Prediction device, method of generating prediction model, and computing device
The reasoning device addresses the challenge of accuracy degradation in inference models by using multiple interpolation methods to generate a stable range for interpolated values, ensuring accurate inference even with unstable time series data.
Patent Information
- Application Number
- JP2023179592
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-10-18
- Publication Date
- 2025-05-01
AI Technical Summary
Inference models face accuracy degradation when interpolating observed values in environments with unstable time series data for observed parameters.
A reasoning device that accepts time series data and uses multiple interpolation methods to generate a range of possible values, reducing bias and maintaining inference accuracy even with limited data.
The solution effectively suppresses the degradation of inference accuracy by providing a reasonable range for interpolated values, even in environments with inconsistent time intervals for data acquisition.
Smart Images
Figure 2025069703000001_ABST
Abstract
Description
[Technical field]
[0001] The present disclosure relates to an inference device that infers a possible range of values for interpolating a plurality of observed values by inputting time series data including a plurality of observed values obtained from a subject for an observable parameter for obtaining bioinformation of the subject into an inference model. The present disclosure also relates to a method and a computing device for generating the inference model. [Background technology]
[0002] Inference models generated using machine learning technology to replace or support clinical judgments of medical professionals in clinical settings are known. For example, Non-Patent Document 1 discloses a model that infers specific biological information based on time series data of observed parameters. An example of specific biological information is the remaining stay time in an intensive care unit (ICU). Non-Patent Document 1 reports that the inference model shows excellent performance for unobserved parameters that are frequently acquired, such as heart rate, but does not show good performance for unobserved parameters that are intermittently acquired (for which a sufficient number of time series data cannot be secured), such as white blood cell count. [Prior art documents] [Non-patent literature]
[0003] [Non-Patent Document 1] G. Harerimana, et al. "A Multi-Headed Transformer Approach for Predicting the Patient's Clinical Time-Series Variables From Charted VitalSigns", IEEE Access (Volume 10), 105993-106004, October 3, 2022 Summary of the Invention [Problem to be solved by the invention]
[0004] In an environment where it is difficult to obtain stable time-series data for observed parameters, it is necessary to suppress the decrease in inference accuracy of an inference model that uses interpolated values of observed values as input. [Means for solving the problem]
[0005] A first example aspect provided by the present disclosure is an inference device, comprising: an interface that accepts time-series data including a plurality of observation values obtained from a subject at different times for an observable parameter for obtaining biological information of the subject; a processor that inputs the time series data into an inference model and outputs a range of possible values that are used to interpolate the multiple observed values as an inference result; Equipped with The processor, generating at least one first interpolant by interpolating the plurality of observed values in a first manner; generating at least one second interpolant by interpolating the plurality of observed values in a second manner different from the first manner; A range defined by the maximum and minimum values of the first and second interpolated values is identified as the inference result.
[0006] In general, the shorter the time interval at which multiple observed values are obtained for the observed parameter, the less the need to interpolate between the multiple observed values using estimated values, and as a result, the deterioration of the inference accuracy of an inference model using time series data of the observed parameter as input can be suppressed. On the other hand, for example, in clinical medical practice, armpit temperature is generally measured once every few hours, and it may be difficult to secure a sufficient number of time series data for the observed parameter. However, according to the configuration of the first aspect, a range defined by a maximum value and a minimum value selected from multiple interpolated values generated through multiple different methods is specified as a reasonable range that the interpolated value can ultimately take. This makes it possible to suppress bias in the interpolation tendency that may occur due to reliance on a specific interpolation method. Therefore, in an environment where it is difficult to secure stable time series data for the observed parameter, it is possible to suppress the deterioration of the inference accuracy of an inference model using values obtained by interpolating observed values as input.
[0007] A second example aspect provided by the present disclosure is an inference device, comprising: an interface that accepts time-series data including a plurality of observation values obtained from a subject at different times for an observable parameter for obtaining biological information of the subject; a processor for inputting the time series data into an inference model to perform at least one inference on a range of possible values to be interpolated for the plurality of observed values; Equipped with The inference model includes a plurality of inference models that are different in response to the intervals of the different time points generated by machine learning; The processor uses one of the plurality of inference models depending on the interval.
