JP2013042904A - Method and device for automating initial setting of abnormal signal detection system for digital data having approximate periodicity - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method and a device for reducing the time of an initial setting in a conventional method, and automating the initial setting of a probability model and a threshold value.SOLUTION: This method for automating the initial setting of an abnormal signal detection system for digital data having approximate periodicity includes: a learning series setting process acquiring a learning series from a partial series of the digital data having the approximate periodicity; a probability model construction process constructing the probability model from the learning series; a data compression process compressing the learning series utilized for the construction of the probability model by use of the constructed probability model; and a threshold value setting process setting the detecting threshold value to the probability model by use of data compressed in the data compression process.

Description

 The present invention relates to a method and an apparatus for automating the initial setting of an abnormal signal detection system for digital data having almost periodicity. Specifically, the present invention relates to a method for automating initial settings and an apparatus for automating an arrhythmia waveform detection apparatus.

 Several methods have been proposed for data compression of electrocardiogram data, which is digital data with an approximate periodicity. For example, a method has been proposed in which a waveform including an excess portion is determined as an abnormal waveform (arrhythmia) when an average compression rate within a fixed-width time window exceeds a given threshold (for example, non-patented). (Ref. 1). In addition, an arrhythmia detection method that uses the mutual information amount of two electrocardiogram data obtained from the same patient instead of the compression rate has been proposed (see, for example, Patent Document 1).

 FIG. 4 shows a flowchart of the method of Non-Patent Document 1. In this method, the detection probability model and threshold value initial setting step (S1), the ECG data is compressed using the previous probability model (S2), and the compression ratio exceeds the previous threshold value. In this case, the arrhythmia is detected based on the flow of the step (S3) for determining that it is an arrhythmia.

 In S1, first, extract all normal rhythms (Y set) from the ECG data for a predetermined number of beats (X beats), and set N sets so that the total time of the set falls within the predetermined time range from the set of Y sets Take out. A method for obtaining a predetermined time range for given data is shown in Non-Patent Document 2.

 Next, the respective anti-dictionaries A1, A2,..., AN are constructed from the N sets of normal tunings W1, W2,. Next, a set of elements common to all of A1, A2,..., AN (anti-dictionary A) is obtained.

 Here, for a given data series, its own pattern does not appear, but a pattern with one leading data and a pattern with one trailing data appear in the given data series. A series is called a minimum prohibited word, and a collection of all the minimum prohibited words for a given data series is called an anti-dictionary.

 Next, a probability model for use in compression is constructed from the anti-dictionary A. In the method of Non-Patent Document 1, a combination of W1, W2,..., WN is called a learning sequence.

 Subsequently, in order to obtain a probabilistic model M and threshold value T with good detection accuracy, it can be set as a learning sequence and threshold value that can be arbitrarily selected from the Y set so that the total time of the set satisfies the predetermined time range Preliminary experiment (initial setting) for detecting arrhythmia for the whole ECG data for all combinations of numerical values, and the probability model M that uses the combination of the probability model and the threshold value with the best detection rate. Set as threshold T.

 In S2, as shown in the compression process (D1) of FIG. 1, the electrocardiogram data is compressed using the probability model M, and an average compression rate R within a time window having a fixed time width is calculated. For example, as shown in FIG. 7, when the upper electrocardiogram data is compressed, lower compression ratio data is obtained.

 In S3, as shown in the detection process (D2) of FIG. 1, when the average compression rate R exceeds the threshold value T, the waveform including the head position of the data within the time window is determined as an arrhythmia. For example, as shown in FIG. 7, in the lower compression ratio data, if the threshold value T = 1.5 is set, the waveform including the head position of the data exceeding the threshold value T = 1.5 is determined as an arrhythmia. The

 On the other hand, in the method of Patent Document 1, in order to calculate the mutual information of two or more electrocardiogram data, a plurality of electrodes are installed in a patient, and two or more electrocardiogram data obtained from different electrodes from the same patient. It is necessary to measure.

 In the measurement of normal ECG data, the multiple ECG data required for the method of Patent Document 1 cannot be obtained. Therefore, the method of Patent Document 1 is not suitable for detecting abnormal waveforms in general ECG data. There is a point.

