmental parameters (such as lighting conditions and background sounds), and high cost. Furthermore, because of complex acquisition procedures, the techniques described above typically require operator attention.

  Fingerprint recognition is an established and most mature technology. However, there are some drawbacks. Fingerprint recognition systems cannot verify the presence of the fingerprint owner's body and thus tend to be erroneous, limiting their suitability for online use cases. Optical sensors are expensive and fragile devices and are generally incompatible with the consumer market. Moreover, the system has negative implications for criminology.

  Retinal scanning technology is characterized by high performance. However, they require high-precision optical sensors, and furthermore, they are not easy to use because they require maneuvering of the head position and manipulation of the human eye, which is a very sensitive organ. Optical sensors are expensive and fragile.

  The iris and face recognition system is an easy-to-use technology because it takes images from a distance and is non-intrusive. However, they require digital photographic equipment and are more sensitive to lighting conditions, pupil size changes, and facial expressions. In addition, the performance of iris recognition may be degraded by the use of dark glasses or contact lenses, and facial recognition may be confused by imitation of the face.

  Speech recognition is the most easy-to-use technique, however, it requires a low noise setting and is very sensitive to the inherent variable speech parameters including intonation. Furthermore, existing conventional recording techniques can be used to deceive speech-based recognition systems.

  Thus, there is a need for an identity recognition system that can be used standalone or integrated with existing security systems that is reliable, robust, difficult to deceive (online or offline), low cost and easy to use. A request exists.

  Over the years, electrocardiogram (“ECG”) measurements have been used for a number of different purposes. The electrocardiogram signal is an electrical signal generated by the heart and can be collected utilizing a prior art surface electrode usually mounted on the patient's chest. The electrocardiogram signal is composed of a plurality of components indicating various functional stages between individual heart beats, and protrudes according to the potential direction of the power generation tissue.

  In an electrocardiogram signal due to normal changes in cardiac tissue structure, cardiac direction, and electrical tissue direction, an individual will show various patient-specific details, all of the cardiac tissue structure, cardiac direction, and electrical tissue direction. Affects the ECG signal measured from the extremities. Many types of systems take advantage of these patient-specific changes.

  For example, U.S. Patent No. 6,053,009 by Blazey et al. Describes using an electrocardiogram signal to assess or profile an individual's physiological and cognitive status. Speaking of identification, Non-Patent Document 1 compares a patient's electrocardiogram with a pre-registered electrocardiogram characteristic parameter for identification. Wiederhold, in US Pat. No. 5,836,831, proposes to identify patients using electrocardiogram signals collected directly or remotely, and “searches” for feature extraction to identify individuals. Present a “preliminary analysis”.

  However, the electrocardiogram signal is composed of electrocardiogram components having characteristics that can be common to one group. None of these documents describe a system or method that removes common characteristics of ECG components to form a signature for patient identification. Thus, there remains a need for systems and methods that identify individuals by these attributes.

  The presentation of the aforementioned references in this background art is not an admission that they are prior art or similar art with respect to the invention disclosed herein. However, all of the references in this background description are incorporated by reference as if fully set forth herein.

US Pat. No. 5,892,824 US Pat. No. 5,534,855 US Pat. No. 6,490,680 US Pat. No. 6,310,966 US Pat. No. 6,483,929 US Pat. No. 6,293,904 US application 2003013509

2001 Conference Paper at the 23rd Annual International IEEE Conference of Medicine and Biology Society (in Istanbul, Turkey), entitled “Development of an ECG Identification System” by Kyoso et al. “Electric potential probes-new directions in the remote sensing of the human body” (Sci. Technol. 13 (2002) 163-169).

  Applicants provide solutions with various devices and methods having multiple forms to the aforementioned problem of biometric identification.

In the first form, each of the aforementioned problems relating to biometric identification is solved by using the following method and its variants. That is,
Creating and storing a first biometric signature that identifies the particular individual by forming a difference between the representation of the heart beat pattern of the particular individual and the stored representation of the common characteristics of the heart beat patterns of the plurality of individuals Steps,
After the creating step, obtain a representation of the selected individual's heart beat pattern and between the stored representation of the selected individual's heart beat pattern and the common characteristics of the plurality of individual heart beat patterns. Forming a difference to create a second biometric signature;
Comparing the second biometric signature with the first biometric signature and determining whether the selected individual is the specific individual.

  The system according to this aspect includes an electrocardiogram signal acquisition module, an electrocardiogram signal processing module including an electrocardiogram signature generator, and an output module.

  Thus, according to this first aspect, the systems and methods disclosed herein convert bioelectric signals into unmatched electronic biometric signatures. The incomparable ECG signature makes it difficult to deceive the system, and the inherent robustness makes it ideal for remote as well as local and online use. Furthermore, biometric signature based systems are characterized by high discrimination performance and support both open and closed search modes.

  In one preferred method according to the first aspect, the stored representation of one or more ECG components measures and stores that representation for multiple individuals and averages all of the stored representations Can be obtained. On the other hand, the common characteristics can be acquired by techniques such as principal component analysis, fuzzy clustering analysis, and wavelet decomposition.

  Since this first electrocardiogram method is robust, there is another advantage. They allow a simple and straightforward collection technique that can be implemented as a low-cost, user-friendly collection device, and further eliminates the need for skilled operators.

  According to variations on these systems and methods, the common characteristic of the electrocardiographic component of one or more patients is an analytical model of the common characteristic of one or more electrocardiographic components instead of or in addition to the empirical model. Can be removed. Similarly, the stored representation is first classified into subgroups, common characteristics within at least one subgroup are identified, patient characteristics are classified according to the subgroups, and one or more subgroup ECG components are identified from the patient signal. Common characteristics can be eliminated by forming patient signatures excluding common characteristics and calculating patient signature correlations to subgroup signatures to identify patients.

  Multiple templates can also be maintained for individual patients, such as by storing multiple signatures that are individually generated at various heart rates. In this embodiment, the patient signature may be correlated with an appropriate template, such as for an appropriate heart rate. Thus, in a variation, the systems and methods disclosed herein may use multiple signature templates that identify individuals in a range of situations and reactions. Meanwhile, or in addition, according to the first embodiment, the patient signal and the registration signal may be normalized based on the heart rate.

