DE10123572C1 - Automated online analysis of series of measurements involves evaluating sampling values relevant to analysis depending on association with model curve determined by mathematical model - Google Patents

Abstract

The method involves automatically associating the sampled values acquired during the series of measurements with a model curve determined by a mathematical model, determining and evaluating the sampled values relevant to the analysis depending on this association and determining the end of the measurements. Independent claims are also included for: (1) an arrangement for automated online analysis of series of measurements; (2) a computer program product for automated online analysis of series of measurements; and (3) a computer-readable storage medium for a computer program.

Description

The invention relates to a method and a Vorrich automatic online analysis of series of measurements of time-dependent signals and an ent talking computer program produce and a correspond suitable computer-readable storage medium. operation area of procedures and computer program product is in particular the automatic online evaluation of various NEN time-dependent signal curves, in which for the meas determination of value determined at the beginning of the measurement known positions or areas of the curve Need to become. An application example for the Vorrich tung is a portable, easy-to-use biosensor Measuring device for determining the concentration for a large number of analytes.

The problem of analysis and interpretation considered here tation of measurement curves occurs in the areas of Senso on which the measured value to be determined is not only  from the current measurement signal, but also from facto ren, such as curve type or measurement time within the course ve depends.

For the determination of measured values from measurement curves is a Numerous different methods are known, the type of the signal influences the choice of method (see: MR. Tränkler pocket book of measurement technology, Oldenbourg- Verlag 1992, H.-R. Tränkler signal processing. In: Sen sors Vol 1. Edited by W. Göpel, J. Hesse, J. N. Zemel. VCH Verlag Weinheim).

Subsequently, only non-periodic signals continue considered. From measurement curves of such non-periodic ones Signals can be used to obtain a measured value, for example that a sample after a certain on positioning time, with a saturation value or a measure ximum or minimum of the curve is taken. Furthermore is known to find the measured value by integrating the Samples over a certain time (area under egg ner curve) or to be calculated from the increase NEN. For example, JP 3262938 (Method for pro cessing output signal of semiconductor biosensor) maximum slope of a trace is used to determine the Measured value. It is also known at known curve, the curve using a approximate suitable function and so the measurement time shorten (K. Hirose and K. Daiichi. Method of meas uring sample by enzyme electrodes. European patent 0 623 681 A1, 1994).

For dynamic systems whose behavior is caused by time series hen of measurable quantities is a method to determine the course of the curve of these measured variables, which are characterized by the system parameters (modes) be based in the German patent application  DE 197 40 565 A1 described. In this procedure to Er Version of the modes of a dynamic system with a Variety of fashions, each one a (t) charak own systematic parameters will be a time row of at least one system variable x (t) of a mo dellation such as B. Unterzo a switching segmentation gene, which is set up in each time period a predetermined minimum length for each system variable le x (t) a predetermined prediction model such as e.g. B. a neural network for a corresponding system mode to capture, after modeling the time a drift segmentation takes place in which in each Period in which the system from a first Sy stem mode changes to a second system mode, a fol mixed forecast models that through a linear, paired overlay of the before sagemodelle of the two system modes is given.

Furthermore, it is known to use Ver continue to analyze and evaluate the pattern recognition For example, fuzzy logic becomes a pattern identifier used (H.-J. Zimmerman. Fuzzy Set Theory and its applications. Kluwer Academic Publishers, Bos ton / Dordrecht / London; T. Tilli. Pattern recognition with Fuzzy logic. Franzis publishing house, Munich. 1993), specifically also for evaluating biosensor signals. Doing it Known methods do not allow the system to be trailed kidneys to process complex curves or the end of the measurement automatically recognized (J. Weitzenberg. Methods of Fuzzy logic for evaluating measurement curves amperometri shear biosensors. Master's thesis, Martin Luther University of Halle-Wittenberg, 1998).

It is also known to use neural networks for pattern recognition to use, especially for the analysis of peaks in the measurement course  ven (US 5,121,433 Neural Net system for analyzing chro matographic peaks).

Hidden Markov models (HMM) are also used Pattern recognition used.