JP 2008-200120 A

Takahiro Ota, Hiroyoshi Morita, 2 others, “arrhythmia detection using anti-dictionary coding method”, May 14, 2010, IEICE Technical Report, MBE, ME and Bio Cybernetics, Volume 110, 52, 35-40, The Institute of Electronics, Information and Communication Engineers Takahiro Ota, Hiroyoshi Morita, “One-pass distortion-free ECG using anti-dictionaries”, IEICE Transactions A, J87-A, No. 9, pp. 1187-1195, September 2004, incorporated association The Institute of Electronics, Information and Communication Engineers

 The method of Non-Patent Document 1 can perform abnormal waveform detection on general electrocardiogram data. However, as the first problem of the method of Non-Patent Document 1, since the entire electrocardiogram data is required for setting the probability model M and the threshold T value, the arrhythmia cannot be detected simultaneously with the measurement. There is a problem.

 In addition, there are many combinations of normal tuning that can be used as learning sequences in the electrocardiogram data. Further, regarding the threshold value, all the values within the range between the lower limit value and the upper limit value of the compression rate when the electrocardiogram data is compressed are applicable. Therefore, as a second problem, in the method of Non-Patent Document 1, there are a large number of combinations of probability models and thresholds that must be examined in the initial setting, and this processing takes time. is there.

 The present invention solves the above two problems, shortens the initial setting time of the method of Non-Patent Document 1, and further provides a method and apparatus for automating the initial setting of the probability model and the threshold value. The purpose is to provide.

 The present invention is a method for automating the initial setting of an anomaly signal detection system for digital data having an approximately periodicity, wherein a learning sequence is obtained by acquiring a learning sequence from a partial sequence of the digital data having an approximately periodicity. A probability model construction step of constructing a probability model from the learning sequence, a data compression step of compressing the learning sequence used to construct the probability model using the constructed probability model, and the data compression There is provided a method for initial setting of an abnormal signal detection system, including a threshold setting step for setting a detection threshold for the probability model using the data compressed in the step.

Further, according to the present invention, the threshold value is represented by the formula 1:
TI = (INT (RI × C) +1) / C
(In the formula, C is 100, INT (RI × C) represents the integer part of RI × C, and RI represents the maximum value in the compression ratio.)
The above method is calculated by:

 Furthermore, the present invention is an apparatus for automating the initial setting of an abnormal signal detection system for digital data having a substantially periodicity, wherein a learning sequence is obtained from a partial sequence of the digital data having the substantially periodicity. A sequence setting means, a probability model construction step for constructing a probability model from the learning sequence, a data compression means for compressing the learning sequence used for constructing the probability model using the constructed probability model, and There is provided an apparatus comprising threshold value setting means for setting a threshold value for detection for the probability model using data compressed in a data compression step.

Further, according to the present invention, the threshold value is represented by the formula 1:
TI = (INT (RI × C) +1) / C
(In the formula, C is 100, INT (RI × C) represents the integer part of RI × C, and RI represents the maximum value in the compression ratio.)
A method for initial setting of the abnormal signal detection system described above is provided.

 Further, the present invention is an arrhythmia waveform detection device, a learning sequence setting means for acquiring a learning sequence from a partial sequence of electrocardiogram data, a probability model construction step of constructing a probability model from the learning sequence, Data compression means for compressing the learning sequence used for constructing the probability model using the constructed probability model, and a threshold for detection for the probability model using the data compressed in the data compression step An arrhythmia waveform detection device comprising: threshold setting means for setting

 In addition, the present invention provides a detection process for test data, wherein an abnormal signal detection is performed on test data using the constructed probability model and the set threshold value pair, and the obtained detection result The detection method further includes a detection pair setting step of selecting a pair of a probability model and a threshold for use in detecting an abnormal signal with respect to an actual signal.

In addition, the present invention provides a detection means for test data, which detects an abnormal signal for test data using the constructed probability model and the set threshold value pair, and the obtained detection result A detection pair setting means for selecting a probability model and threshold pair for use in detecting an abnormal signal for an actual signal,
The above apparatus is further provided.

 According to the method and apparatus of the present invention, it is possible to shorten the initial setting time in an abnormal signal detection system for digital data having almost periodicity such as an arrhythmia waveform in electrocardiogram data. In addition, according to the method and apparatus of the present invention, the initial setting of the probability model and the threshold value can be automated in the abnormal signal detection system for digital data having almost periodicity such as an arrhythmia waveform in electrocardiogram data. Therefore, according to the method and apparatus of the present invention, by preparing test data, everything from data input to arrhythmia detection can be automated.