  According to the second form disclosed herein, the process for identification may set a dynamic threshold. This dynamic threshold may be based on the desired level of confidence in the identification, as determined by the confidence score.

  According to a third aspect disclosed herein, the systems and methods disclosed herein may utilize a “Q factor” to determine whether to reduce signal contamination due to noise. Similarly, Q-factor or other signal quality measurements may be utilized to determine the length of patient sample required to identify the patient with a desired confidence level. It may also be used to register samples with a desired confidence level so that the samples are appropriate for future comparisons.

  In another embodiment relating to “Q factor” calculation, the systems and methods disclosed herein calculate patient signatures and / or standard deviations of registered signatures for noise and calculate them. From this, it may be determined whether the signal property is appropriate for identification.

  Similarly, the systems and methods disclosed herein may measure the impedance of contacts or probes to determine signal properties. The signal property measurement according to this configuration can also be used to inform the patient to adjust the contact with the sensor or probe or to adjust the placement relative to the sensor or probe.

  According to a fourth aspect, patient and database signatures may be encrypted as a security warning against unauthenticated access and use of the signatures.

  According to a fifth form, the electrocardiogram signal can be collected by electrodes placed in contact with a body part that produces a consistent signal. For a certain body arrangement, even a slight change in electrode arrangement can cause a drastic change in the received signal form and can cause a clear signal component to be generated or eliminated. Thus, in this form, the systems and methods disclosed herein may utilize an electrode placement site that produces a patient-specific consistent signal that is robust despite variations in electrode placement within the site. These sites include arms and legs (including fingers and toes). The robustness of electrode placement within these sites arises from consistent ECG signal emission that does not vary as long as the electrode is close to the limb tip.

  According to this same fifth form, one sensing probe, known as a very high impedance sensing probe, may be used to collect signals, including signals from a single body point such as a finger. On the other hand, these very high impedance probes can remotely sense the electrocardiogram signal, thereby eliminating electrode placement difficulties while maintaining signal consistency.

  According to a sixth aspect, the systems and methods disclosed herein may include elements and steps that protect against registration fraud and reduce the likelihood that a database registrant will not be able to represent patient identification. .

  According to a seventh aspect, the systems and methods disclosed herein can identify a patient by comparing the patient's fitness score with a database registrant's fitness score.

  According to an eighth aspect, the systems and methods disclosed herein may use weighted correlation techniques to attribute different weights to different ECG signal components to generate signatures. On the other hand, or in addition, the signatures may be normalized using various metrics including root mean square calculation or L1 metrics.

  The systems and methods according to each of the foregoing forms preferably perform their tasks automatically to identify recognition. Further, these systems and methods can be incorporated into a wide range of devices and systems. A few non-limiting examples are as follows: Smart card, passport, driver's license device, biologon identification device, personal digital assistant (PDA), cell implantation identification device, anti-theft device, electrocardiogram monitor device, e banking device, e transaction device, pet identification device, physical access device , Logical access devices, and devices that combine electrocardiograms and fingerprint monitors, blood pressure monitors, and / or other forms of biometric devices.

  Further, the systems and methods disclosed herein compare the width of a patient's QRS complex, or more generally, the patient's QRS-related signature components to those of a registered group, or an analytical ECG model, etc. Can be used to identify an individual's age.

  In another example, the systems and methods disclosed herein identify an individual being administered, such as by registering and calculating a series of drug-related signature templates, or by analytically deriving. Can be used. This method can also be used to identify or capture patients attempting to deceive the system by using drugs to alter their ECG signals.

  Building and room access control, surveillance system access, wireless device access, control and user verification, mobile phone activation, computer access control (via laptop, PC, mouse, and / or keyboard), data (including document control) Access, public transport passenger recognition, elevator access control, firearm locking, vehicle control systems (including ignition start and via door lock), smart card control and smart card credit verification, online line (including copyright protection work) Access to components, electronic tickets, nuclear material access and control, robot control, aircraft access and control (passenger identification, flight control, Maintenance worker access), vending machine access and control, Londro matte washer / dryer access and control, locker access, children inoperable lock, TV and / or video access control, decryption key access and use, cash Slot machine, slot machine maintenance access, game console access (including online transaction functions), computer network security (including network access and control), POS buyer identification, online transactions (including customer identification and account access), cash Payment service or wire transfer identification, building maintenance access and control, and implant medical device programming controller The other usage example of using systems and methods disclosed herein is intended to include. Other applications will be apparent to those skilled in the art and are within the scope of this disclosure.

  For any application, an apparatus according to any or all of the above-described forms may operate continuously or on demand. The device may be configured to obtain a representation of the heartbeat pattern of the selected individual by contacting one or more electrodes with the individual or by providing a sensor spaced from the individual. When the device is provided in a smart card, the device can be operated for a limited time after successive recognitions and then inoperable until the next successive recognition is performed. The device may be configured to operate with an encryption key or a digital signature.

  With respect to the methods disclosed herein, the method steps described above may be performed sequentially or in a different order. The systems and methods disclosed herein may be utilized for humans or for patients of another animal.

  Each of these forms may be used interchangeably, interchangeably and in combination. Additional embodiments as well as modifications, variations, and extensions are also described herein.

  The systems and methods disclosed herein may provide a solution to the aforementioned problem of biometric identification.