[0008] In general, if the time interval at which multiple observed values of the observable parameter are obtained is constant, it is possible to suppress a decrease in the inference accuracy of an inference model in the range of possible non-observable values interpolated between the multiple observed values. On the other hand, for example, armpit temperature is generally measured manually by medical professionals in clinical medical settings, and the time interval at which observed values of the observable parameter are obtained may not be constant. However, according to the configuration of the second aspect, an appropriate one is selected from multiple inference models generated by different machine learning corresponding to the time interval to infer an interpolated value, so that it is possible to suppress a decrease in the inference accuracy of an inference model that uses an interpolated value of an observed value as an input when the time interval at which observed values of the observable parameter are obtained is not constant.
[0009] A third example aspect provided by the present disclosure is a method for generating an inference model for inferring biometric information by a computing device, the method comprising: acquiring observation data including a plurality of observation values acquired from a living body at a first time interval for an observable parameter for obtaining the biological information; generating training data including a plurality of observations at a second time interval longer than the first time interval by thinning out at least one of the plurality of observations; performing supervised learning using the learning data with the thinned-out observed value as a correct answer; a range of possible values of at least one value that interpolates a plurality of observed values of the observable parameter obtained from the living body at the second time interval; By varying the second time interval, a plurality of types of learning data are generated from the observation data.
[0010] A fourth example aspect provided by the present disclosure is a computing device for generating an inference model for inferring biometric information, comprising: an interface that receives observation data including a plurality of observation values obtained from a living body at a first time interval for an observable parameter for obtaining the biological information; A processor; Equipped with The processor, generating training data including a plurality of observations at a second time interval longer than the first time interval by thinning out at least one of the plurality of observations; performing supervised learning using the learning data with the thinned-out observed value as a correct answer; configuring the inference model to infer a range of at least one value that interpolates a plurality of observation values obtained from a living organism at the second time interval for the observable parameter; By varying the second time interval, a plurality of types of learning data are generated from the observation data.
[0011] By performing machine learning using learning data in which the second time interval is set to match the time interval of the multiple observed values included in the time series data acquired from the subject, the generated inference model can acquire the ability to identify, with at least one inference, a reasonable range that the non-observed value to be interpolated between the multiple observed values acquired at the time interval. If the time interval of the multiple observed values included in the time series data is not constant, multiple inference models can be prepared by performing machine learning using learning data with different second time intervals. This makes it possible to configure an inference device that can identify, with at least one inference, for each different time interval, a reasonable range that the non-observed value to be interpolated between the multiple observed values acquired at the time interval. [Brief description of the drawings]
[0012] [Figure 1] 1 illustrates an example of the functional configuration of an inference system according to one embodiment. [Diagram 2] 2 shows an example of a flow of processing performed in the inference device of FIG. [Diagram 3] 2 shows another example of processing performed in the inference device of FIG. 1. [Figure 4] 2 shows another example of the processing flow performed in the inference system of FIG. 1. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0013] The embodiments will be described in detail with reference to the accompanying drawings. In each drawing, the scale is appropriately changed so that each element to be described can be recognized.
[0014] 1 illustrates an example of a functional configuration of an inference system 1 according to an embodiment. The inference system 1 includes an inference device 2.
[0015] The inference device 2 according to this example is a device that predicts a possible range of armpit temperature values to be interpolated between a period between a plurality of time points based on time series data TS including a plurality of armpit temperature observation values obtained from a subject S at the plurality of time points. Armpit temperature is an example of an observed parameter. Other examples of observed parameters include respiratory rate obtained by visually observing chest movement, Glasgow Coma Scale (GCS) obtained by visually confirming the state of consciousness, urine volume measured in a urine collection bag, index values obtained through blood tests (red blood cell count, white blood cell count, blood lactate level, arterial blood oxygen partial pressure, etc.), and blood pressure obtained by non-invasive measurement.
[0016] As used herein, the term "time series of an observable parameter" refers to the change over time in the values of an observable parameter obtained at multiple time points, which may or may not be spaced apart by a regular interval.
