US 20060052679 A1
The present invention generally relates to a method for continuous monitoring of the concentration of an analyte by determining its change over time in the living body of a human or animal. A measurement variable value correlating with the desired concentration of the analyte are measured as the measurement signal (zt) and the change over time of the concentration is determined from the measurement signal as the useful signal (yt) using a calibration. A filter algorithm is used to reduce errors of the useful signal, which result from noise contained in the measurement signal. The filter algorithm includes an operation in which the influence of an actual measurement value on the useful signal is weighted using a weighting factor (V).
1. A method for continuous monitoring concentration of an analyte by determining the analyte's change over time in the living body of a human or animal, the method comprising:
measuring at sequential points in time, measurement values of a measurement variable correlating with a desired concentration of the analyte;
measuring the measurement variable as a measurement signal (zt);
determining the change over time of the concentration of the analyte from the measurement signal as a useful signal (yt) by means of a calibration;
providing a filter algorithm in the time domain for determination of the useful signal (yt) from the measurement signal (zt), wherein the filter algorithm reduces errors of the useful signal resulting from noise contained in the measurement signal, wherein the filter algorithm includes an operation in which the influence of an actual measurement value on the useful signal is weighted by means of a weighting factor (V);
determining a signal variation parameter (σt) on the basis of signal variations detected in close chronological relation to the measurement of the actual measurement value; and
adapting dynamically the weighting factor as a function of the signal variation parameter determined for the point in time of the actual measurement.
2. The method according to
3. The method according to
4. The method according to
5. The method according to
6. The method according to
7. The method according to
8. The method according to
9. The method according to
10. A device for continuous monitoring of a concentration of an analyte by determining the analyte's change over time in the living body of a human or animal, the device comprising:
a measurement unit, by which measurement values of a measurement variable correlating with the desired concentration are measured as the measurement signal (zt) at sequential points in time;
an analysis unit, by which the change over time of the concentration is determined by means of a calibration as a useful signal (yt) from the measurement signal, and
a filter algorithm in the time domain for determination of the useful signal (yt) from the measurement signal (zt) to reduce errors of the useful signal, which result from noise contained in the measurement signal;
wherein the filter algorithm includes operations, in which the influence of an actual measurement value on the useful signal is weighted using a weighting factor (V), such that a signal variation parameter (σt) is determined on the basis of signal variations detected in close chronological relationship with the measurement of the actual measurement value, wherein the weighting factor is dynamically adapted as a function of the signal variation parameter determined for the point in time of the actual measurement
The present application is a Continuation of U.S. patent application Ser. No. 10/945,798, filed Sep. 21, 2004 which claims priority to German Patent Application No. 10343863.7, filed Sep. 23, 2003, which are hereby incorporated by reference in their entirety.
The present invention generally relates to a method and a device for continuous monitoring of the concentration of an analyte. In particular, the invention relates to determining the analyte's change over time in the living body of a human or animal. The term “continuous monitoring (CM)” is used hereafter for this purpose.
A CM method and device is described, for example, in U.S. Pat. No. 5,507,288.
A main task is the continuous monitoring of the concentration of glucose in the body of the patient, which is of great medicinal significance. Studies have led to the result that extremely grave long-term effects of diabetes mellitus (for example, blinding because of retinopathy) can be avoided if the change over time of the concentration of the glucose is continuously monitored in vivo. Continuous monitoring allows to dose the required medication (insulin) precisely at each point in time and to keep the blood sugar level always within narrow limits, similarly to a healthy person.
The present invention relates in particular to CM of glucose. Further information can be taken from document (1) and the literature cited therein. The content of this document is incorporated herein by reference.
The present invention is, however, also suitable for other applications in which the change over time of an analyte in the living body (useful signal) is derived from a measurement signal, which comprises measurement values, measured at sequential points in time, of a measurement variable correlating with the concentration desired. The measurement signal may be measured invasively or non-invasively.