FIG. 2 is a simplified block diagram of a system for use with the forms disclosed herein, comprising a signal acquisition module, a signal processing module, and an output module. FIG. 2 is a block diagram of a form of a signal acquisition module of the system of FIG. FIG. 2 is a block diagram of a signal processing module configuration of the system of FIG. 1. It is a graph which shows the total average electrocardiogram signal waveform calculated from the database of 20 patients. FIG. 5 shows a group of 10 patient ECG signal waveforms that join the database and contribute to the average waveform of FIG. FIG. 6 shows an ECG signature waveform or group of templates derived from the signal waveform of FIG. FIG. 5 shows a scatter plot and distribution histogram of squared correlation values for 20 patients who contributed to the total average waveform of FIG. Fig. 4 shows a table of z-scores based on the desired degree of confidence in the identification cutoff. The correlation distribution is shown. The distribution of Z conversion correlation is shown. An identification performance curve (static) is shown. An identification performance curve (dynamic) is shown. The nature of the signal as a function of NSR is shown. Fig. 5 shows the fitness score distribution as a function of signal nature for a 5 second split. Fig. 5 shows the fitness score distribution as a function of signal nature for a 20 second split. The fitness score as a function of the recording period (for Q = 0.8) is shown. Shows the fitness score as a function of recording period (for Q = 0.5). 1 shows a functional element diagram of a preferred system. Figure 2 shows a functional element diagram of a preferred signal processor.

[Definition]
Unless otherwise indicated, the terms “identify”, “identify” and “identify” have the meanings of “proving identity”, “proving identity” and “proof of identity”, respectively. Including concepts.

  “Closed search” means a search in which a single stored signature is examined to verify the identity of the individual.

  “Open search” means a search where a plurality of stored signatures are examined to identify a patient.

[First embodiment]
According to a first form, bio-electrical signals are collected, processed and analyzed to identify personal identification. A preferred embodiment of the system and method according to this first aspect is shown by way of example in FIG. FIG. 1 shows a system called electro-biometric identification (E-BioID). In this preferred form, the stored representation of the common characteristics of one or more ECG components of a plurality of individuals is an average of the one or more ECG components of those individuals. However, other embodiments utilize stored representations of various types of common characteristics. For example, it is achieved by principal component analysis, fuzzy clustering analysis, wavelet decomposition, or obtained by an analysis model.

  In a preferred embodiment, the basic elements of the E-BioID system include a signal acquisition module 12, a signal processing module 14, and an output module 16, which are mounted in a single housing. In another preferred embodiment, the system is provided for remotely analyzing locally acquired electro-biometric signals. Each of the elements shown in FIG. 1 can be readily implemented by one of ordinary skill in the art based on principles and techniques already known in combination with the present disclosure.

  FIG. 2 shows a preferred configuration of the signal acquisition module in the E-BioID system. The data acquisition module preferably includes one or more sensors 22, a preamplifier 24, a bandpass filter 26, and an analog to digital (A / D) converter 28. Each of these elements can be easily implemented by one of ordinary skill in the art based on principles and techniques already known in combination with the present disclosure.

  The sensor 22 may be of any type that can detect a beating pattern. For example, a metal plate sensor connected as an “additional device” of a standard computer keyboard may be used. In another form, a single sensor may acquire a signal from a single contact point, such as touching itself with a finger. On the other hand, the sensor may not touch the patient at all.

FIG. 3 shows the preferred elements of the signal processing module 14 in the E-BioID system. The signal processing module preferably includes a digital signal processor (DSP) 32, a dual port RAM (DPR) 34, an electrically erasable programmable read only memory (E ² PROM) 36, and an I / O port 38. Each of these elements can be easily implemented by those skilled in the art based on principles and techniques already known in combination with the present disclosure. The signal processing module 14 is connected to the signal acquisition module 12 and the output module 16 via the port 38.

  In another embodiment, the signal processing module can be implemented on a personal computer by a suitable program. The personal computer is an extensible computer platform that allows easy integration of the system into existing computer facilities in a home, office or university / business environment.

  The output module 16 is preferably composed of a dedicated display unit such as an LCD or CRT monitor, and may include a repeater for the operation of an external electrical device such as a locking mechanism. On the other hand, the output module may include a communication line that relays the recognition result to a remote location for further activity.

[Signal acquisition, processing and analysis]
The bioelectric signal, i.e. the heartbeat signal, is obtained in a simple way, in which the patient is instructed to contact at least one sensor for a few seconds. One or more sensors, which may be metal plates, direct the bioelectric signal to the amplifier 24, which amplifies the bioelectric signal to the desired voltage range. In a preferred embodiment, the voltage range is 0-5 volts.

  The amplified signal passes through filter 26 to remove contributions outside the preferred frequency range of 4-40 Hz. On the other hand, a wider range of 0.1-100 Hz may be utilized with the notch filter to reject major frequency interference (50/60 Hz). The digitization of the signal is preferably done by the 12-bit A / D converter 28, preferably at a sampling frequency of about 250 Hz.

  In module 14, the signal is normalized with the magnitude of the 'R' peak to account for signal magnitude variables generally related to extrinsic electrical characteristics. The normalized data is converted into an electronic-biometric signature that is compared to a pre-stored electronic-biometric template. The comparison result is quantified, optionally assigned a confidence level, and transmitted to the output module 16. The output module 16 feeds back the recognition contents to the user of the E-BioID system, and may activate an external device such as a lock or siren, a virtual device such as a network login confirmation, or a communication link.

  On the other hand, or in addition, the signal may be normalized to the pulse rate. This is useful because the electrocardiogram signal is affected by variations in pulse rate, a well-known electrocardiogram variation. Pulse rate variation can result in latent amplitude and shape variations of the 'P' and 'T' components with respect to the 'QRS' component of the ECG signal (these components are shown in FIG. 5). However, pulse rate variation can be compensated automatically by retrospective pulse rate driven adjustment of the signal complex. In addition, the adaptive operating mode of the system can track and compensate for pulse rate induced variation. This can be done by compressing or expanding the time scale of one cycle of the heartbeat waveform. More sophisticated prescriptions that define the relationship between waveform characteristics (eg, ST, PQ segment duration) and pulse rate may also be utilized. Thus, this variation method is based on electrocardiogram signal identification, and the analysis is performed in synchrony with the heartbeat to remove characteristics common to the general population, thereby making it an electro-biometric, or Biometrics and signatures are constructed to extend patient-specific characteristics that are not normally detectable with raw ECG signals.

  In another embodiment, the E-BioID system is implemented as a fully integrated compact device with many functional elements implemented in an ASIC-based system.