In the past decade, these have become the standard me method for speech recognition (L. R. Rabi ner. A tutorial on hidden markov models and selected applications in speech recognition. Proceedings of the IEEE, 77 (2): 257-286, 1989; L.R. Rabiner and B.H. Juang. Fundamentals of Speech Recognition. Prentice Hall, 1993; US 5,307,444 Voice Analyzing System Ising Hidden Markov Model and Having Plural Neural Network Predictor).

HMM also come for other problems, such as B. for the analysis of periodic ECG curves (US Pat. No. 5,778,881), DNA sequence recognition (P.F. Baldi and S. Brunak. Bio informatics: The machine learning approach. The MIT Press, 1998) for handwriting recognition (US Pat. No. 5,636,291 Continuous parameter hidden Markov model approach to automatic handwriting recognition) or for error diagnosis nose (US 5,465,321 HMM for fault detection in dynamic sytems).

It is state of the art for HMM specific problem positions to use certain algorithms. The For ward algorithm (L.R. Rabiner and B.H. Juang. Funda mentals of speech recognition. Prentice Hall, 1993) calculates the probability that an HMM will given sequence of symbols. To make it true most probable sequence of states for a given HMM and counter To determine the sequence of symbols, the Vi terbi algorithm (G.D. Forney Jr. The Viterbi Algo rithm. Proceedings of the IEEE, vol 61, no. 3, 1973) used. To train HMM using sample The Baum-Welch algorithm (L. R. Rabi  ner and B. H. Juang. Fundamentals of Speech Recogni tion. Prentice Hall, 1993). Also stand The technology is the visualization or clustering of Data with a so-called self-organizing map (T. Kohonen. Self-organizing maps. Springer publishing house, Heidelberg, 1995).

A method for estimating low memory FM Need and for online estimation of FM is in the German published patent application DE 199 34 845 A1 ben. In this process, at least one with min at least three indices, depending on the auxiliary size iterated from the last input / output value from which (or those) current estimates of the model parameters can be calculated so that when estimating online no part of the time series has to be saved, and the Memory requirements do not depend on the length of the time series hangs.

The known methods are not for online recognition suitable measuring positions (samples), if there are any among the curves to be analyzed ten where several local maxima or minima available and the measuring position depending on the curve shape or between such local maxima or minima. Furthermore, the automatic detection of the end is one Online measurement not possible or not always possible.

The object to be achieved by the invention is in, to address the disadvantages described, in particular a method, an apparatus and a compu To provide the program product through which it is possible in the acquisition of measured values from a time series of samples the initially unknown time point of the desired measurement position and / or the end of the measurement to determine online.  

This object is achieved by the Features in claims 1, 11, 19 and 21.

Advantageous embodiments of the invention are in the Subclaims included.

A particular advantage of the invention is that in the procedure for automatic online analysis of Series of measurements the samples recorded in the series of measurements during the measurement one automatically by a math matic model assigned to the model curve and depending on this assignment the for the Analy The relevant sample values are determined and evaluated which and the end of the measurement is determined.

The inventive device for automatic on line analysis of series of measurements is distinguished by that they have at least one sensory agent capture of data, at least one computer system with an execution environment in which an application runs, which are the samples taken in a series of measurements during the measurement one automatically by a math assigns the model curve to the determined model and depending on this assignment the for the Analy relevant samples are determined and evaluated as well the end of the measurement determines at least one means of storage Securing the recorded data and the mathematical Mo dells, and means of exchange and / or representation of data, the components mentioned above Means for data transmission are interconnected.

Another advantageous part of the invention is a computer program product for automatic online ne analysis of series of measurements, directly in the respective  internal memory of a digital computer and / or Microcontrollers can be loaded and software code includes sections with which the steps according to An 1 are executed if the product is on a Computer and / or microcontroller is running.

For the automatic online analysis of series of measurements advantageously a computer-readable storage medium used on which a program is stored that it enables a computer after it is in the memory cher of the computer has been loaded, an automati carry out cal online analysis of series of measurements, whereby the samples recorded in the measurement series during the measurement automatically one by a mathematical Model determined model curve assigned and depending of this assignment the rele for the analysis The relevant samples are determined and evaluated in this way how the end of measurement is determined.