An invasive measurement method is described, for example, in
U.S. Pat. No. 6,584,335.
Here a hollow needle carrying a thin optical fiber is stuck into the skin, light is irradiated under the skin surface through the optical fiber, and a modification of the light through interaction with interstitial liquid which surrounds the optical fiber is measured. In this case, the measurement signal comprises measurement values obtained from light which is returned through the optical fiber into a measurement device after the interaction. For example, the measurement signal may comprise spectra of the light which are measured at sequential points in time.
Another example of invasive measurement methods is the monitoring of concentrations by means of an electrochemical sensor which may be stuck into the skin. An electrical measurement variable, typically a current, is thus determined as the measurement variable which is correlated with the concentration of the analyte.
Different non-invasive methods are discussed in Document (1). These include spectroscopic methods in which light is irradiated directly (i.e., without injuring the skin) through the skin surface into the body and diffusely reflected light is analyzed. Methods of this type have achieved some importance for checking the change over time of oxygen saturation in the blood. For the analysis of glucose alternative methods are preferred, in which light is irradiated into the skin in a strongly localized manner (typically punctually) and the useful signal (course of the glucose concentration) is obtained from the spatial distribution of the secondary light coming out of the skin in the surroundings of the irradiation point. In this case the measurement signal is formed by the intensity profile, measured at sequential points in time, of the secondary light in the surroundings of the irradiation point.
A common feature of all methods of this type is that the change of the concentration over time (useful signal) is determined from the measurement values measured at sequential points in time (measurement signal) using a microprocessor system and a suitable algorithm. This analysis algorithm includes the following partial algorithms: a filter algorithm, by which errors of the useful signal resulting from signal noise contained in the measurement signal are reduced and a conversion algorithm, in which a functional relationship determined by calibration, which relationship describes the correlation between measurement signal and useful signal, is used.
Typically, these parts of the analysis algorithm are performed in the described sequence, i.e., first a filtered measurement signal is obtained from a raw measurement signal by filtering and the filtered signal is then converted into the useful signal. However, this sequence is not mandatory. The raw measurement signal can also be first converted into a raw useful signal and then filtered to obtain the final useful signal. The analysis algorithm may also include further steps in which intermediate variables are determined. It is only necessary in the scope of the present invention that the two partial algorithms a) and b) are performed as part of the analysis algorithm. The partial algorithms a) and b) may be inserted anywhere into the analysis algorithm and performed at any time.
The present invention relates to cases in which time domain filter algorithms are used. Kalman filter algorithms are particularly common for this purpose. More detailed information on filter algorithms of this type is disclosed by the following literature citations, some of which also describe chemical and medical applications:
S. D. Brown: The Kalman filter in analytical chemistry, Analytica Chimica Acta 181 (1986), 1-26.
K. Gordon: The multi-state Kalman filter in medical monitoring, Computer Methods and Programs in Biomedicine 23 (1986), 147-154.
K. Gordon, A. F. M. Smith: Modeling and monitoring biomedical time series, Journal of the American Statistical Association 85 (1990), 328-337.
U.S. Pat. No. 5,921,937
EP 0 910 023 A2
WO 01/38948 A2
U.S. Pat. No. 6,317,662
U.S. Pat. No. 6,575,905 B2
As noted, the filter algorithm is used for the purpose of removing noise signals which are contained in the raw measurement signal and would corrupt the useful signal. The goal of every filter algorithm is to eliminate this noise as completely as possible, but simultaneously avoid to disturb the measurement signal. This goal is especially difficult to achieve for in vivo monitoring of analytes, because the measurement signals are typically very weak and have strong noise components. Special problems arise because the measurement signal typically contains two types of noise, which differ significantly in regard to the requirements for the filter algorithm: measurement noise: such noise signal components follow a normal distribution having a constant standard deviation around the correct (physiological) measurement signal non-physiological signal changes, which are caused, for example, by movements of the patient and changes of the coupling of a measurement sensor to the skin to which it is connected. They are typically neither distributed normally around the physiological measurement signal, nor is the standard deviation from the physiological measurement signal constant. For such noise components of the raw signal the term NNNC (non-normal, non-constant)-noise is used hereafter.