  In another embodiment, the device can be incorporated into a watch worn on the wrist, in which case the signal is measured between the wrist of the arm on which the watch is worn and another arm of the wearer. The back of the watch is made of a conductive medium that contacts the back of the wrist (eg, a metal plate), and the surface of the watch is provided with another metal plate that needs to be touched with the fingers of another hand. The watch may indicate confirmation of the wearer's identity and / or transmit a signal to activate a physically or logically locked device such as a door, computer, safe or the like. The watch may also transmit the wearer's personal information.

[Principle of operation]
Biometric recognition requires that a newly acquired biometric signature be compared to a signature template in a registered or recorded biometric signature template database. Here, two phases of system operation are required: registration and recognition.

[Registration phase]
In a preferred embodiment, each new patient is instructed to touch the first sensor with the left hand finger and touch another sensor with the right hand finger. In another embodiment, the patient may touch a sensor, usually made of metal, at other parts of the body, preferably the hands or feet. In another embodiment, the patient may touch a single sensor at a single point on the body. On the other hand, the patient may not touch the sensor. The system preferably monitors and starts recording the patient's pulse rate and continues for at least 20 seconds. Depending on the level of accuracy required, shorter intervals can also be used. When recording is complete, the system performs a self-test that compares the consistency of the signatures by comparing at least two biometric signatures derived from the two parts of the registered segment. The two sites may be two halves or two larger overlapping segments. Two sites are used to derive two biometric signatures. If the self-test result is successful, the patient is enrolled, and if not successful, the procedure is repeated. Successful records are used to construct an ECG signal or series of ECG signals that are added to the ECG signal database.

  The ECG signal is converted into a set of ECG signature templates by removing characteristics common to all or a subset of patients participating in the data set, thereby enhancing patient characteristic identification characteristics.

  In a preferred embodiment, the system forms an overall average electrocardiogram template, which is calculated by a synchronous average of the normalized ECG signals of patients in the entire pool. The total average represents the common characteristics described above, and thus the difference of the total average from each of the ECG signals will result in a clear, patient-specific set of electronic biometric template signatures. In other embodiments, other means of removing common properties such as principal component analysis, fuzzy clustering analysis or wavelet decomposition may be utilized.

  On the other hand, common characteristics are removed by using an analytical model for common characteristics of one or more ECG components, rather than using empirical models calculated from registered data. Can be done.

  In another preferred embodiment, the database is divided into a plurality of subsets so as to increase the similarity within the subsets and the gaps between the subsets. In this embodiment, an apparent total average or other common characterization for one or more subsets is calculated. This database partitioning can itself be done using standard pattern classification schemes such as linear classification, Bayesian classification, fuzzy classification or neural networks. For large databases, it is possible to partition the database into subsets, not only to ensure the validity of the total average as an appropriate representation of the similarity between ECG signals, but also to simplify and shorten the search process. Useful. Patient signatures will be generated by removing common characteristics found within the appropriate subgroups.

  FIG. 4 shows an example of a total average composed of a pool of 20 patients added to the database.

  FIG. 5 shows 10 examples of an electrocardiogram signal, and FIG. 6 shows an electro-biometric template signature derived from the electrocardiogram signal by removing characteristics common to all patients included in the database. In particular, the individual signatures of FIG. 6 are obtained by subtracting the waveform of FIG. 4 from the corresponding signal of FIG. It can be seen that even though the original ECG signals are quite similar, the derived electro-biometric signatures are significantly different. It can be seen that these differences inherently reflect the unique ECG disparity that underlies the recognizability of the E-BioID system.

[Recognition phase]
In the recognition phase, the patient is involved in the system as in the registration phase, however, a shorter recording time on the order of a few seconds is sufficient.

  In the preferred embodiment, the system performs a matching procedure (closed search). The system processes the acquired signals and forms an electro-biometric patient signature by removing common characteristics found in the entire database, found in a partitioned subgroup of the database, or provided by an analytical ECG model. The signature is adjusted according to the pulse rate, and the adjusted electronic-biometric signature is compared to the patient's registered electronic-biometric signature template.

  In another preferred embodiment, the system performs an identification procedure (open search). The system repeats the comparison process for the entire database or database sub-groups to identify matching identifications.

[Comparison process]
In a preferred embodiment, the comparison is performed by calculating a correlation coefficient ρ between the electron-biometric signature σj and the electron-biometric signature template Φi. It becomes as follows.

The correlation coefficient keeps its original sign while being squared: η = sign (ρ) * | ρ | ² . In another embodiment, the comparison may be based on other similar measurements, such as RMS error between the electronic-biometric signatures.

  Depending on the closed or open search or mode of operation, the comparison yields one or more correlation coefficients. In the closed search, the square correlation coefficient (η) whose code is maintained is used to make a recognition determination. A value greater than a given threshold is accepted as a positive identification or match. Borderline, values near the threshold may indicate the need to extend or repeat the recording. In an open search, if the highest coefficient is above the selected threshold, the largest sign-preserving square correlation coefficient of all the sign-preserving square correlation coefficients will give the most likely patient identification. Will occur.

  The preset threshold is derived from the required confidence level. A higher desired confidence level requires a higher threshold. In one embodiment, a sign preserving square correlation value greater than 0.8 indicates a conforming characteristic and a sign preserving square correlation value less than 0.7 indicates a non-conforming characteristic. Therefore, a sign-maintaining square correlation value greater than 0.8 can be considered a true fitness value, and a value less than 0.7 can be considered a non-match value.

  The upper graph of FIG. 7 shows a scatter plot of the sign-maintained square correlation value, with the 0.8 threshold marked with a dotted line. A clear separation between the fit value (circle) and the non-fit (star) is obvious. The histograms of the other two graphs show different views of the effective recognition ability of the E-BioID system, where non-conforming values are concentrated near zero (no correlation) and the conforming value is 1.0 (absolute correlation). It can be seen that it is densely distributed nearby.

  In another embodiment, a more sophisticated decision scheme such as a multi-parameter scheme (eg, a fuzzy logic scheme) may be utilized, which multi-parameter scheme utilizes one or more distance measurements. For example, multiple correlation values can be derived from a partitioned data analysis.