FIG. 5 is a flowchart of an arrhythmia detection processing portion of a conventional method described in Non-Patent Document 1. The figure showing the relationship of the position which acquires a learning series from electrocardiogram data in the method of this invention. The figure showing the relationship between the length of the learning series from ECG data, and the predetermined time range in the method of the present invention. FIG. 3 is a flowchart of a conventional method described in Non-Patent Document 1. The flowchart which shows the probability model which concerns on one Embodiment of the method of this invention, and the setting method of a threshold value. The figure which showed the process in the abnormal signal detection apparatus of the arrhythmia based on one Embodiment of the apparatus of this invention. The figure which shows the compression data and threshold value which are obtained from electrocardiogram data.

 Hereinafter, a method for automating the initial setting of an abnormal signal detection system for digital data having substantially periodicity according to the present invention will be described with reference to the drawings. A case where electrocardiogram data is particularly used will be described as an example of digital data having almost periodicity.

 In the method of the present invention, the abnormal signal detection system is initialized according to the steps shown in FIG. In the method of the present invention, first, in the learning sequence setting step (ST1), a learning sequence is acquired from a partial sequence of digital data having almost periodicity. In the method of the present invention, the learning sequence is constructed using only a portion of the data, not the entire data. For example, in order to set a learning sequence, as shown in FIG. 2, a normal tuning rhythm is acquired for each predetermined number of beats (X beats) in order from the beginning of a series of electrocardiogram data. As shown in FIG. 2, in the electrocardiogram data, normal tuning 1 in units of a predetermined number of beats (X beats) is represented as (W1, W2,..., WN), respectively. A combination of normal tuning 1 (W1, W2,..., WN) in units of the predetermined number of beats (X beats) is set as a learning sequence. On the other hand, the abnormal waveform 2 is excluded from the learning sequence. Further, as shown in FIG. 2, the normal rhythm of A beat that is less than X beats is also excluded from the learning sequence.

 Therefore, in the method of the present invention, the learning sequence refers to a series of digital data having an almost periodicity, in which data acquired in units of a predetermined cycle (X cycle) is concatenated. Further, the learning sequence does not include data that is less than a predetermined cycle number unit. Further, the learning sequence does not include data of a portion having no periodicity.

 In the learning sequence setting step (ST1), as shown in FIG. 3, K learning sequences can be created. The created learning sequences are, for example, L1,. Hereinafter, a case where there are a plurality of learning sequences LJ (J = 1,..., K) will be described.

 Next, in the probability model construction step (ST2), a probability model is constructed for the learning sequence set in ST1. In order to construct a probabilistic model, first, for the learning sequence LJ (J = 1,..., K), the anti-dictionaries A1,. Is generated.

 In this specification, for a given data series, its own pattern does not appear, but a pattern with one leading data and a pattern with one trailing data appear in the given data series. Such a series is called a minimum prohibited word. A collection of all minimally prohibited words for a given data series is called an anti-dictionary. The anti-dictionary can be generated as described in Non-Patent Document 1. Briefly, an anti-dictionary can be generated as follows:

 First, a dictionary having a data structure in which all patterns appearing in the data series are registered is generated. Next, 1 data is added to the end of the pattern, and it is checked whether or not the pattern satisfies the three conditions of the minimum prohibited word. If the condition is satisfied, the pattern is registered in the anti-dictionary. This is performed for all patterns registered in the dictionary. For example, T. Ota and H. Morita, “On the Construction of an Antidictionary with Linear Complexity Using the Suffix Tree”, IEICE Transactions on Fundamentals of Electronics Communications and Computer Sciences, Vol. E90-A, No. 11, pp. 2533-2539, November 2007, The Institute of Electronics, Information and Communication Engineers (IEICE).

 Subsequently, a set AJ of elements common to all of A1,..., AN is obtained. Next, a probability model MJ is constructed from the anti-dictionary AJ. Briefly, a probability model can be generated as follows:

 The probabilistic model based on the anti-dictionary is a deterministic finite automaton that does not accept a minimum forbidden word as a subsequence. As a generation method, first, all the minimum prohibited words included in the anti-dictionary are expressed by a data structure called a tree structure often used for data retrieval. Each contact point of the generated tree corresponds to one of the head data string of a certain minimum forbidden word (including the minimum forbidden word itself). Next, a new data string in which one arbitrary data (Z) is added to the end of the data string corresponding to the arbitrary node A is prepared, and the node B corresponding to the new data string is searched. If there is no corresponding node, the search is performed until node B is found while cutting the first data of the new data string. If node B is found, node B is set as a transition destination from node A to data Z. If this process is performed when one piece of data that is likely to appear is added to all nodes, a probability model is generated. For example, M. Crochemore, F. Mignosi, and A. Restivo, “Automata and Forbidden Words”, Information Processing Letters, Vol. 67, No. 3, pp. 111-117, October 1998. There is a method shown in.