The present invention is based on the technical problem to achieve a better precision of CM methods by improving the filtering of noise signals.
According to the present invention this is achieved by means of a filter algorithm which includes an operation in which the influence of an actual measurement value on the useful signal is weighted using a weighting factor (“controllable filter algorithm”), a signal variation parameter (related in each case to the actual point in time, i.e. time-dependent) is determined on the basis of signal variations detected during the continuous monitoring in close chronological connection with the measurement and the weighting factor is adapted dynamically as a function of the signal variation parameter determined for the point in time of the actual measurement.
The present invention, including preferred embodiments, will be described in greater detail hereafter on the basis of the figures. The details shown therein and described in the following may be used individually or in combination to provide preferred embodiments of the present invention.
The following detailed description of the embodiments of the present invention can be best understood when read in conjunction with the following drawings, where like structure is indicated with like reference numerals and in which:
Skilled artisans appreciate that elements in the figures are illustrated for simplicity and clarity and have not necessarily been drawn to scale. For example, the dimensions of some of the elements in the figures may be exaggerated relative to other elements to help improve understanding of the embodiment(s) of the present invention.
The following description of the preferred embodiment is merely exemplary in nature and is in no way intended to limit the invention or its application or uses.
The components of a CM device according to the present invention are shown in
Variations of the flow of the interstitial liquid from the body into the photometric measurement unit 12 lead to regular, relatively small signal variations, which are referred to as “fluidic modulation”. After approximately three minutes, at the point in time identified with the arrow 18, an inhibition of the liquid flow occurred, which may be caused, for example, by movement of the patient or by the entrance of a cell particle into the catheter 10. This inhibition of the flow leads to a large drop of the raw measurement signal A. This is an example of the fact that not all noise signals are distributed normally, with essentially constant standard deviation, around the signal corresponding to the actual physiological measurement value. Rather also interfering contributions of the type shown here exist, for which these conditions do not apply (NNNC noise). Therefore, the signal requires filtering even in such cases in such a manner that a useful signal results which corresponds as closely as possible to the actual physiological concentration of the analyte. An example for such a useful signal is shown in
The basis of a filter algorithm operating in the time domain, which the present invention relates to, is a system model that describes the change over time of the variables of interest and their relationship to one another. The functional relationship which describes the development of the system from time t to time t+1 is as follows:
Therein, yt and ut are vectors, which are referred to as state vectors and vectors of input variables, respectively. The state vector yt contains the variables of physiological interest and optionally check variables, which allow to check the measurement, as will be described in greater detail below. In the CM method, these include the desired analyte concentration, for example, the glucose concentration gt in the blood. The speed of change of the analyte concentration gt′=dgt/dt is suitable as a check variable. The state variable yt may also contain model variables related to the measurement method. For example, in the case of a measurement result of the type shown in
Input variables which, in the field of automatic control, correspond to control variables and are therefore not measured themselves are entered into the vector ut. In the case of glucose monitoring, for example, the administered insulin quantity given and the bread exchange units supplied are suitable input variables, because they both influence the glucose concentration in the blood. If these input variables are used, the vector ut has two elements: insulin dose and bread exchange units. A characteristic feature of input variables is that no prediction of their future values is necessary in the scope of the filter algorithm.
The mentioned variables of the state vector yt and the input vector ut are, of course, only to be understood as examples. The present invention relates to greatly varying systems which require different system models. It is not necessary to use the models in a discrete form. The continuous form with the corresponding differential equations may also be used.
A feature of filter algorithms in the time domain, is that they include an alternating sequence of predictions and corrections. A prediction of the system state (“predictor step”) is followed by a subsequent correction of this prediction on the basis of a further measurement value (“corrector step”).