  In a preferred embodiment, the system improves its performance over time by adding an electrocardiogram signal to the patient database file when changes in the signal occur. For subsequent recognition, the system processes the newly acquired signal, calculates the pulse rate, forms an electronic-biometric signature, and selects the registered electronic-biometric signature template with the most similar pulse rate. The new electronic-biometric signature is then compared to the selected registered electronic-biometric signature template.

  In another preferred embodiment, the system uses signals collected during long-term system operation to track possible variations in the registered patient's electrocardiogram signals and if consistent variations occur If so, the registered signal is automatically adjusted to reflect these variations. This tracking process compensates for gradual variations in the ECG signal over time, however, it does not compensate for rapid, sensitive variations such as those expected in connection with clinical cardiac conditions. In another embodiment, such sensitive fluctuations may be reported to a patient indicating a request for medical consultation.

[Second form]
Biometric identification methods benefit from the proper determination of the identification threshold. The identification threshold can be derived from a correlation analysis between the target signature and the registered database signature. The threshold value can be determined using the distribution of empirical data for optimal discrimination performance. However, the fixed threshold implicitly assumes deterministic signatures and static noise, while in practice the signatures are variable and the noise also depends on external influences that are almost unpredictable. Thus, including the first form, the biometric identification method is adversely affected by signal and noise variations in the database and test readings. In general, this results in reduced correlation for both conformance and nonconformity.

  Therefore, in the second embodiment, the biometric identification method and system, including the one according to the first embodiment, can use a dynamic threshold that can compensate for the effects of signal fluctuations and noise interference. In this form, a dynamic data independent identification threshold is presented. In a preferred embodiment, the dynamic threshold is recalculated for each identification trial using a statistical approach that normalizes the correlation data and thereby allows the quantified statistically significant identification threshold to be calculated. The The threshold has been shown to be resistant to variable signal and noise conditions.

  The preferred method according to this second aspect is based on the determination of confidence bounds for a correlation-based score between a trial signature and a set of registered signatures. Confidence limits define the upper and lower limits for a given value with a given degree of statistical confidence. A two-sided limit defines both the upper and lower bounds, while a one-sided limit defines a cut-off only on the upper or lower side with the understanding that there is no limit below or above the value of the variable. . Confidence limits can be statistically determined in a number of different ways if they meet certain statistical criteria that are compatible with individual statistical methods.

  The most statistical method relies on normally distributed variable values, ie it follows a bell-shaped Gaussian distribution. Normal distribution variables have sufficient statistical properties, and their statistical limits can be determined in a straightforward manner based on variable means and deviations.

  If the variable is not a normal distribution, a normalization conversion that converts the original variable into a new variable with a normal distribution and determines a confidence limit may be used. Appropriate mathematical transformations may be determined using statistical considerations or by empirical investigation of sufficiently large databases. Inverse transformations are also required to express confidence bounds in the form of original variables.

  Signal cross-correlation analysis may be utilized for the fitting procedure. The value ranges from -1 (absolute inverse correlation) to 0 (uncorrelated) to 1 (absolute positive correlation). In general, a significant positive correlation indicates strong true identification, and thus a one-sided, upper confidence limit should be used to define a dynamic discrimination threshold.

  By definition, correlation is a bounded variable and is therefore not normally distributed. A mathematical transformation is needed to normalize the correlation distribution that determines the upper confidence bound. On the other hand, an empirical technique that does not depend on such conversion may be used.

  The preferred method described more fully below is particularly suitable for correlation analysis. It is based on Fisher's Z transform and converts the correlation to a normal distribution variable.

  Another method may utilize square correlation. Since the raw correlation is not additive, other statistical functions for mean or correlation have no statistical significance. Although squared correlations are additive, they are also not normally distributed and therefore require additional transformations. If the correlation predecessor changes the distribution of those values, additional transformations may be required to account for these changes. These additional transformations include, but are not limited to, logarithm, square, square root, and transcendental functions.

  Yet another method involves some prior empirical testing, where a large number of objects are preferably correlated to a huge database. The possibility of false identification is determined directly by searching this database, otherwise the transformation can be determined empirically. However, since this method is not dynamic and must be done prior to the actual test, the effect of the test conditions cannot be easily compensated and requires the development of a mathematical model for the effects of noise. .

  The preferred method according to the second embodiment, the Fisher-transform method, converts the correlation between the target signature and the registered signature in order to obtain a score distribution closer to the normal distribution. Including. As described above, data that meets the normalization conditions can be used to derive confidence limits for the parameters.

  Fisher's Z-transform was designed to normalize the correlation. The transformation can be expressed as:

  Zf is a conversion value, arctanh is a hyperbolic arctangent function, and r is a conversion coefficient. arctanh is expressed in radians.

  Once all correlations have been converted, the one-sided confidence limits for the converted scores are the average of all the converted correlations and the standard deviation of all the converted correlations, excluding the target correlation. It is determined by calculating and calculating as follows.

ここで、ｚは正規分布’ｚ スコア’であり、Ｚｆ_ｍｅａｎはデータベースによる変換された相関関係の平均であり、ｓｄ_Ｚｆはデータベースによる変換された相関関係の標準偏差である。

Here, z is the normal distribution 'z score', Zf _mean is the average of the correlations converted by the database, and sd _Zf is the standard deviation of the correlations converted by the database.

  The lower case z herein shows the value of the normal distribution z-score, which is derived based on the desired degree of confidence in the cutoff. Such a table of scores is shown in FIG.

  In the table of FIG. 8, the standard deviation is accumulated by an appropriate z-score and added to the average, and the overall quantity is converted back to a correlation by taking a hyperbolic tangent.

  For example, a 95% confidence limit can be determined using a z score of 1.65. Therefore, if the average of the converted values is 0.05 and the standard deviation is 0.25, the 95% confidence limit is 0.72. That is, a correlation value exceeding 0.72 occurs only by chance on the occasion of 5% or less.

  The reverse procedure can be used to determine the likelihood that a particular object identification will be due to chance. The following equation is solved for z-score.