[0017] The inference device 2 includes an input interface 21. The input interface 21 is configured as a hardware interface capable of receiving the above-mentioned time-series data TS. The time-series data TS may be input from an appropriate sensor or measuring device attached to the subject S, or may be input via an appropriate input device and user operation. The time-series data TS may be in the form of analog data or digital data, depending on the specifications of the input source. When the time-series data TS is in the form of analog data, the input interface 21 includes an appropriate conversion circuit including an A / D converter.
[0018] The inference device 2 includes a processor 22 and an inference model 23. The inference model 23 is a computer algorithm that can be executed by the processor 22. The inference model 23 is configured to output, as an inference result, a range of possible armpit temperature values that are interpolated between the periods between the multiple time points, based on time series data TS including multiple observed values of armpit temperature obtained from the subject S at the multiple time points.
[0019] The inference device 2 includes an output interface 24. The processor 22 is configured to input the time series data TS received by the input interface 21 to the inference model 23, and to output, from the output interface 24, result data RS corresponding to the inference result output from the inference model 23.
[0020] An example of a specific operation of the inference model 23 will be described with reference to Fig. 2. The time series data TS received by the input interface 21 includes an observation value v1 of armpit temperature acquired at time t1 and an observation value v2 of armpit temperature acquired at time t2.
[0021] The processor 22 is configured to generate a plurality of interpolated data sets in different ways based on the time series data TS. In this example, a first interpolated data set IT1, a second interpolated data set IT2, and a third interpolated data set IT3 are generated.
[0022] The first interpolated data set IT1 is obtained by interpolating three non-invasive blood pressure values in the period between time t1 and time t2. The interpolation method is based on the Last Observation Carried Forward (LOCF) method. In this method, the interpolation is performed such that an observation value at a certain time point is maintained until the next observation value is obtained. The number of interpolated values can be determined arbitrarily.
[0023] The second interpolated data set IT2 is obtained by interpolating three non-invasive blood pressure values in the period between time t1 and time t2. The interpolation method is based on linear interpolation. In this method, values corresponding to the observed values at any time point that are located on a straight line connecting two observed values are interpolated. The number of interpolated values can be determined arbitrarily.
[0024] The third interpolated data set IT3 is obtained by interpolating three non-invasive blood pressure values in the period between time t1 and time t2. The interpolation method is based on spline interpolation. In this method, a value corresponding to an observation value at any time point, which is located on a polynomial curve connecting two observation values, is interpolated. The number of interpolated values can be determined arbitrarily.
[0025] If the interpolation methods are different, the number of the multiple interpolated data sets generated by the processor 22 may be determined arbitrarily. Any interpolated data set selected from the multiple interpolated data sets generated may be an example of a first interpolated data set. Similarly, any other interpolated data set selected from the multiple interpolated data sets may be an example of a second interpolated data set. At least one interpolated value included in the first interpolated data set is an example of a first interpolated value. At least one interpolated value included in the second interpolated data set is an example of a second interpolated value.
[0026] Note that the expression "different interpolation methods" used in this specification is not intended to refer only to cases where the types of interpolation methods are different. As an example, an interpolation data set generated by spline interpolation using a cubic polynomial and an interpolation data set generated by spline interpolation using a quintic polynomial are interpreted as "different interpolation methods" even though they are the same type of "spline interpolation". As another example, multiple interpolation data sets that use the same type of interpolation method but have different numbers of values interpolated in the same period are also interpreted as "different interpolation methods".
[0027] Next, the processor 22 generates integrated data IG by integrating the first interpolation data set IT1, the second interpolation data set IT2, and the third interpolation data set IT3. Furthermore, the processor 22 identifies, as an inference result, a range RG defined by a maximum value MX and a minimum value MN among the multiple interpolation values included in the integrated data IG.