In particular, it is to be regarded as an advantage of the method according to the invention that it makes it possible to determine the initially unknown point in time of the measurement position sought when extracting measured values from a time series of samples. Since this point in time can be significantly dependent on the current curve shape and, moreover, it is often only possible to determine when the measurement can be ended on the basis of the curve shape during the measurement (cf. FIG. 1 (a) -1 (f)), it represents an entirely The decisive advantage is that the type of curve to be used for the measurement value formation and the detection of the end of the measurement is already detected during the measurement process.

The method is characterized in that be the most probable model during the measurement is determined for the current curve shape. It he is the first to allow the model to be determined the end of the measurement and for the detection of the for the Measurement of the relevant area or the relevant Use measuring position. There is also opposite known methods, e.g. B. by means of fuzzy logic the models based on statistical trai training data and special expert knowledge.

The invention is based on one in the Figures illustrated embodiment explained in more detail be tert.

Show it:

Fig. 1 different waveforms with mapped measurement positions for a sensor type;

Fig. 2 assignment of sections of a measurement curve to the states of an HMM. The number of states E _i varies from curve type to curve type;

Fig. 3 basic sequence of the method (working phase);

Fig. 4 block diagram of a biosensor measuring device.

In a preferred embodiment, the object is achieved in that the time course of the sample values of the measurement curve with a trainable, statistical signal model described with a representative number of in front of measured curves that have been determined and evaluated by experts course is trained (training phase) and that the trained model for the evaluation of emerging  Sample time curves is used online to time points to determine which of the current Ab the measured value is to be calculated.

The measurement curves belonging to the series of measurements are marked with Modeled using a statistical signal model. This offers the advantage of not knowing the curves a detailed physico-chemical signal mo dells to be able to process. As a statistical signal Hidden Markov (HMM) models are chosen. because at the individual sections of a measurement curve stands assigned in the HMM. A corresponding model tion enables the detection of the for the measurement of the correct model using the curve ver run even with different curve types. Through the Assignment of relevant curve sections to the Zu levels of the HMM is a recognition of the end of the measurement guaranteed. You can also use the states the position of the necessary to determine the measured value Determine the sample values within the curve. For the Generating the HMM are both statistical trainings data as well as special expert knowledge.

In a particular embodiment of the invention, the statistical signal model is a discrete HMM. For this purpose, the sample-time curves are divided into several sections, each section being assigned one or more states of the discrete HMM. The for the measurement value formation 10 relevant sections are assigned to special conditions. A state for end detection is also provided.

The first to nth derivatives of the sample-time curve are used as features for generating observation symbols 13 for the HMM. The derivatives are combined into an n-dimensional feature vector. A vector quantization (VQ) 3 is necessary for processing in a discrete HMM. For this purpose, a self-organizing map (SOM) according to Kohonen is used.

An HMM is used for each class of sample-time curves trained with the help of the Baum-Welch algorithm. Since the He can only maximize the model parameters locally the selection of suitable initial parameters, especially for the emission probabilities of great importance. Therefore, in all training curves they become the measured value relevant sections with the help of a developed software based on expert knowledge ma recently marked, d. H. for further processing placed. In all areas, the middle icon frequencies for all training curves estimated and as initial emission probabilities for the assigned conditions used in training.

When analyzing a newly emerging, still unknown and incomplete scan sequence, the most likely model for the current measurement curve is first determined with the help of the forward algorithm 4 . The most probable state sequence 12 for the selected model and the current curve is then calculated using the Viterbi algorithm 5 . Based on the state sequence 12 , the end of the measurement and the measured value within the curve are detected and the times of occurrence of the signals to be calculated are determined.

The invention is illustrated below using the example of measurement cure ven amperometric biosensors explained.

The concentration of the analyte to be determined is in amperometric biosensors to a time-varying current I (t, C), which is made up of a basic current component I ₀ and the concentration-dependent signal current component I _D. Therefore, I ₀ and I _D must be determined separately for each measurement. For this purpose, the sensor is first immersed in a buffer solution (determination of I ₀ ) and then in the actual measurement solution (determination of I _D ). This shows a typical run-in curve in the buffer solution, at the flattest point of which I _{0 is} averaged. In the measuring solution there are characteristic signal curves, the shape of which depends on the sensor type, measuring method and the concentration of the analyte to be measured ( Fig. 1 (a) - (f)). The measuring time is in the range of seconds to minutes. The course of the curve is used to decide when the measurement is ended.

modeling

Such a biosensor measurement curve can be divided into several sections, which can be represented as a sequence of states (see FIG. 2). Different models but the same procedure are used to determine I ₀ and I _D. This method is explained below using the example of the determination of I _D.

The first section of each curve, the typical setting phase I, is with a number of n states E ₁ , depending on the curve type. , ., E _n modeled. Be the area II after the adjustment phase, a state V is assigned, which models the beginning of the measuring section. The actual measuring section III follows, which is relevant only for the measurement 10 . The measured value is determined in section III by averaging or maximum formation. This zone (usually a certain peak, a saddle or a plateau) can differ in its position and shape depending on the type of curve. The measuring section III is modeled with the state M.

The measuring section III is followed by an area IV in which the Current usually remains constant or even increases or decreases rapidly. The state N is drawn to it assigns. This area IV becomes permeable enough fen, this is a sure sign to end the Measurement. Reaching the end of the measurement is done by modeled an excellent final state F.

Feature extraction 2

To determine the 1st and 2nd derivative of the curve the measurements in a fixed size window, which z. B. contains 30 points, with a square radio approximation, from which the derivatives analy be determined table. For vector quantization comes a self-organizing map with z. B. 16 vectors for use.

training phase

For the training phase must be complete, by experts evaluated, measurement curves of each occurring curve type for To be available.

In each of these curves, the times are first for the start of the adjustment phase (and thus the measurement solution), the end of the adjustment phase, the time of measurement and fixed the end of the measurement. A fixed time range from Z. B. 4 s around the marked time of measurement is considered Measuring section M fixed. This gives you automa table the times for the sections before V and after N the measuring section. In all areas, the mitt symbol frequencies for all training curves estimates and as initial emission probabilities used for the assigned states in training.

The HMM is initially only trained up to state N. and then extended by the final state F, wel  the same emission probabilities as the Condition N is maintained. Based on all training curves, the average length of stay in state N using the Mar Estimations are estimated and from this the transition probability calculated from state N to state F.

working phase

In the online analysis of a scanning sequence I (t), the forward algorithm must first be used to determine which of the trained HMMs represents the current measurement curve. Once this model has been determined, the end of the measurement must be detected and, if necessary, the measurement section determined. To do this, proceed as follows (see also Fig. 3):

• 1. set t = 0
• 2. Repeat steps (a) to (f) until the last state of the state sequence 12 q ₁ ,. , ., q _{t is} equal to the final state F, or the maximum time t _{max has been} exceeded:
  • a) t = t + 1
  • b) measurement of a new sample value I (t).
  • c) With feature extraction 2 and VQ 3 , a new observation symbol 13 o _{t is} generated.
  • d) Determine with the forward algorithm 4 for the observation symbols 13 o ₁ ,. , ., o _t the probability P (o ₁ ,..., o _t | HMM _i ) for each model i.
  • e) Choose the model with the maximum probability. This model is assumed to be the correct interpretation of the curve shape.
  • f) Use the Viterbi algorithm 5 to calculate the most likely state sequence 12 q ₁ ,. , ., q _t for the selected model. Each state q _{j of} the state sequence 12 is assigned the corresponding current value 11 I _j for later evaluation.
• 3. If the final state F has been reached 6, the actual measured value formation 10 takes place: on the basis of the state sequence 12 q ₁ ,. , ., q _t , all current values 11 I _k ,. , ., I _{k + n of} the curve, to which the measurement state M has been assigned, is used for the measurement determination. This can be an averaging over these values or a maximum formation. If the specified maximum time has been exceeded 7 and the final state F has not been reached, it must be checked whether at least the measuring state M has been reached 8. This case has to be treated separately: A measured value can be formed under certain circumstances, but a warning is issued. If the measurement state M was also not reached, the measurement is discarded. 9. The reasons for the latter case can be: a new or unknown curve type or a defective or used sensor.

The device according to the invention is described in more detail below using the exemplary embodiment of a portable, easy-to-use biosensor measuring device for determining the concentration for a large number of analytes with reference to FIG. 4.

The device comprises the pluggable biosensible biosensor 14 , two circuit complexes 16 and 19 which are substantially different in their function, and also connected, if necessary, to a personal computer 22 .

The circuit complex 16 is used for analog preprocessing and conversion of the sensor signal, which is passed via a data line 28 to an I / U converter 27 and then to an A / D converter 26 for signal conversion. To set the biosensor operating point, a polarization voltage is generated by means of a D / A converter 17 and fed to the biosensor 14 via the data line 15 .

The circuit complex 19 is used for the digital further processing of the biosensor signal, it comprises the micro controller 24 , a memory 23 , a display unit 25 and a keypad 20th An interface 21 realizes the coupling between microcontroller 24 and personal computer 22 if necessary.

The microcontroller 24 controls the A / D wander 26 , the D / A converter 27 via the internal bus 18 and communicates with the memory 23 in order to obtain the sample values and to save them for further processing by means of HMM.

The other essential task of the microcontroller 24 is the simultaneous (parallel) analysis of the incoming sample values by means of HMM and, after the time for the measured value taking and the end of the measurement have been detected / recognized by means of HMM, the calculation of the measured value. The microcontroller program contains all data of the trained HMM as well as the above-mentioned algorithms for processing the data with a trained HMM.

The HMM is trained on the basis of representative measurement curves evaluated by experts by means of the personal computer 22 ; After successful training, the data of the trained HMM are transmitted via the interface 21 into the microcontroller 24 and stored there.

The computer program product, which enables the biosensor measuring device to carry out the method according to the invention, consists of two packages, one of which all program parts running on the microcontroller 24 (generation, acquisition, evaluation and display of the sample or measurement values) and the other includes all of the program running on the personal computer 22 (HMM training).

In a further embodiment of the invention, the circuit complex is present 16 times and coupled to the circuitry 19 in the manner described; this allows the query of a biosensor 14 with n channels for n different analytes. The software and the display unit 25 are adapted to the number of channels of the biosen sensor 14 .

LIST OF REFERENCE NUMBERS

I curve section of the adjustment phase
II curve section characterizing the beginning of the measuring section
III measurement section
IV curve section characterizing the end of the measuring section

1

model curve

2

feature extraction

3

vector

4

Forward algorithm

5

Viterbi algorithm

6

Test for reaching the final state

7

Test for exceeding the maximum time

8th

Test for reaching the measuring state

9

Cancellation of the procedure

10

measurement value

11

Sequence of current values

12

state sequence

13

observation symbols

14

biosensor

15

Data line between D / A converter and biosensor

16

Circuit complex for analog preprocessing and conversion

17

D / A converter

18

internal bus

19

Circuit complex for digital processing

20

keypad

21

interface

22

personal computer

23

Storage

24

microcontroller

25

display unit

26

A / D converter

27

I / U-converter

28

Data line between biosensor and I / U converter

Claims (22)

1. A method for automatic online analysis of measurement series, characterized in that the sample values recorded in the measurement series are automatically assigned a model curve determined by a mathematical model during the measurement and, depending on this assignment, determines the sample values relevant for the analysis and be evaluated and the end of the measurement is determined. 2. The method according to claim 1, characterized in that the mathematical model is a trainable statis is a signal model. 3. The method according to claim 2, characterized in that the statistical signal model is a discrete left Right Hidden Markov Model (HMM) is. 4. The method according to claim 2, characterized in that the model in the training phase using experts knowledge is initialized interactively.   5. The method according to claim 2, characterized in that represent the statistical signal model with a tative number of previously determined and from Exper th evaluated curve profiles is trained. 6. The method according to claim 3, characterized in that as observation symbols for the HMM the 1st to nth Derivation of the curve of the samples used the. 7. The method according to any one of claims 2 to 6, characterized in that
in the working phase of the trained model, the most likely model is determined from a group of different models for a measurement curve that is currently (online) emerging during a measurement, using the forward algorithm,
then the most probable sequence of states for this selected model is calculated using the Viter bi algorithm,
the position of the points in time of the sample values necessary for determining the measured value is determined and
the end of the measurement is detected. 8. The method according to any one of claims 3 to 7, characterized in that as observation symbols for the discrete HMM the 1st and 2nd derivative of the measurement signal-time curve be used. 9. The method according to any one of claims 1 to 8, characterized in that when using the method on the determination of concentration in liquid media by means of amperometric biosensors, the concentration value sought is formed on the basis of expert knowledge from two current values, which two re levant Areas each correspond to a corresponding time series of samples, the first relevant area during the immersion of the sensor in buffer solution when the basic current I _{0 is} detected and the second relevant area when the sensor is immersed in the measuring or calibration solution when the measuring current is being acquired I _{D is} recorded and temporarily stored, the temporal course of the sampling values both when determining the base current I ₀ and when determining the measuring current I _{D is described} with at least one trainable, statistical signal model which is generated online from the possible curve types during the measurement process selects the most suitable, the end of the measurement is detected and the points in time are determined at which the base current I ₀ and the measurement current I _{D are} to be calculated from the respectively current sample values and used for calculating the concentration. 10. The method according to claim 9, characterized in that the statistical signal model before use Evaluation of newly created measurement curves with a right presentative number of previously determined and of Experts evaluated traces of the curve trained will (training phase). 11. Device for the automatic online analysis of measurement series,
characterized in that
them over
at least one means for sensory data acquisition,
at least one computer system with an execution environment in which an application is running which automatically assigns the sample values recorded in a series of measurements during the measurement to a model curve determined by a mathematical model and, depending on this assignment, the sample values relevant for the analysis determines and evaluates and determines the end of measurement,
at least one means for storing the recorded data and the mathematical model, and
Means for exchanging and / or displaying data
has, wherein the components mentioned are connected to each other with tel for data transmission. 12. The device according to claim 11, characterized in that the means for the sensory recording of data is connected to the device or one with the Device connectable single component. 13. Device according to one of claims 11 or 12, characterized in that the means for the sensory detection of data is a biosensor ( 14 ). 14. The apparatus according to claim 11, characterized in that the computer system consists of a microcontroller ( 24 ) and / or a personal computer ( 22 ). 15. The apparatus according to claim 11, characterized in that the device is divided into at least one circuit complex for analog preprocessing and conversion ( 16 ) of a sensor signal and at least one circuit complex for digital further processing ( 19 ) of a sensor signal, the circuit complex for analog Preprocessing and conversion ( 16 ) of the sensor signal at least one D / A converter ( 17 ), an A / D converter and an I / U converter ( 27 ) and the circuit complex for digital further processing ( 19 ) of the sensor signal at least one button field ( 20 ), a memory ( 23 ), a microcontroller ( 24 ) and a display unit ( 25 ). 16. The apparatus according to claim 15, characterized in that the device comprises a plurality of circuit complexes for analog preprocessing and conversion ( 16 ) of a sensor signal. 17. The apparatus of claim 11,
characterized in that
the means to exchange data at least
an internal bus ( 18 ) and / or
a keypad ( 20 ) and / or
an infrared interface ( 21 )
include. 18. The apparatus according to claim 11,
characterized in that
at least the means for displaying data
a display and / or
a monitor
comprise as a display unit ( 25 ). 19. Computer program product for automatic online Analysis of series of measurements, directly in the respective internal memory of a digital computer and / or microcontrollers can be loaded and soft includes code sections with which the steps are carried out according to claim 1 when the pro product on the computer and / or microcontroller running.   20. Computer program product according to claim 19, characterized in that the computer program product consists of two modules, one of which on a microcontroller ( 24 ) running program parts and the other on a personal computer ( 22 ) includes program parts le. 21. Computer-readable storage medium on which a pro gram is stored that it is a computer possible after it is in the computer's memory has been loaded an automatic online Carry out analysis of series of measurements, the in the series of samples recorded during the Measurement automatically one by a mathematical Model determined model curve assigned and in Depending on this assignment the for the Ana relevant samples determined and evaluated be tet and the end of the measurement is determined. 22. Computer-readable storage medium according to claim 21, characterized in that the computer-readable storage medium comprises two data carriers, program parts running on the first data carrier on a microcontroller ( 24 ) and program parts running on the second data carrier on a personal computer ( 22 ) are stored.