ここで、ｚは正規分布’ｚ スコア’であり、Ｚｆｃは変換された対象の相関関係であり、Ｚｆ_ｍｅａｎはデータベースによる変換された相関関係の平均であり、ｓｄ_Ｚｆはデータベースによる変換された相関関係の標準偏差である。

Where z is the normal distribution 'z score', Zfc is the correlation of the transformed object, Zf _mean is the average of the transformed correlation by the database, and sd _Zf is the transformed correlation by the database The standard deviation of the relationship.

  As a result, z-score can be converted to a one-sided probability value using an interpolation method if necessary with reference to a table of cumulative normal distribution. For example, referring to the brief table above, a z-score of 1.80 indicates a 3.75% probability that the object will be very highly correlated.

  As described above, if the noise in the registered signature, ie, the target signature, is random, the overall correlation with the target value will be reduced. Thus, true identification, if present, will have a lower correlation with the subject. It should be noted that high raw correlation does not fluctuate much due to the maximum correlation sealing effect, so reducing raw correlation values increases raw correlation variability, however, the conversion Compensated. Thus, the dynamic threshold with the desired certainty can be recalculated for each identification attempt using the method described above. It is important that the overall random noise still tends to drive the pre-correlation to zero and reduce the overall true variability, thereby reducing the confidence limit. However, a true fit is still significant unless the signal to noise ratio falls below a certain limit.

  The following example of the second form is based on a database of 38 patients. All patients are healthy individuals and participate in the study on a voluntary basis.

[Example 1: Normalization of correlation]
By correlating all pairs in the database, a set of 703 cross-correlations was obtained. The raw correlation distribution and the z-transformed correlation distribution are shown in FIG. The raw correlation is not a normal distribution (upper figure), but the transformed correlation appears to be close to the normal distribution (lower figure).

[Example 2: Implementation]
The biometric identification method was implemented utilizing an analysis of 38 registered signatures and 38 test signatures. FIG. 10 shows FAR and FRR implementation curves as a function of static threshold, and FIG. 11 shows implementation curves as a function of dynamic threshold. It is clear that the dynamic threshold shows significantly better results (eg EER _static = 3%, EER _dynamic = 0%)

[Third embodiment]
As mentioned above, the dynamic identification threshold is a data driven threshold and is preferably recalculated in each identification session to establish confidence limits and demonstrate the statistical significance of the identification process. However, the overall score will still decrease due to the degradation of the signal properties due to background noise, lowering the dynamic threshold, thereby reducing the identification confidence. This problem requires an evaluation of the nature of the signal in both the registration and identification phases to facilitate high performance recognition.

  The third form solves this problem by calculating the Q value, an index of the nature of one type of signal. The signal quality index Q is a quantitative description of the nature of the ECG signature. It is based on random errors in two or more ECG waveforms and is derived with reference to their signal average ECG.

  The Q value may be used to verify the nature of the signal between the registration phase and the identification phase, thereby ensuring sufficient system performance. In the case of a lower Q factor than required by a pre-defined threshold (based on the desired level of identification confidence per se), the measurement may be extended or repeated until the confidence requirement is met.

One suitable method for deriving Q is a series of steps as follows.
(1) The input electrocardiogram signal is classified into an electrocardiogram waveform having conventional waveform morphology characteristics (for example, P, QRSS, T element).
(2) The ECG waveform is arranged with respect to the R wave peak (“time fixed”).
(3) An average electrocardiogram is derived from the arranged electrocardiogram waveforms. The preferred method takes an arithmetic average, although other methods such as harmonic average, geometric average, weighted average, or intermediate value may be utilized. Yet another method involves transforming the original signal by other methods such as principal component analysis.
(4) Each original ECG waveform is processed against the average ECG, and some differences are derived with respect to the average ECG. Although other methods (eg, division of the original ECG by the average ECG) can be used, the preferred method performs subtraction. If the average electrocardiogram is stable and is a true signature of the patient's electrocardiogram, the resulting difference is a signature of the noise (electrocardiogram noise) inherent in the individual individual's ECG waveform.
(5) The individual sample points corresponding to the time axis in the individual ECG noise waveforms are processed together to derive the magnitude of variability. The most preferred method is to determine the variance. Another method that can be utilized includes standard deviation or distance.
(6) The average is obtained from these variability magnitudes. The most preferred method is to take an arithmetic average. Another method may include taking an average after conversion (eg, log) or taking another average (vaporization, harmony, intermediate). Another summary score, such as a maximum value, may also be available.

  Keeping in mind that the signal can be normalized prior to analysis, the average may itself be used as the Q index as directly related to the SNR. On the other hand, various other scaling transformations may be applied to the average to convert the average into an index with the desired minimum, maximum, and linearity characteristics.

[Example 1 according to third embodiment: Q (signal property) and NSR (signal-noise ratio)]
If X represents an electrocardiogram data matrix, each column representing one electrocardiogram waveform represents Xi (n), where i is the index of the electrocardiogram waveform and n represents a discrete time unit. The average of all electrocardiogram waveforms is denoted by x (n). Calculate error terms for every point at time n.

ここで、分散はδ_ｅ ^２（ｎ）で示される。可変平均を０から１の距離に変換する好適なスケーリングコンバージョンは、以下のように規定される。

Here, the dispersion is represented by δ _e ² (n). A suitable scaling conversion that converts a variable average from a distance of 0 to 1 is defined as:

  The simulation shown in FIG. 12 shows that the signal is evaluated with respect to the noise level using the Q factor. This simulation uses an actual electrocardiogram recording with an added level of Gaussian white noise added to the signal. FIG. 12 shows the Q value as a function of the signal-to-signal noise ratio (NSR). It can be seen that when Q begins to drop from the plateau, it decreases monotonically with increasing NSR, and eventually the ECG alignment procedure breaks down (NSR--35 dB, Q-0.2).

[Example 2 according to third embodiment: score as a function of signal properties]
Theoretically, a match score close to 1 tends to be a positive match and a non-match score tends to zero indicating a complete lack of correlation. In practice, however, the true fitness score is affected by temporal variations in the ECG signature and, more importantly, background noise. Therefore, higher signal properties are sought for high score discrimination in a short time. It should be noted that a high quality signal increases the upper boundary on the fitness score, but does not affect the lower boundary depending on the variability of the cardiological signature. The examples shown in FIGS. 13 and 14 show score distributions as a function of signal nature based on a database of 38 patients. FIG. 13 shows a short data segment of 5 seconds each. In contrast, FIG. 14 shows longer sections of 20 seconds each (FIG. 14). Obviously, the longer section compensates for the effect of noise to a considerable extent and the score distribution is flat.