[0028] In general, the shorter the time interval at which multiple observed values are obtained for the observed parameter, the less the need to interpolate between the multiple observed values using estimated values, and as a result, the deterioration of the inference accuracy of the inference model 23 using the time series data of the observed parameter as input can be suppressed. On the other hand, there may be cases where it is difficult to secure a sufficient number of time series data for the observed parameter, such as the armpit temperature in this example. However, according to the configuration of this embodiment, the range defined by the maximum and minimum values selected from multiple interpolated values generated through multiple different methods is specified as a valid range that the interpolated value can ultimately take. This makes it possible to suppress bias in the interpolation tendency that may occur due to relying on a specific interpolation method. Therefore, in an environment where it is difficult to secure stable time series data for the observed parameter, it is possible to suppress the deterioration of the inference accuracy of the inference model using values obtained by interpolating observed values as input.
[0029] In particular, in this embodiment, a plurality of interpolated data sets are generated using different types of interpolation methods, which can further reduce the bias in the interpolation tendency described above.
[0030] 1, the inference system 1 may include an output device 3. A processor 22 of the inference device 2 is configured to output, from an output interface 24, control data CT that causes the output device 3 to visualize a possible range of values that interpolate an observed value inferred by an inference model 23. Examples of the output device 3 include a display device and a printing device. With this configuration, a user can confirm the validity of an inference result through visualization by the output device 3.
[0031] In this case, the output interface 24 is configured as a hardware interface capable of outputting the control data CT. The control data CT may be in the form of analog data or digital data, depending on the specifications of the output device 3. When the control data CT is in the form of analog data, the output interface 24 includes an appropriate conversion circuit including a D / A converter.
[0032] An example of a specific operation of the inference model 23 will be described with reference to Fig. 3. In this example, the time series data TS received by the input interface 21 includes an observation value v1 of the armpit temperature acquired at time t1, an observation value v2 acquired at time t2, and an observation value v3 acquired at time t3. In this example, the time intervals at which the observation values are acquired are not constant. That is, the time interval T12 between time t1 and time t2 is different from the time interval T23 between time t2 and time t3.
[0033] The inference device 2 according to the present example is configured to vary the number of non-observed values to be interpolated so as to have a positive correlation with the time interval over which the observed values are obtained. In other words, the longer the time interval over which the observed values are obtained, the more non-observed values to be interpolated.
[0034] In the example shown in Fig. 3, two non-observed values are interpolated between the observations v1 and v2 acquired in the time interval T12. On the other hand, three non-observed values are interpolated between the observations v2 and v3 acquired in the time interval T23. The symbol N shown in Fig. 3 represents the number of non-observed values to be interpolated.
[0035] As an example, the plausible range of each interpolated value may be inferred using the method described with reference to Fig. 2. That is, a range defined by a maximum and a minimum value selected from a plurality of interpolated values generated through a plurality of different methods is identified as the inference result.
[0036] 1, the inference system 1 includes an inference model generation device 4. An inference model 23 of the inference device 2 is generated through machine learning executed by the inference model generation device 4. The inference model generation device 4 includes an input interface 41, a processor 42, and an output interface 43. The inference model generation device 4 is an example of a computing device.
[0037] With reference to FIG. 4, a method for generating an inference model 23 according to another example executed by the inference model generating device 4 will be described.
[0038] A first observation data set OB1 including a plurality of observation values of armpit temperature acquired from a living subject at a first time interval T1 is received by the input interface 41. The first observation data set OB1 includes four observation values v11 to v14.
[0039] The processor 42 generates a first thinned data set TH1 by thinning at least one of the observations included in the first observation data set OB1. In this example, two observations v12 and v13 obtained between the earliest observation v11 and the latest observation v14 among the four observations are thinned. As a result, the first thinned data set TH1 includes a plurality of observations in a second time interval T2 that is longer than the first time interval T1.
[0040] Similarly, a second observation data set OB2 including a plurality of observation values of armpit temperature acquired from a living subject over a first time interval T1 is received by the input interface 41. The second observation data set OB2 includes four observation values v21 to v24.
[0041] The processor 42 generates a second thinned data set TH2 by thinning out two observation values v22, v23 obtained between the earliest observation value v21 and the latest observation value v24 among the four observation values included in the second observation data set OB2 in the same manner as in the case of the first observation data set OB1. Note that it is preferable that the observation values v21 and v24 left in the second thinned data set TH2 are substantially equal to the observation values v11 and v14 left in the first thinned data set TH1, respectively.