In a predictor step, the actual value of the state variable yt at the point in time t is predicted using the system equation (1):
In this equation, ŷt identifies the value of the state vector at the point in time t which is estimated (predicted) using the data of the previous point in time (t−1); Wt identifies a system error vector.
In the case of a recursive filter algorithm, the calculation of each predictor step is not performed by taking all preceding points in time (t−1, t−2, t−3 . . . ) into consideration, but rather by using a weighted sum of smoothed signal values. In the example of a linear Kalman algorithm, the corresponding equation may be written as follows:
In this equation, At is the system matrix and B is the input matrix. In the general (non-linear) case, ft is to be preset or is to be calculated from data determined up to this point.
In the corrector step, the prediction is corrected on the basis of an actual measurement value according to
In this equation, Δt is a variable which represents a measure of the deviation of an actual measurement value zt from the predicted value and is referred to as the “innovation”.
Further it is taken into consideration that typically the system variables cannot be observed directly. The linkage between the measurement values and the state variables is provided by means of a measurement model (measurement function ht) according to:
The noise of the measurement values is taken into consideration by vt.
In the case of a linear Kalman algorithm (cf. equation 2a), the measurement equation is
Ht referring to the measurement matrix.
For example, in the continuous monitoring of glucose using an electrochemical sensor, a current i is measured which is correlated with the glucose concentration gt. In that example ht describes the correlation of the state variable gt with the measurement variable i (current), which is an element of the vector zt.
In the given example of photometric glucose detection using filter-assisted compensation of the fluidic modulation, a non-linear measurement model is used which links the photometric measurement signal zt to the system variables of glucose concentration gt, amplitude At, and frequency ωt of the fluidic modulation:
According to equation (3), the influence of the actual measurement value (contained in the innovation Δt) on the filtered useful signal value yt is weighted by the factors αt and βt. The described algorithm is therefore a controllable filter algorithm.
In the case of a Kalman filter, αt=1 for every point in time and βt=Kt. Kt refers to the Kalman gain. Accordingly, the corrector equation is as follows:
Further details regarding the Kalman gain Kt and more detailed information on the algorithm may be taken from the relevant literature, as cited above. Expressed descriptively, the Kalman gain is a measure of the weight given to additional measurement values. The Kalman gain is calculated anew in every iteration step of the filter algorithm according to an equation which may be written in simplified form (for the linear case) as follows:
Here, Pt designates the Kalman error covariance matrix. V designates the measurement error covariance matrix in the conventional Kalman algorithm.
Equation (6) shows that the elements of Kt may assume only values between 0 and 1. If the assumed measurement error V is relatively large in relation to the Kalman error covariance Pt, Kt is small, i.e., the particular actual measurement value is given relatively little weight. In contrast, if V is small in relation to Pt (multiplied by Ht), a strong correction occurs due to the actual measurement value.
An important improvement of the filtering is achieved in the scope of the present invention in that, on the basis of signal variations detected in close chronological relationship with the measurement of the actual measurement value zt, a signal variation parameter, designated here as σt, is determined and the weighting of the influence of the actual measurement value zt is dynamically adapted in the context of the corrector step as a function of σt. This is shown in graphic form in
The present invention does not have the goal of weighting different filter types—like a filter bank—by applying weighting factors. For this purpose, a series of system models analogous to equation (2) would have to be defined, one model for each filter of the filter bank. This is not necessary in the present invention, whereby the method is less complex.
No precise mathematical rules may be specified for the functional relationships used in steps 23 and 24, because they must be tailored to each individual case. However, the following general rules apply:
The standard deviation, which may be calculated as follows, is suitable as the signal variation parameter, for example.