[Example 3 according to third embodiment: signal properties and recording period]
The nature of the signal can be quantified using the Q parameter. If the Q value is smaller, but Q is not below a certain limit at which the ECG alignment process breaks down, a longer record is required to maintain statistical advantage. 15 and 16 show the increase in identification score as a function of recording period for a given Q value.

  Thus, according to this third aspect, the method and system disclosed herein can calculate the nature of the signal using a Q-factor or other means, and the system provides a reduced sample of noise. Or longer samples based on the quality factor or other signal quality criteria and the desired degree of identification confidence.

[Fourth form]
According to a fourth aspect, the methods and systems disclosed herein may encrypt stored signatures. This safety feature prevents misuse of data in the database even though the various methods and systems herein operate normally based on stored signatures rather than raw ECG data. Designed to be Thus, an additional layer of security may be utilized by encrypting the signature itself. For this purpose, various wavelength changing techniques including a PKI (public key infrastructure) technique used for credit card data may be used. This fourth step is necessary because an unauthenticated individual needs to decrypt the signature and convert the signature into a raw data signal, which is impossible without knowing which common characteristics have been removed from the raw data. According to the form, inappropriate use of registered patient data becomes more difficult. Thus, one advantage of the systems and methods disclosed herein is that it becomes very difficult to misuse stored information.

[Fifth embodiment]
Biometric identification systems are generally vulnerable to registration fraud. The system and method according to this fifth aspect solves this problem by utilizing electrocardiogram data from embryologically related individuals registered in a database. Close relatives have electrocardiograms that share common characteristics. By correlating patient signatures with the general population and / or correlating with registrants that are deliberately associated with the patient, the system determines with confidence whether the patient is their intended one. it can. This technique may be utilized in addition to verifying personal identification via conventional methods such as photo recognition and / or fingerprint matching. However, unlike those methods that are non-Euclidean and not amenable to similarity-based clustering, this technique determines which of the registration processes by determining the probability of the embryological relationship based on the registrant's ECG signature. Even at the stage, fraud can be determined.

[Sixth embodiment]
The systems and methods disclosed herein also utilize very high impedance probes for measuring electrocardiograms. Since reliability and availability are important for an electrocardiogram-based biometric identification system, it is useful to measure the electrocardiogram at a single point or even without touching the patient. In order to increase the reliability and ease of use of biometric identification, potential probes may operate in conjunction with biometric methods and systems, including those described herein. Very high impedance probes come in various forms. See, for example, Harland et al. The very high input impedance probe according to this configuration preferably has very low noise characteristics and does not require a current conduction path to operate. As a result, they work well with the methods and systems described above, even when used by an amateur without the assistance of an expert system operator. Thus, these probes can be used in airport-based biometric identification systems, such as collecting an electrocardiogram signal as an individual wears clothes and passes a scanner (similar to a metal detector). Similarly, a single probe may be used to collect an electrocardiogram from an individual fingertip in an ATM or a game machine. Utilizing a single probe gives the patient more freedom of movement and makes it easier for the patient to respond to identification and registration management. This is particularly useful when the biometric identification system described herein is used to control a patient's machine operation, especially when the machine requires physical contact to operate. This is the case (for example, small arms and vehicles). The single probe and remote probe ECG capture system according to this embodiment may be supplemented by a noise reduction scheme that reduces body noise and EMG.

  According to a seventh aspect, the biometric identification method and system provides a relevance score for a patient (formed by comparing a patient signature with that of a database registrant), a registrant signature in a database registration. It may be correlated with the relevance scores of multiple registrants (formed by comparison with those of the person). Thus, rather than analyzing the patient's correlated fitness score distribution, this identification technique analyzes the patient's fitness score correlation distribution and those of the registrant. With respect to the fifth aspect, the method and system according to this aspect are useful for identifying individuals involved. This is because individuals associated with a group of registrants have a Gaussian distribution of fitness scores with a substantially higher median than a Gaussian distribution of fitness scores for individuals not associated with the registrant. Therefore, by examining the distribution of the fitness score, the probability of the patient's embryological relationship with the registrant can be confirmed.

[Eighth form]
Finally, in addition to or in addition to the correlation techniques described above, the methods and systems described herein may utilize weighted correlation for identification. According to this embodiment, the correlation can give different weights to different signature differences. For example, the signature difference due to QRS group characteristics may be weighted more than the signature difference due to T or P complex characteristics. Since T is highly variable, QRS is stable, and P is intermediate, systems and methods may utilize the root mean square of signature values as part of the weighting function. Thus, the signature can be normalized using root mean square calculation, L1 metrics, or another normalization technique.

[Preferred embodiment that can be used with all forms]
FIG. 17 shows a functional diagram of a preferred system. Similarly, FIG. 18 shows a functional diagram of a preferred signal processor. The term “processor” is used herein in general usage, and processing may be done by physical discrete components, such as a coprocessor on an IC chip, or the processor may be physically An integrated unit may be included.

[General example that can be used with all forms: registration algorithm]
The following is an exemplary algorithm for the registration phase that can be utilized with any of the foregoing forms.
i. x _i (n) represents a 20 second, 250 Hz digital sample of the i th new patient, and n represents a discrete unit of time.
ii. x _i (n) is bandpass filtered in the range of 4 Hz-40 Hz.
iii. The filtered signal is denoted y _i (n).
iv. The filtered signal y _i (n) is searched for QRS complex and identifies the 'R' peak as an anchor point.
v. The filtered signal y _i (n) is maintained or inverted to obtain a positive 'R' peak.
vi. The identified QRS complex is calculated to establish the average pulse rate reading PR _i .
vii. The filtered signal y _i (n) is segmented around the anchor point and takes 50 samples before each 'R' anchor point and 90 samples after the anchor point.
viii. Individual data segments are normalized by the size of the 'R' anchor point.
ix. The segments are aligned and averaged around the anchor point to produce the patient's electrocardiogram signal denoted s _i (n).
x. The patient's electrocardiogram signal s _i (n) is adjusted according to the average pulse rate PR _i by normalizing the 'P' and 'T' latency associated with the pulse rate. The adjusted electrocardiogram signal is denoted by v _i (n).
xi. The patient's electrocardiogram signal v _i (n) adjusted for the pulse rate is added to the database and brought to the total average T (n).
xii. A set of electro-biometric signatures Φ _i is constructed by subtracting the total average T (n) from each of the adjusted ECG signals stored in the system database.