[0042] In this example, a third observation data set OB3 including a plurality of observations of armpit temperature taken from a living subject over a first time interval T1 is also received by the input interface 41. The third observation data set OB3 includes four observations v31 to v34.
[0043] The processor 42 generates a third thinned data set TH3 by thinning out two observation values v32, v33 obtained between the earliest observation value v31 and the latest observation value v34 among the four observation values included in the third observation data set OB3 in the same manner as in the case of the first observation data set OB1. Note that it is preferable that the observation values v31 and v34 left in the third thinned data set TH3 are substantially equal to the observation values v11 and v14 left in the first thinned data set TH1, respectively.
[0044] Next, the processor 42 performs supervised learning using the two observation values v12, v13 thinned out as described above using the first thinned data set TH1 as learning data TR, the two observation values v22, v23 thinned out as described above using the second thinned data set TH2 as learning data TR, and the two observation values v32, v33 thinned out as described above using the third thinned data set TH3 as learning data TR as correct answers.
[0045] That is, the above machine learning makes the neural network learn different patterns of changes from one observed value to another observed value that may occur during the second time interval T2. The inference model 23 generated by this machine learning is configured to infer, in response to input of time series data TS including two observed values acquired from the subject S during the second time interval T2, a possible range of each of two unobserved values that may be interpolated between the two observed values during the first time interval T1. The inference model 23 is configured to identify a maximum value MX and a minimum value MN among a plurality of unobserved values that may be interpolated between the two observed values, and to output a range RG defined by them as an inference result.
[0046] By performing machine learning using the learning data TR in which the second time interval T2 is determined so as to match the time interval of the multiple observed values included in the time series data TS acquired from the subject S, the generated inference model 23 can acquire the ability to specify, at least once inference, a reasonable range that the non-observed value to be interpolated between the multiple observed values acquired at that time interval can take. When the time interval of the multiple observed values included in the time series data TS is not constant, multiple inference models 23 can be prepared by performing machine learning using the learning data TR in which the second time interval T2 differs. This makes it possible to configure an inference device 2 that can specify, for each of the different time intervals, a reasonable range that the non-observed value to be interpolated between the multiple observed values acquired at that time interval can take at least once inference.
[0047] In general, if the time interval at which multiple observed values are obtained for an observable parameter is constant, the inference of the range of possible non-observable values to be interpolated between the multiple observed values can be uniformly applied to all intervals by one inference model. On the other hand, as in the case of armpit temperature in this example, the time interval at which observed values are obtained for an observable parameter may not be constant. However, according to the configuration of this example, an appropriate one of multiple inference models generated by different machine learning corresponding to the time interval is selected to infer the interpolated value, so that when the time interval at which observed values of the observable parameter are obtained is not constant, it is possible to suppress a decrease in the inference accuracy of the inference model that uses an interpolated value of the observed value as an input.
[0048] The training data TR having a different second time interval T2 is preferably generated by changing the thinning interval from the same observation data set. For example, by thinning out the observation values v12 and v14 from the observation values v11 to v14 included in the first observation data set OB1, it is possible to generate training data TR having a shorter second time interval T2 than the first thinned data set TH1.
[0049] With this configuration, an inference model that can handle cases where the time intervals between multiple observation values included in the time series data TS are not constant can be efficiently generated from a limited number of observation data sets.
[0050] Each of the processor 22 of the inference device 2 and the processor 42 of the inference model generating device 4 having the various functions described above can be realized by a general-purpose microprocessor that operates in cooperation with a general-purpose memory. Examples of the general-purpose microprocessor include a CPU, an MPU, and a GPU. Examples of the general-purpose memory include a ROM and a RAM. In this case, the ROM can store a computer program that executes the above-mentioned processing. The ROM is an example of a non-transient computer-readable medium that stores a computer program. The general-purpose microprocessor specifies at least a part of the program stored on the ROM, expands it on the RAM, and executes the above-mentioned processing in cooperation with the RAM. The computer program may be pre-installed in the general-purpose memory, or may be downloaded from an external server via a communication network and then installed in the general-purpose memory. In this case, the external server is an example of a non-transient computer-readable medium that stores a computer program.
[0051] Each of the processors 22 and 42 may be realized by a dedicated integrated circuit capable of executing the above computer program, such as a microcontroller, an ASIC, or an FPGA. In this case, the above computer program is pre-installed in a storage element included in the dedicated integrated circuit. The storage element is an example of a computer-readable medium storing a computer program. Each of the processors 22 and 42 may also be realized by a combination of a general-purpose microprocessor and a dedicated integrated circuit.
[0052] The configurations referred to above are merely examples for facilitating understanding of the present disclosure. Each configuration example may be appropriately modified or combined with other configuration examples within the scope of the present disclosure.
[0053] In the above embodiment, the inference device 2 and the inference model generation device 4 are described as devices independent of each other. In this case, the input interface 21 and the output interface 43 are configured as hardware interfaces. However, both devices may be realized as different functional modules in a single device. In this case, both interfaces may be realized as software interfaces.
[0054] The inference model 23 does not necessarily have to be installed in the inference device 2. Although not shown in the figure, the inference device 2 may be communicatively connected to an external server device via a communication network. In this case, the inference model 23 may be installed in the external server device.
[0055] In the above embodiment, the inference model generating device 4 generates the inference model 23 through machine learning using a neural network. However, the inference model 23 may be generated through other machine learning algorithms. Examples of other machine learning algorithms include linear regression, logistic regression, decision tree, random forest, and support vector machine. [Explanation of symbols]
[0056] 2: inference device, 21: input interface, 22: processor, 23: inference model, 3: output device, 4: inference model generation device, 41: input interface, 42: processor, MN: minimum value, MX: maximum value, OB1: first observation data set, OB2: second observation data set, RG: range, S: subject, T1: first time interval, T2: second time interval, TR: learning data, TS: time series data
Claims
1. an interface that accepts time-series data including a plurality of observation values obtained from a subject at different times for an observable parameter for obtaining biological information of the subject; a processor that inputs the time series data into an inference model and outputs a range of possible values that are used to interpolate the multiple observed values as an inference result; Equipped with The processor, generating at least one first interpolant by interpolating the plurality of observed values in a first manner; generating at least one second interpolant by interpolating the plurality of observed values in a second manner different from the first manner; identifying a range defined by a maximum value and a minimum value of the first interpolation value and the second interpolation value as the inference result; Reasoning device.
2. The first method and the second method have different types of interpolation methods. The inference device of claim 1 .
3. an interface that accepts time-series data including a plurality of observation values obtained from a subject at different times for an observable parameter for obtaining biological information of the subject; a processor for inputting the time series data into an inference model to perform at least one inference on a range of possible values to be interpolated for the plurality of observed values; Equipped with The inference model includes a plurality of inference models that are different in response to the intervals of the different time points generated by machine learning; The processor uses one of the plurality of inference models depending on the interval. Reasoning device.
4. The processor causes the range to be visualized on an output device.
4. An inference device according to claim 1 or 3.
5. A method for generating an inference model for inferring biometric information by a computing device, comprising: acquiring observation data including a plurality of observation values acquired from a living body at a first time interval for an observable parameter for obtaining the biological information; generating training data including a plurality of observations at a second time interval longer than the first time interval by thinning out at least one of the plurality of observations; performing supervised learning using the learning data with the thinned-out observed value as a correct answer; a range of possible values of at least one value that interpolates a plurality of observed values of the observable parameter obtained from the living body at the second time interval; generating a plurality of types of learning data from the observation data by varying the second time interval; How to generate an inference model.
6. A computing device for generating an inference model for inferring biometric information, comprising: an interface that receives observation data including a plurality of observation values obtained from a living body at a first time interval for an observable parameter for obtaining the biological information; A processor; Equipped with The processor, generating training data including a plurality of observations at a second time interval longer than the first time interval by thinning out at least one of the plurality of observations; performing supervised learning using the learning data with the thinned-out observed value as a correct answer; configuring the inference model to infer a range of at least one value that interpolates a plurality of observation values obtained from a living organism at the second time interval for the observable parameter; generating a plurality of types of learning data from the observation data by varying the second time interval; Computing device.