 Therefore, in the method of the present invention, the probability model is a deterministic finite automan that does not accept a sequence including elements of A1,..., AN as subsequences in the anti-dictionary A1,. is there.

 Next, in the threshold value setting step (ST3) for the probability model, the learning sequence used to construct the probability model is compressed using the constructed probability model. When there are a plurality of learning sequences LJ (J = 1,..., K), K probability models M1,..., MK are used to compress the learning sequences L1,. To do. Briefly, the learning sequence can be compressed as follows.

 In order to perform compression, the nodes on the automaton are transitioned from the initial nodes of the deterministic finite automaton (probabilistic model) by the data series. In the transition, when a node has only one node as a transition destination in the transition, one data used for the transition is deleted, and when a node has two or more transition destinations, each node The number of transitions is counted, and the appearance probability of each data is performed using a compression method called an entropy code. Specifically, there is a method shown in T. OTA, “Antidictionary Data Compression Using Dynamic Suffix Trees”, Ph.D. Thesis, The University of Electro-Communications, 2007.

 Next, threshold values T1,..., TK for the probability models M1,..., MK are determined by Equation 1 for each maximum value R1,.

Formula 1 TI = (INT (RI × C) +1) / C
In the formula, C is 100, INT (RI × C) represents the integer part of RI × C, and RI represents the maximum value in the compression ratio.

 By the above process, in the initial setting of the abnormal signal detection system for digital data with almost periodicity, digital data with periodicity and abnormal signals are distinguished from digital data with almost periodicity such as ECG. The threshold can be set automatically. Specifically, as shown in FIG. 7, the above process can automatically set the lower compression rate data and the threshold value (T = 1.5 in the case of FIG. 7) from the upper electrocardiogram data. it can.

 In the method of the present invention, when there are a plurality of learning sequences LJ (J = 1,..., K), as shown in FIG. 5, a detection step (ST4) for test data can be further included.

 In the detection process (ST4) for test data, an abnormal signal is output for test data with a length within a predetermined time range using a probability model and threshold pair (M1, T1), ..., (MK, TK). Experiments for detection (in the case of an electrocardiogram, arrhythmia detection) can be performed. Here, it is preferable to use the test data that has been subjected to normal tuning and other discrimination for each waveform by a doctor or the like. The detection accuracy is determined using the frequencies tp, fn, tn, and fp shown in Table 1, and the sensitivity (sensitivity) = tp / (tp + fn) × 100 (%), specificity = tn / (tn + fp) × 100 (%).

 In the detection probability model and threshold setting step (ST5), the probability model / threshold pair (MJ, TJ) that maximizes the sum of the sensitivity and specificity obtained in the detection step (ST4) for the test data. Set as a probabilistic model for detection and threshold (M, T). If there are multiple pairs that take the maximum value, the one with the longest learning sequence is selected.

 In the above method, the predetermined number of beats, the predetermined time range, and the length of the test data may be changed according to the patient's condition and the measurement accuracy of the electrocardiogram data.

 In the above method, C can be changed to a power of 10 that is 10 or more.

 Moreover, in the said invention, the width | variety of the time window concerning average compression rate acquisition can be changed according to the measurement accuracy of electrocardiogram data.

 In the above description of the method, the electrocardiogram data has been described as the digital data having a substantially periodicity, but digital data having an arbitrary approximate periodicity may be used instead. In this case, instead of detecting normal tuning rhythm and arrhythmia, it is possible to detect normal signals and abnormal signals in digital data, respectively.

 FIG. 6 shows a circuit configuration diagram of a probability model and threshold automatic setting device using the method of the present invention. An example is shown in which electrocardiogram data is used as digital data having almost periodicity, and an arrhythmia in the electrocardiogram is detected as an abnormal signal.

 In FIG. 6, the PC 5 reads electrocardiogram data in which a normal waveform and an abnormal waveform have been identified from the input / output device 4 as an input. On the other hand, a detection probability model and threshold pair are output to the abnormal signal detection device 6 as outputs. The calling order of the respective means 13 to 18 in the calculation block is stored in the program (ROM) 12.