[Example: Recognition algorithm]
The following is an exemplary algorithm for the recognition phase.
i. x _j (n) represents a 10 second, 250 Hz digital sample of the patient being tested.
ii. x _j (n) is bandpass filtered in the range of 4 Hz-40 Hz.
iii. The filtered signal is denoted y _j (n).
iv. The filtered signal y _j (n) is searched for the position of the QRS complex and uses the R peak as an anchor point.
v. The filtered signal y _j (n) is maintained or inverted to obtain a positive 'R' peak.
vi. Identified QRS complex is calculated to establish the average pulse rate reading PR _j.
vii. The filtered signal y _j (n) is segmented around anchor points and takes 50 samples before each anchor point and 90 samples after the anchor point.
viii. The segments are aligned around the anchor point and averaged to produce a patient electrocardiogram signal denoted s _j (n).
ix. The patient's electrocardiogram signal s _j (n) is normalized according to the average pulse rate PR _j . The electrocardiogram signal of the patient whose pulse rate is adjusted is denoted by v _j (n).
x. An electro-biometric signature σ _j is constructed by subtracting the total average T (n) from the electrocardiogram signal v _j (n) with adjusted pulse rate.
xi. The correlation coefficient between the electron-biometric signature σ _j and all registered electron-biometric signatures Φ _i is calculated and squared to maintain its original sign.
xii. The maximum code-preserving square correlation value is selected and compared to a previously set threshold.
xiii. If the selected maximum sign-preserving squared correlation value is greater than a previously set threshold, a positive match is indicated and the patient is identified.

  Accordingly, ECG signal acquisition, processing, and analysis methods and apparatus for electro-biometric identification recognition may include any subset of the following registration and recognition steps.

[Register]
Acquisition, digitization and storage of ECG signals from patients;
a. ECG signal database organization;
b. Splitting the template database into multiple subsets based on ECG signal similarity;
c. Construction of one or more overall averages;
d. Derivation of patient-specific electronic-biometric signatures.

[recognition]
[Verification]
The newly captured e-biometric signature is compared to the patient-specific registered e-biometric signature template;
e. Correlation and confidence analysis of associated stored electronic-biometric signature templates and newly captured patient's electronic-biometric signatures;
f. Display and registration of recognition results and / or activation of physical or virtual local / remote mechanisms.

[identification]
The newly captured electron-biometric signature is compared with all of the electron-biometric signature templates that are added to the database;
g. Correlation and confidence analysis of all stored electronic-biometric signature templates and newly captured patient's electronic-biometric signatures;
h. Display and registration of recognition results and / or activation of physical or virtual local / remote mechanisms.

  Others can easily modify and adapt the embodiments herein for various applications without undue experimentation and without departing from the generic concept. Such adaptations and modifications are to be included and intended to be included within the meaning and scope of equivalents of the disclosed embodiments.

  It should be understood that the style or terminology used herein is for the purpose of description and is not limiting. The means, members, and steps for carrying out various disclosed functions may take a variety of alternative forms and still fall within the scope of the words or equivalents of the claims.

  Thus, the expressions “means for” and “means for” found in the above specification and / or claims and followed by a functional description, or the term of a method step, are present or future And whether it is an exact equivalent or embodiment of the embodiments disclosed in the above specification, other means or steps that perform the listed functions, i.e. perform the same functions, may be utilized. It is intended to define and cover any structural, physical, chemical or electrical element or configuration, or any method step, and their representation should give the broadest interpretation. Is intended.

DESCRIPTION OF SYMBOLS 12 ... Signal acquisition module, 14 ... Signal processing module, 16 ... Output module, 22 ... Sensor, 38 ... I / O port.

Claims (5)

In the method of identifying an individual,
Creating first and second biometric signatures that identify the particular individual by forming a difference between the representation of the heartbeat pattern of the particular individual and the stored representation of the common characteristics of the heartbeat patterns of the plurality of individuals a storing Te, each of said first and second biometric signature is in the different heart beating, and the step,
After the creating step, obtain a representation of the selected individual's heart beat pattern and between the stored representation of the selected individual's heart beat pattern and the common characteristics of the plurality of individual heart beat patterns. Forming a difference to create a third biometric signature;
Selecting one of the first and second biometric signatures based on the selected individual's pulse rate;
Said third biometric signature is compared to that said selected one of said first and second biometric signature method the selected individuals including and determining whether the above specific individual. The creating and storing step includes creating and storing a plurality of first and second biometric signatures, each comprising a representation of an individual individual heart beat pattern and a common characteristic of the heart beat pattern To identify individual individuals by forming a difference with the stored representation of
It said step of comparing is a plurality of said selected ones of said first and second biometric signature, for each method of claim 1, characterized in that it is carried out. The comparing step is
Said third biometric signature obtains a correlation coefficient correlated with those said selected one of said first and second biometric signature, the first and second biometric signature identified The method of claim 1, comprising comparing a correlation coefficient associated with the selected one of the selected ones with a correlation coefficient threshold.   The method of claim 1, wherein the step of obtaining a representation of a selected individual's heart beat pattern comprises obtaining a plurality of representations of a heart beat pattern. The step of creating and storing first and second biometric signatures of a particular individual obtains a plurality of representations of the heart beat pattern of the particular individual over a predetermined period, and a plurality of representations of the heart beat pattern of the particular individual The method of claim 1 including the step of creating first and second biometric signatures each successive from each of the first and second biometric signatures.

Images