If one assumes that the determination of the standard deviation is based on the actual measurement values z and four preceding measurement values z1 to z4, and if the difference between z and the preceding values is referred to as δz (δzn=z−zn), the average value ε is calculated as
The standard deviation of the four values of the difference δ1, δ2, δ3, δ4 in relation to the linear smoothing function is
On the basis of this standard deviation at, a dynamic (time-dependent) measurement error covariance Vt, which is included in a filter core with the Kalman algorithm, may be calculated, for example, according to
In this case, σo and γ are constant parameters which characterize the filter, and which may be set to tailor the chronological behavior of the filter, in particular its adaptivity, to a particular application.
In the example of a controllable recursive filter, the weighting factors αt, βt from equation (3) are a function of the signal variation parameter in such a manner that with increasing σt, factor αt becomes larger and factor βt becomes smaller.
As already explained, equations (7) through (10) only represent one of numerous possibilities for calculating a signal variation parameter and, based thereon, a weighting factor for a controllable filter algorithm in the time domain. The standard deviation, which may, of course, be calculated using a varying number of measurement values, can be replaced by variables which represent a measure for the signal variations in a period of time preceding an actual measurement value. The term “signal variation parameter” is used generally to identify a mathematical variable which fulfills these requirements.
Three typical graphs of a signal S are plotted against time t in
Evidently, in the case of curve b the strong variations are filtered well within the circle 25, but in the rectangle 26, the signal b reflects the physiological variations of the raw signal insufficiently. The useful signal c, in contrast, follows the physiological variations in the region 26 well, while the filtering of the non-physiological variations in the region 25 is insufficient. The conventional Kalman filter algorithm therefore allows no setting which leads to optimal filtering for the different conditions shown. In contrast, the present invention does not even require knowledge of the maximum variations of measurement values. The filter algorithm adapts itself automatically to the changes in the signal course and provides a filtered signal which corresponds to the curve b in the circle 25 and to the curve c in the rectangle 26.
Significant additional reliability may be achieved if the filtering extends not only to the desired analyte concentration, but rather additionally to at least one further variable, which is designated “check variable”. This may be a variable derived from the analyte concentration, in particular its first, second, or higher derivative versus time. Alternatively, an additional measurement variable, such as the flow of the interstitial liquid at the sensor shown in
This check variable may, as explained above (for gt′, At, and ωt), be included in the filter algorithm as a system variable. The filtering then also extends to the check variable, for which corresponding reliable smoothed useful signal values are available as the result of the filtering. These may then be compared to threshold values, in order to perform plausibility checks, for example. In the case of the glucose concentration, for example, it is known that the glucose concentration physiologically does not change by more than 3 mg/dl/min under normal conditions. A higher filtered value of the time derivative gt′ is a sign of a malfunction. Therefore the query 30 shown in
In order that the invention may be more readily understood, reference is made to the following examples, which are intended to illustrate the invention, but not limit the scope thereof.
It is noted that terms like “preferably”, “commonly”, and “typically” are not utiliized herein to limit the cope of the claimed invention or to imply that certain features are critical, essential, or even important to the structure or function of the claimed invention. Rather, these terms are merely intended to highlight alternative or additional features that may or may not be utilized in a particular embodiment of the present invention.
For the purposes of describing and defining the present invention it is noted that the term “substantially” is utilized herein to represent the inherent degree of incertainty that may be attributed to any quantitative comparison, value, measurement, or other representation. The term “substantially” is also utilized herein to represent the degree by which a quantitative representation may vary from a stated reference without resulting in a change in the basic function of the subject matter at issue.
Having described the invention in detail and by reference to specific embodiments thereof, it will be apparent that modifications and variations are possible without departing from the scope of the invention defined in the appended claims. More specifically, although some aspects of the present invention are identified herein as preferred or particularly advantageous, it is contemplated that the present invention is not necessarily limited to these preferred aspects of the invention.
As any person skilled in the art will recognize from the previous description and from the figures and claims, modifications and changes can be made to the preferred embodiment of the invention without departing from the scope of the invention as defined in the following claims.