 As a detailed procedure, in the PC 5, the CPU 11 first uses the learning sequence setting means 13 of the calculation block 7 from the input identified electrocardiogram data to generate a learning sequence LJ (J = 1,..., K). Generate. The generated learning sequence is stored in the anti-dictionary table (RAM) 8.

 Subsequently, the CPU 11 uses the anti-dictionary construction means 14 of the calculation block 7 from the learning sequence LJ (J = 1,..., K) stored in the learning sequence / anti-dictionary table (RAM) 8 to Build the dictionary AJ (J = 1, ..., K). The constructed anti-dictionary AJ is stored in a learning sequence / anti-dictionary table (RAM) 8.

 Subsequently, the CPU 11 uses the probability model construction means 15 from the anti-dictionary AJ (J = 1,..., K) stored in the learning sequence / anti-dictionary table (RAM) 8 to correspond to the probability model MJ corresponding to each. (J = 1, ..., K) is constructed. The constructed probability model MJ is stored in a probability model / threshold table (RAM) 9.

 Subsequently, the CPU 11 uses the probability model MJ (J = 1,..., K) stored in the probability model / threshold value table (RAM) 9 and the data compression unit 10 to learn a sequence / anti-dictionary table (RAM) 8. The learning sequence LJ used to construct the MJ stored in is compressed. After the compression, the maximum value RJ of the compression rate is obtained using the maximum value detecting means 16 of the calculation block 7. Further, the threshold value TJ is obtained from the RJ and Formula 1 using the threshold value setting means 17. The obtained threshold value TJ is stored in a probability model / threshold value table (RAM) 9.

 In order to distinguish digital data with periodicity and abnormal signals from digital data with approximate periodicity, such as an electrocardiogram, in the initial setting of the abnormal signal detection system for digital data with approximate periodicity by the procedure described above. The threshold value can be set automatically.

 In the apparatus of the present invention, the learning sequence LJ can be one, but when generating a plurality of learning sequences LJ (J = 1,..., K) as described above, as shown in FIG. In addition, the following processing can be performed.

 The CPU 11 has identified the probability model and threshold pair (MJ, TJ) (J = 1, ..., K) stored in the probability model / threshold value table (RAM) 9 and the normal waveform and abnormal waveform. Based on the test data, arrhythmia detection is performed using the test data detection means 18 of the calculation block 7 to obtain sensitivity SeJ and specificity SpJ. Here, as the test data, the identified electrocardiogram data used for learning sequence generation can be used.

 Subsequently, the pair of (MJ, TJ) that maximizes the sum of SeJ and SpJ is output from the PC 2 to the abnormal signal detection device 6 as a detection probability model and threshold pair (M, J).

 The method of selecting (M, T) may be changed to another index such as the maximum value of specificity or sensitivity instead of the maximum value of the sum described above.

 According to the apparatus of the present invention, by preparing test data for any input data, it is possible to automate the initial setting and reduce the processing time in the apparatus using the method of Non-Patent Document 1. . For this reason, it is possible to increase the utility value of the device using the method of Non-Patent Document 1.

 Examples of the arrhythmia detection method using the method of the present invention will be described below with reference to Tables 2 to 8. In this example, ventricular extrasystole (PVC) was detected as an arrhythmia. First, Table 2 shows the specifications of the electrocardiogram data according to this example. Table 3 shows the total number of beats and the number of PVCs in the electrocardiogram file (file A) of the detected patient.

 In this example, the predetermined time range is 20,722 samples (about 58 seconds) to 72,762 samples (about 202 seconds). The width of the time window for obtaining the average compression rate is 5 samples (about 0.01 seconds). The predetermined number of beats X is 11 beats, and the length of the test data is 72,000 samples (200 seconds).

 Table 4 shows the specifications of six learning sequences consisting of 11 × 17 beats, 11 × 18 beats,..., 11 × 23 beats in the predetermined time range obtained in ST1 shown in FIG.

 Next, Table 5 shows the number of minimum prohibited words, the anti-dictionary size, and the probability model size included in the anti-dictionary obtained from the six learning waveforms obtained in ST2 shown in FIG. In the table, 1 byte is 8 bits.

 Next, Table 6 shows threshold values determined by ST3 shown in FIG. 5 for six learning sequences consisting of 11 × 17 beats, 11 × 18 beats,..., 11 × 23 beats within a predetermined time range. .

 Table 7 shows the detection results for the test data shown in ST4 of FIG. 5 for the learning sequence of 11 × 17 beats, 11 × 18 beats,..., 11 × 23 beats. As shown in Table 7, for the test data, the sum of the sensitivity and specificity of 11 × 19 beats is large for the other five learning sequences. The probability model generated from the 11 × 19 beat learning waveform with the largest sum of sensitivity and specificity and its threshold value 1.40 are the same as the probability model and threshold value for file A of the arrhythmia detection system of Non-Patent Document 1. To do.

 Table 8 shows the results of a PVC detection experiment for file A using the arrhythmia detection system of Non-Patent Document 1, using a probability model M generated from an 11 × 19 beat learning waveform and a threshold T = 1.40. .

 As shown in Table 8, it can be seen that PVC can be detected with a sensitivity of 100% and a specificity of 95.0% by using the threshold generated by the method of the present invention.

 The method and apparatus of the present invention can be operated in real time with a small amount of memory, and is particularly effective for use in a portable measuring instrument. Therefore, the method and apparatus of the present invention can enhance the convenience of portable medical devices used for abnormal signal detection and home medical care in manufacturing machinery and the like.

1 Normal tuning (included in learning sequence)
2 Abnormal waveform
3 Normal tuning (not included in the learning sequence)
4 I / O devices
5 PC (computer)
6 Abnormal signal detector
7 Computation block
8 Learning series / anti-dictionary table (RAM)
9 Stochastic model and threshold table (RAM)
10 Data compression section
11 CPU
12 Program (ROM)
13 Learning sequence setting method
14 Anti-dictionary construction means
15 Stochastic model construction means
16 Maximum value detection means
17 Threshold setting method
18 Test data detection

Claims (7)

A method for automating the initial setting of an anomaly signal detection system for digital data having an approximate periodicity,
A learning sequence setting step of acquiring a learning sequence from a partial sequence of digital data having an approximately periodicity,
Probabilistic model construction step of constructing a probabilistic model from the learning sequence;
Using the constructed probability model, a data compression step of compressing the learning sequence used for constructing the probability model;
Using the data compressed in the data compression step to set a detection threshold for the probability model; and a threshold setting step;
A method for initial setting of an abnormal signal detection system, including: The threshold is given by Equation 1:
TI = (INT (RI × C) +1) / C
In the formula, C is 100, INT (RI × C) represents the integer part of RI × C, and RI represents the maximum value in the compression ratio.
The method for initial setting of the abnormal signal detection system according to claim 1, wherein An apparatus for automating the initial setting of an abnormal signal detection system for digital data having a substantially periodicity,
A learning sequence setting means for acquiring a learning sequence from a partial sequence of digital data having an approximately periodicity,
Probabilistic model construction step of constructing a probabilistic model from the learning sequence;
Data compression means for compressing the learning sequence used for the construction of the probability model using the constructed probability model;
A threshold value setting means for setting a threshold value for detection with respect to the probability model using the data compressed in the data compression step;
A device comprising: The threshold is given by Equation 1:
TI = (INT (RI × C) +1) / C
In the formula, C is 100, INT (RI × C) represents the integer part of RI × C, and RI represents the maximum value in the compression ratio.
The method for initial setting of the abnormal signal detection system according to claim 1, wherein A device for detecting an arrhythmia waveform,
Learning sequence setting means for acquiring a learning sequence from a partial sequence of electrocardiogram data;
Probabilistic model construction step of constructing a probabilistic model from the learning sequence;
Data compression means for compressing the learning sequence used for the construction of the probability model using the constructed probability model;
A threshold value setting means for setting a threshold value for detection with respect to the probability model using the data compressed in the data compression step;
A device for detecting an arrhythmia waveform. Using the constructed probability model and the set threshold value pair to perform an abnormal signal detection on the test data;
From the obtained detection result, selecting a pair of detection model and threshold for use in detecting an abnormal signal for an actual signal, a detection pair setting step,
The method according to claim 1 or 2, further comprising: A detection means for test data, which performs an abnormal signal detection on test data using the constructed probability model and the set threshold value pair;
From the obtained detection result, a pair of detection pair setting means for selecting a probability model and threshold pair for use in detecting an abnormal signal for an actual signal;
5. The apparatus according to claim 3 or 4, further comprising: