Method and state machine system for detecting operating state of sensor

Abstract

The invention relates to a method for detecting an operating state of a sensor, the method comprising, in a state machine: receiving continuous monitoring data relating to the operation of the sensor; providing a trained learning algorithm for detecting an operating state of the sensor indicative of the sensor function, where the learning algorithm is trained according to a training data set comprising historical data; detecting an operational state of the sensor by analyzing the continuous monitoring data with a trained learning algorithm; and providing output data indicative of the detected operating state of the sensor. Further, a state machine system is provided having one or more processors configured for data processing and performing a method for detecting an operational state of a sensor.

Description

[0001] This application is a divisional application, and the invention name of its parent case is "Method and state machine system for detecting the operating state of a sensor", the application date is June 29, 2018, and the application number is 201880043556.5. Technical Field

[0002] The present disclosure relates to a method and a state machine system for determining the operating state of a sensor. Background Art

[0003] Document US2014 / 0182350 A1 discloses a method for determining the end-of-life of a CGM (Continuous Glucose Monitoring) sensor, the method including using an end-of-life function to evaluate a plurality of risk factors to determine the end-of-life state of the sensor, and providing an output related to the end-of-life state of the sensor. The plurality of risk factors are selected from a list including: the number of days the sensor has been used, whether the signal sensitivity has decreased, whether there is a predetermined noise pattern, whether there is a predetermined oxygen concentration pattern, and the error between the reference BG (Blood Glucose) value and the EGV sensor value.

[0004] Document EP 2335584 A2 relates to a method for self-diagnostic testing and setting a pause operation mode of a continuous analyte sensor in response to the result of the self-diagnostic testing.

[0005] In document US2015 / 164386 A1, Electrochemical Impedance Spectroscopy (EIS) is used in combination with a continuous glucose monitor and continuous glucose monitoring (CGM), enabling in vivo sensor calibration, total (sensor) fault analysis, and intelligent sensor diagnosis and fault detection. An equivalent circuit model is defined, and circuit elements are used to characterize sensor behavior.

[0006] Document US2010 / 323431 A1 discloses a control circuit and method for controlling a bistable display having bistable segments, each bistable segment being capable of switching between an on state and an off state by applying a voltage. The voltage is provided from a charge pump to a display driver and supplied to respective segments in the bistable segments via an output according to a display instruction provided by a system controller. Both the bistable segment voltage level of at least one of the outputs of the display driver and the charge pump voltage level of the voltage are detected and compared with an effective bistable segment voltage level and an effective charge pump voltage level, respectively. If either of the detected voltage levels is invalid, a fault signal may be provided to the system controller. Summary of the Invention

[0007] The object of the present disclosure is to provide a state machine system and method for detecting the operating state of a sensor, which will allow for a safer prediction of potential operating state problems.

[0008] To solve this problem, a method for detecting the operating state of a sensor according to independent claim 1 is proposed. Additionally, a state machine system for performing the method for detecting the operating state of a sensor according to independent claim 12 is provided. Alternative embodiments are the subject of the dependent claims.

[0009] According to one aspect, a method for detecting the operating state of a sensor is provided. In a state machine, the method includes: receiving continuous monitoring data related to the operation of the sensor; providing a trained learning algorithm for detecting the operating state of the sensor indicative of the sensor function, wherein the learning algorithm is trained according to a training data set including historical data; detecting the operating state of the sensor by analyzing the continuous monitoring data with the trained learning algorithm; and providing output data indicative of the detected operating state of the sensor.

[0010] According to another aspect, a state machine system is provided. The state machine system has one or more processors configured for data processing and for performing a method for detecting the operating state of a sensor, the method including: receiving continuous monitoring data related to the operation of the sensor; providing a trained learning algorithm for detecting the operating state of the sensor indicative of the sensor function, wherein the learning algorithm is trained according to a training data set including historical data; detecting the operating state of the sensor by analyzing the continuous monitoring data with the trained learning algorithm; and providing output data indicative of the detected operating state of the sensor.

[0011] According to the proposed technique, a machine learning process is applied to detect the operating state of a sensor. Thereby, a predictive method for determining the operating state of a sensor by using a trained learning algorithm is achieved, the trained learning algorithm being trained according to a training data set and being applied to analyze continuous monitoring data related to the operation of the sensor.

[0012] For example, anomalies and / or faults regarding the operation of the sensor can be predicted, thereby avoiding potential problems in the operation of the sensor.

[0013] The learning algorithm is trained according to a training data set including historical data. As used in this application, the term "historical data" refers to data collected, detected, and / or measured prior to the process of determining the operating state. The historical data may have been detected or collected before starting to collect the continuous monitoring data received for operating state detection.

[0014] The training dataset can be collected, detected, and / or measured by the same sensor and / or a different sensor. A sensor different from the sensor for which the operating state is to be detected can be of the same sensor type.

[0015] The training dataset can include training data indicating the sensor state to be detected or predicted. For example, the training dataset can indicate one or more of the following: a manufacturing defect state, a fault state, a blood glucose indication state, and an anamnestic indication state.

[0016] Detection can include at least one of the following: detecting the manufacturing defect state of the sensor, which indicates a defect in the manufacturing process of the sensor; detecting the fault state of the sensor, which indicates a fault of the sensor; detecting the abnormal state of the sensor, which indicates an abnormality in the operation of the sensor; detecting the blood glucose indication state of the sensor, which indicates the blood glucose index of the patient for whom continuous monitoring data is provided; and detecting the anamnestic indication state of the sensor, which indicates the anamnestic patient state of the patient for whom continuous monitoring data is provided. Detection of the manufacturing defect state of the sensor can be performed after the sensor is manufactured. Alternatively or additionally, detection of the manufacturing defect state can be applied to an intermediate sensor product (non-final sensor) while the manufacturing process is still running. Similarly, detection of the fault state of the sensor can be part of or related to the manufacturing process. Alternatively, through the proposed technology, for example, in the case of applying the sensor for measurement, the fault state of the sensor can be predicted after the manufacturing process is completed. Detection of the abnormal state of the sensor can be performed during the measurement process, for example, in real time while the sensor is detecting the measurement signal. Similarly, one of the detection of the blood glucose indication state and the detection of the anamnestic indication state can be performed while the measurement process is running. Alternatively, such detection can be applied after the measurement process is completed.

[0017] For example, in response to detecting the blood glucose indication state of the sensor, the blood glucose index of the patient can be determined. The blood glucose index is a number associated with a specific type of food, which indicates the effect of the food on a person's blood glucose (also known as blood sugar) level. A value of one hundred can represent a standard, i.e., an equivalent amount of pure glucose. Additionally or alternatively, other blood glucose parameters can be determined, such parameters including the rate of change of the blood glucose level, acceleration, event patterns due to, for example, patient movement, meals, mechanical stress on the sensor regarding the anamnestic indication state of the sensor. Regarding the anamnestic indication state, potential anamnestic data, such as hba1c or demographic data, such as the age and / or gender of the patient, can be determined.

[0018] Providing a trained learning algorithm may include providing at least one learning algorithm selected from the group consisting of: K-nearest neighbor, support vector machine, naive Bayes, decision trees (such as random forests), logistic regression (such as polynomial logistic regression), neuronal network, decision trees, and Bayesian networks. One of naive Bayes, random forests, and polynomial logistic regression may be preferred. In a preferred embodiment, the random forest algorithm may be applied, for which the correlations and interactions between parameters are analyzed or automatically incorporated.

[0019] In this embodiment, a method includes training a learning algorithm based on a training data set including historical data.

[0020] The method may also include training a learning algorithm based on a training data set including historical data.

[0021] Training may include training a learning algorithm based on a training data set including at least one of in vivo historical training data and in vitro historical training data.

[0022] Training may include training a learning algorithm based on a training data set including continuous monitoring historical data.

[0023] Training may include training a learning algorithm based on a training data set including test data from the group consisting of: manufacturing test data, patient test data, personalized patient test data, population test data including multiple patient data sets. The training data set may be derived from one or more of such different test data to optimize the training data set for one or more operating states of the sensor.

[0024] Training may include training a learning algorithm based on a training data set including training data indicating one or more sensor-related parameters from the group consisting of: the current value of the sensor, particularly the current value of the working electrode in the case of a continuously monitored sensor; the voltage value of the sensor, particularly the voltage value of the counter electrode or the voltage value between the reference electrode and the working electrode in the case of a continuously monitored sensor; the temperature of the sensor environment during measurement; the sensitivity of the sensor; the offset of the sensor; and the calibration state of the sensor. One or more of the sensor-related parameters may be selected depending on the operating state to be detected. Regarding the calibration state of the sensor, for example, it may indicate when the last calibration was performed.

[0025] The one or more sensor-related parameters may include at least one of non-cross-correlated sensor-related parameters and cross-correlated sensor-related parameters. Two or more sensor-related parameters may be cross-correlated. In this case, the cross-correlated sensor-related parameters may be selected for detecting the operating state by taking into account all the cross-correlated sensor-related parameters. Differently, in the case of non-cross-correlated sensor-related parameters, a single one of the non-cross-correlated sensor-related parameters may be selected for detecting the operating state. The non-cross-correlated sensor-related parameters may independently allow the detection of the operating state.

[0026] The method may further include validating the trained learning algorithm according to a validation data set, the validation data set including measured continuous monitoring data and / or simulated continuous monitoring data, which indicate at least one of the following for the sensor: a manufacturing defect state, a failure state, a blood glucose indication state, and a memory indication state.

[0027] The method may further include at least one of the following: receiving continuous monitoring data including compressed monitoring data, and training a learning algorithm according to a training data set including compressed training data, wherein the compressed monitoring data and / or the compressed training data are determined by at least one of a linear regression method and a smoothing method. The compressed data may be the result of reducing the size of the monitoring data or the training data. Regarding the smoothing method, a kernel smoothing or spline smoothing model or time series analysis known per se may be applied. At different stages of compression, the monitoring data / training data may include data per second (measurement signal), data per minute, and / or statistical data including eigenvalue (such as sensor parameters, variance, noise, or rate of change).

[0028] The continuous monitoring data may be provided by a sensor that is a fully or partially implanted sensor for continuous glucose monitoring (CGM). Generally, in the context of CGM, an analyte value or level indicating the glucose value or level in the blood may be determined. The analyte value may be measured in interstitial fluid. The measurement may be performed subcutaneously or in vivo. CGM may be implemented as an almost real-time or quasi-continuous monitoring process, so as to frequently or automatically provide / update the analyte value without user intervention. In an alternative embodiment, the analyte may be measured by a biosensor in a contact lens through ocular fluid or by a biosensor on the skin in sweat via transdermal measurement. The CGM sensor may stay for several days to several weeks and then must be replaced.

[0029] Regarding the state machine system, the above alternative embodiments may be applied mutatis mutandis. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Additional embodiments will be described hereinafter with reference to the figures. Shown in the figures are: Figure 1 An embodiment of a state machine system is shown; Figure 2 A flowchart of an embodiment of a method for determining an operating state of a sensor is shown; Figure 3 An overview of data collection for a learning algorithm is shown; Figure 4 A graph of the current density measured at the working electrode of a sensor is shown; Figure 5 An error - free measurement is shown; Figure 6 A measurement showing a fluidics error is shown; Figure 7 A measurement showing a maxed out current error is shown; Figure 8 The degree of correlation between different parameters for use in a learning algorithm is shown; Figure 9 An illustration of using hyperparameters to adapt the model features of a random forest model is shown; Figure 10 An illustration of the prediction error of logistic regression is shown; Figure 11 A receiver operating characteristic curve for logistic regression is shown; Figure 12 An example of a tree for a random forest model is shown; Figure 13 An exemplary illustration of the error of a random forest is shown; Figure 14 A comparison of the accuracies of different exemplary learning algorithms is shown. Detailed Description

[0031] Figure 1 An embodiment of a state machine system 1 is shown, which may also be referred to as a state analysis system. The state machine system includes one or more processors 2, a memory 3, an input interface 4, and an output interface 5. In the illustrated embodiment, the input interface 4 and the output interface 5 are provided as separate modules. Alternatively, both the input interface 4 and the output interface 5 may be integrated in a single module.

[0032] In another embodiment, additional functional elements 7 may be provided in the state machine system 1.

[0033] Continuous monitoring data related to the operation of sensor 7 is received in one or more processors 2 via input interface 4. Sensor 7 may be connected to input interface 4 of state machine system 1 via a wire. Alternatively or additionally, a wireless connection such as Bluetooth, WiFi or other wireless technologies may be provided.

[0034] In the illustrated embodiment, sensor 7 includes a sensing element 8 and sensor electronics 9. In this embodiment, sensing element 8 and sensor electronics 9 are provided in the same housing of sensor 7. Alternatively, sensing element 8 and sensor electronics 9 may be provided separately and may be connected using a wire and / or wirelessly.

[0035] In one embodiment, continuous monitoring data may be provided by sensor 7, which is a fully or partially implanted sensor for continuous glucose monitoring (CGM). Generally, in the context of CGM, an analyte value or level indicating the glucose value or level in the blood may be determined. The analyte value may be measured in interstitial fluid. This measurement may be performed subcutaneously or in vivo. CGM may be implemented as an almost real-time or quasi-continuous monitoring process, thus providing / updating the analyte value frequently or automatically without user intervention. In an alternative embodiment, the analyte may be measured using a biosensor in a contact lens through ocular fluid or using a biosensor on the skin in sweat via transdermal measurement.

[0036] The CGM sensor may remain in place for several days to weeks and then must be replaced. A transmitter may be used to send information about the analyte value or level indicating the glucose level from the sensor to a receiver (such as sensor electronics 9 or input interface 4) via wireless and / or wired data transmission.

[0037] Output data indicating the detected operating state of sensor 7 is provided to one or more output devices 10 via output interface 5. It is contemplated that any suitable output device may be used as output device 10. For example, output device 10 may include a display device. Alternatively or additionally, output device 10 may include an alarm generator, a data network and / or one or more other processing devices. In another embodiment (not shown), more than one output device 10 is provided.

[0038] One or more output devices 10 may be connected to output interface 5 of state machine system 1 via a wire. Alternatively or additionally, a wireless connection such as Bluetooth, WiFi or other wireless technologies may be provided.

[0039] In an alternative embodiment, output device 10 or one of the more than one output devices 10 is integrated in state machine system 1.

[0040] In an embodiment, one or more additional input devices 11 are connected to the input interface 4. Such additional input devices 11 may include one or more additional sensors to collect training data and / or validation data for use by a learning algorithm. Additionally or alternatively, the additional input devices 11 may also include sensors for obtaining different data types. An example of such different data types is temperature data. The sensor data of such different data types may be additionally analyzed for detecting the operating state of the sensor 7. Additionally or alternatively, the sensor data of such different data types may be used as training data and / or validation data. Alternatively or additionally, one or more additional input devices 11 may include a data network, an external data storage device, user input devices such as a keyboard, a mouse, one or more additional processing devices, and / or any other device suitable for providing relevant data to the state machine system 1.

[0041] Figure 2 is a flowchart illustrating an embodiment of a method for detecting the operating state of a sensor.

[0042] In step 20, continuous monitoring data related to the operation of the sensor 6 is received at the input interface 4 of the state machine system 1.

[0043] The continuous monitoring data may indicate one or more sensor-related parameters. Such sensor-related parameters may include the current value of the working electrode of the sensor, the voltage value of the counter electrode of the sensor, the voltage value between the reference electrode and the working electrode, the temperature of the sensor environment during measurement, the sensitivity of the sensor, the offset, and / or the calibration state of the sensor. The sensor-related parameters may include non-cross-correlated sensor-related parameters, cross-correlated sensor parameters, or a combination thereof.

[0044] In one embodiment, the continuous monitoring data may include compressed monitoring data. In this case, the compressed monitoring data is determined by at least one of a linear regression method and a smoothing method.

[0045] In step 21, a trained learning algorithm is provided. The learning algorithm is trained according to a training data set including historical data. The trained learning algorithm may be provided in the memory 3 of the state machine system 1. Alternatively, the trained learning algorithm may be provided in one or more processors 2 from the memory 3. In an alternative embodiment, the trained learning algorithm is provided via the input interface 4. For example, the trained learning algorithm may be received from an external storage device. In another embodiment, the trained learning algorithm may be provided in one or more additional functional elements 7, or the trained learning algorithm may be provided in one or more processors 2 from one or more additional functional elements 7.

[0046] In different embodiments, the order of steps 20 and 21 can be reversed. In a particular embodiment, a trained learning algorithm is provided as long as sensor 7 is put into operation. As another alternative, steps 20 and 21 can be performed wholly or partially simultaneously.

[0047] In step 22, using one or more processors 2, the continuous monitoring data is analyzed with the trained learning algorithm. In embodiments where the trained learning algorithm is not provided in processor 2, processor 2 can access the trained learning algorithm to analyze the continuous monitoring data. By analyzing the continuous monitoring data, the operating state of sensor 7 is detected.

[0048] The operating state detected for the sensor in step 22 can be one of several different states. For example, a manufacturing defect state of the sensor indicating a defect in the manufacturing process of the sensor, a fault state of the sensor indicating a fault of the sensor, an abnormal state of the sensor indicating an anomaly in the operation of the sensor, a blood glucose indication state of the sensor indicating the blood glucose index of the patient for which the continuous monitoring data is provided, and / or a memory indication state of the sensor indicating the memory patient state of the patient for which the continuous monitoring data is provided can be detected.

[0049] Next, in step 23, output data indicating the detected operating state of the sensor is provided at output interface 5.

[0050] In an embodiment, the method for detecting the operating state of a sensor can further include training a learning algorithm based on a training data set including historical data.

[0051] Still referring to Figure 2 , in step 24, a training data set including historical data is provided.

[0052] The historical training data can include in vivo historical training data, which indicates sensor-related parameters obtained while sensor 7 is operating on a living subject. Alternatively or additionally, the historical training data can include in vitro historical training data, which indicates sensor-related parameters obtained while sensor 7 is not operating on a living subject.

[0053] The training data set provided in step 24 can include continuous monitoring historical data.

[0054] The training data set can include manufacturing test data, patient test data, personalized patient test data, and / or population test data including multiple patient data sets.

[0055] Training data can indicate one or more sensor-related parameters. Such sensor-related parameters can include the current value of the working electrode of the sensor, the voltage value of the counter electrode of the sensor, the voltage value between the reference electrode and the working electrode, the temperature of the sensor environment during measurement, the sensitivity of the sensor, the offset, and / or the calibration state of the sensor. The sensor-related parameters can include non-cross-correlated sensor-related parameters, cross-correlated sensor parameters, or a combination thereof.

[0056] In one embodiment, the training data set can include compressed training data. In this case, the compressed training data is determined by at least one of a linear regression method and a smoothing method.

[0057] In step 25, a learning algorithm is trained according to the training data set provided in step 24.

[0058] The learning algorithm can be selected from suitable algorithms. Such learning algorithms include: K-Nearest Neighbor, Support Vector Machine, Naive Bayes, Decision Tree (such as Random Forest), Logistic Regression (such as Polynomial Logistic Regression), Neural Network, Decision Tree, and Bayesian Network. The learning algorithm can be selected based on whether it is suitable for use with the continuous monitoring data analyzed in step 22.

[0059] The training of the learning algorithm in step 25 can be performed in the state machine system 1. In this case, in step 24, the training data set can be provided in the memory 3 of the state machine system 1. Alternatively, the training data set can be provided in one or more processors 2 from the memory 3. In an alternative embodiment, the training data set is provided via the input interface 4. For example, the training data set can be received from an external storage device. In a further embodiment, the training data set can be provided in one or more additional functional elements 7, or can be provided in one or more processors 2 and / or the memory 3 from one or more additional functional elements 7.

[0060] In an alternative embodiment, the training of the learning algorithm in step 25 can be performed outside the state machine system 1. In this embodiment, in step 24, the training data set is provided in any suitable manner that enables the training of the learning algorithm.

[0061] Another embodiment can include step 26, in which the trained learning algorithm is verified according to a validation data set. The validation data set includes measured continuous monitoring data and / or simulated continuous monitoring data. This data indicates at least one of the following for the sensor: manufacturing defect state, fault state, blood glucose indication state, and memory indication state.

[0062] The verification of the trained learning algorithm in step 26 can be performed in the state machine system 1. In this case, a verification data set can be provided in the memory 3 of the state machine system 1. Alternatively, the verification data set can be provided in one or more processors 2 from the memory 3. In an alternative embodiment, the verification data set is provided via the input interface 4. For example, the verification data set can be received from an external storage device. In a further embodiment, the verification data set can be provided in one or more additional functional elements 7, or can be provided in one or more processors 2 and / or the memory 3 from one or more additional functional elements 7.

[0063] In an alternative embodiment, the verification of the trained learning algorithm in step 26 can be performed outside the state machine system 1. In this embodiment, the verification data set is provided in any suitable manner that enables the verification of the learning algorithm.

[0064] In one embodiment, the verification data set can include compressed verification data. In this case, the compressed verification data is determined by at least one of a linear regression method and a smoothing method.

[0065] Additional aspects are described next.

[0066] The measurements for collecting continuous monitoring data are performed using a plurality of continuous glucose monitoring sensors.

[0067] Based on an established sequence of work steps in the field of data mining (refer to Shmueli et al., Data Mining for Business analytics – Concepts, Techniques, and Applications with XLMiner, 3rd Edition, New York: John Wiley & Sons, 2016), which will be used to provide support for model development, all or part of the following steps can be implemented: 1. Draft the problem 2. Obtain data 3. Analyze and clean the data 4. Reduce the dimension, if necessary 5. Specify the problem (classification, clustering, prediction) 6. Share data during training. Verification and test data sets. 7. Select data mining techniques (regression, neural networks, etc.) 8. Different versions of the algorithm (different variables) 9. Interpret the results 10. Integrate the model into the existing system.

[0068] The process for data collection, which can be applied in alternative embodiments, is described next.

[0069] At the test site, the current value of the working electrode of the sensor, the voltage value of the counter electrode of the sensor, and the voltage value between the reference electrode and the working electrode can be recorded once per second for each channel. The temperature of the solution in which the sensor is located can be detected once per minute. These parameters can be stored in an Extensible Markup Language (XML) file. Then, the data processing program CoMo captures this XML file and provides it as a so-called experiment in the form of a SAS dataset. At the lowest stage, this experiment consists of data for one second of reference. As Figure 3 shown, this data is compressed to minute values by means of CoMo. In this step, descriptive statistics are additionally generated, such as the minimum, average, and maximum values per minute. Then, compression to step values is performed. As Figure 4 shown, pyramid-shaped steps can be observed. The last compression stage, i.e., the basic statistics, corresponds to the reporting of characteristic values for each sensor.

[0070] First, the data from the highest compression stage, i.e., the basic statistics, can be used, since access to more complex data can be reserved for cases where the classification using simpler data provides insufficient results. In addition, the classification of time-resolved data (as occurred in the first and second stages) would require different programming languages, such as Python.

[0071] In one example, multiple test sequences, such as 16 test sequences, were identified in 256 data entries, which were distributed to the test sites, and the results were multiplied by the number of channels.

[0072] For error identification for each sensor, the graph of the current intensity at the working electrode per minute for each channel according to Figure 4 is considered. For a 7-day measurement, each day is represented by a separate curve. Since the sensor requires a one-day preparation in the form of pre-expansion to operate, only six days are illustrated. From Figure 4 it can be clearly seen that on the third day, channel 4 is significantly different from the other days and thus no longer follows the typical pyramid shape. Therefore, channel 4 is identified as problematic.

[0073] Once all channels have been analyzed and identified, the test sequences can be exported from SAS to memory. In the last step, the test sequences can be read from this memory in R and stored as a reference.

[0074] The entire dataset is divided into three parts: a test dataset representing continuous monitoring data, a training dataset, and a validation dataset.

[0075] In an alternative embodiment, two types of errors representing the operating state of the sensor are to be identified by the model. These two types of errors are fluidic errors and maximum current errors. As Figure 5 shown, channels without errors can initially be regarded as references. As Figure 4 shown, a pyramid shape can be observed. However, these days they are not graphically superimposed but arranged continuously. Since the decision as to whether a channel is identified as problematic is made by means of the current intensity, the current intensity is also used for the analysis regarding the individual errors.

[0076] In this embodiment, fluidic errors are the focus of error detection. Therefore, data from a period of time with a large number of such defects are selected. One difficulty associated with this type of error is the wide variety of manifestations in which it can occur. However, as Figure 6 shown, it can be observed that the measured values tend to decrease. The cause of this error lies in the test site unit, so this defect can also be called a test site error. It is speculated that the cause is bubbles in the test system, which may be caused by, for example, temperature fluctuations. Bubbles in the liquid may form due to an inflow interruption.

[0077] When the sensor is inserted into the channel at the start of the test, a maximum current error may occur. When a current higher than the threshold is detected, the sensor at the test site is marked with this type of error. Now, the staff at the test site can re-insert the sensor into the channel to handle this error. Alternatively, the sensor may ultimately be marked as problematic. Figure 7 A typical maximum current error is shown. Compared with Figure 6 it, significantly higher current values can be identified at the start of the measurement.

[0078] In order to be able to label the data in a meaningful way, different error codes can be provided for the individual errors according to Table 1.

[0079] Table 1:

[0080] Parameter Analysis In an alternative embodiment, the strength of the linear relationship between variables can be determined by means of a correlation coefficient, the value of which can be between -1 and 1. In the case of a value of 1, there is a high positive linear correlation. When looking Figure 8 at it, it can be seen that parameter S360 is correlated with a large number of other parameters.

[0081] As described above, variables such as current can exist, which can be measured directly at the test site. In an embodiment, when compressing data, a linear model and a spline model are used, which estimate various parameters. Since the data set to be used later includes the compressed data, a combined model is considered.

[0082] Measured Value For descriptive statistics regarding measured values representing sensor-related parameters, an analysis of the normal distribution conditions that can be graphically achieved by means of a quantile-quantile plot (QQ plot) according to DIN 53804-1 may be of interest. The X-axis of the QQ plot is defined by the theoretical quantiles, while the Y-axis is defined by the empirical quantiles. The parameters of the normal distribution result in a straight line, which is depicted as a straight line in the QQ plot. Additionally, there are various normal distribution tests, such as the chi-square test or the Shapiro-Wilk test. These hypothesis tests define the null hypothesis as the existence of a normal distribution, and conversely, the alternative hypothesis assumes the non-existence of a normal distribution. These test methods are highly sensitive to deviations. In an embodiment, therefore, the normal distribution can be analyzed for each parameter by means of a QQ plot.

[0083] The measured values can include the sensor current for different glucose concentrations. These can be determined as the median value over certain time periods and can additionally or alternatively be averaged. The measured values can also include the sensitivity of the sensor. Additionally or alternatively, the measured values can include parameters characterizing a graph depicting the measured values, such as the sensor current. These can include, for example, drift and / or curvature. Additionally or alternatively, the values can include statistical values regarding other measured values. Different models such as linear models and / or spline models can be employed to approximate the measured values. The measured values and parameters can be determined in their entirety or any of them at different glucose concentrations and / or for different time periods.

[0084] Learning Algorithm In replacement embodiments, several modeling methods for the learning algorithm were selected (see, e.g., AFew Useful Things to Know About Machine Learning by Domingos, Commun. ACM 55.10, pp. 78 - 87. DOI: 10.1145 / 2347736.2347755, 2012), and their advantages and disadvantages were analyzed. Additionally, the methods can be analyzed in terms of their compatibility with the problem in order to enable method selection. Exemplary methods are described next (Encyclopedia of Machine Learning by Sammut et al., 1st edition, Springer-Verlag GmbH, 2011). Table 2 summarizes the advantages and disadvantages of these methods.

[0085] K-Nearest Neighbor The goal of this method is to classify an object into the category to which similar objects in the training volume have been classified, and thus output the category with the highest frequency as the result. To determine the proximity of the object, similarity measures such as, for example, the Euclidean distance are used. This method is very suitable for significantly larger data volumes, which are not present in the present example. This is also the reason why this model was not included in the control consideration.

[0086] Support Vector Machine In this method, a hyperplane is calculated that classifies the objects into categories. To calculate the hyperplane, the distance around the category boundary is maximized, so the support vector machine is one of the "large margin classifiers". An important assumption of this method is the linear separability of the data; however, it can be extended to higher-dimensional vector spaces by means of the kernel trick. For classification with less overfitting, a large data volume is required, which is not present in some embodiments.

[0087] Naive Bayes The naive assumption is that the current variables are statistically independent of each other. In most cases, this assumption is incorrect. Nevertheless, Naive Bayes gives good results in many cases, such that a relatively high correct classification rate is obtained even with a small correlation between the attributes. The characteristic of Naive Bayes lies in its simple operation mode, so it can be used for model selection.

[0088] Logistic Regression Regarding logistic regression, the analytical calculation possibility for how much the characteristics of the dependent variable can be attributed to the values of the independent variables is calculated.

[0089] Neural Network An artificial neural network is based on the biological structure of neurons in the brain. A simple neural network consists of neurons arranged in three layers. These layers are the input layer, the hidden layer, and the output layer. Between the layers, all neurons are interconnected by weights, and the weights are gradually optimized during the training phase. Currently, neural networks are widely used in many fields, so there are a large number of model variations. There are multiple hyperparameters that must be determined from empirical values to optimize such a network. In some embodiments, these hyperparameters are not determined for reasons of time efficiency.

[0090] Decision tree A decision tree is a sorted hierarchical tree, characterized by its simple and understandable appearance. Nodes closer to the root are more significant for classification than nodes closer to the leaves. In one embodiment, due to the fact that decision trees often encounter problems caused by overfitting, the random forest method is selected for model selection. This method consists of multiple decision trees, whereby each tree represents a partial number of variables.

[0091] Bayesian network A Bayesian network is a directed graph that illustrates a multivariate likelihood distribution. The nodes of the network correspond to random variables, and the edges show the relationships between them. A possible application could be to explain the causes of disease symptoms in diagnosis. To develop a Bayesian network, it is necessary to be able to describe the dependencies between variables as detailed as possible. Due to the errors handled in some embodiments, the generation of such a graph is not feasible.

[0092] Table 2:

[0093] Method Selection In an alternative embodiment, the models are initially considered theoretically and analyzed in terms of their assumptions. Next, a first implementation is carried out, and then it can be optimized by means of various methods.

[0094] In the first step, a binary problem with a linear model can be used, which includes a total of three variables. Subsequently, based on the actual problem, all classes and parameters can be used to train the learning algorithm represented by the model. Finally, the model characteristics can be adapted to the available data by means of hyperparameters such as the number of decision trees in the case of a random forest. Figure 9 The illustration of this process is exemplified by the random forest model. The abbreviation ACC represents accuracy, which decreases with the first adaptation but then increases again with the optimization steps by means of cross-validation.

[0095] Naive Bayes: The model that can be used in the embodiments is based on Bayes' theorem and can be used as a simple and fast method for classifying data. In such an embodiment, the conditions are that the data present are statistically independent of each other and are in a normal distribution. Since this method can determine the relative frequencies of the data in only a single pass, it is considered a simple and fast method.

[0096] According to Bayes' theorem, the following formula is used to calculate the conditional likelihood:

[0097] When it is assumed that the attributes exist independently of each other, the Naive Bayes classifier can be defined as follows:

[0098] This function always predicts the most likely class y for the attribute xi by means of the maximum a posteriori rule. The maximum a posteriori rule behaves similarly to the maximum likelihood method but with the knowledge of the prior term. When there is metric data in the dataset, a distribution function is needed to calculate the conditional likelihood for P(xi|y). In an embodiment, Naive Bayes may also resort to the normal distribution (Guide to Intelligent Data Analysis: How to Intelligently Make Sense of Real Data by Berthold et al., 1st Edition, Springer-Verlag GmbH, 2010). Although there is no normal distribution in the case of many CGM variables, Naive Bayes can still be used because it can achieve a high correct classification rate despite a slight deviation from the normal distribution.

[0099] P(X j |y) = N(x i , μ, σ 2 ), The mean μ and variance σ are calculated for each attribute xi and each class y.

[0100] Since a relatively small dataset is sufficient for good predictions in the case of this model, in one embodiment only four measurements can be used as input. In one embodiment, as a first consideration, a subset of the available parameters can be selected, including A2, I90, and D.

[0101] Under the condition that I90 appears in a class, Naive Bayes can be used to determine the probability of an error.

[0102] In one embodiment, no statement should be made regarding the error type. Thus, without the need for new data identification, four test sites that only include fluidics errors can be selected. In this case, generally, error code 0 can be identified as no error, and error code 1 can be identified as an error. Table 3 illustrates an excerpt of the input data set for one embodiment of Naive Bayes.

[0103] Table 3:

[0104] As shown in Table 4, the model output can include prior values calculated for the categories. In the next step, the mean value and standard deviation of each variable for category 0 (no error) and category 1 (error) can be calculated. They can be used to determine the distribution function of the variables based on the normal distribution.

[0105] Table 4:

[0106] The quality of the model can be evaluated with the aid of various parameters of the output. As shown in Table 5, in one embodiment, accuracy, sensitivity, and specificity may be of primary significance from this output.

[0107] Table 5:

[0108] In one embodiment, accuracy allows for a first impression of the model results and can thus be used to evaluate the quality.

[0109] In certain embodiments, to be able to evaluate the importance of accuracy, the Kappa value can be used. The Kappa value is a statistical measure of the correspondence of two quality parameters - in this embodiment, the observed accuracy and the expected accuracy. After the observed accuracy and the expected accuracy have been calculated, the Kappa value can be determined as follows:

[0110] There are different methods to interpret the Kappa value. Table 6 summarizes one such method, which can be learned from "The Measurement of Observer Agreement for Categorical Data" by Landis et al., Biometrics 33, pp. 159–174, 1977: Table 6:

[0111] In one embodiment, a positive predictive value, a negative predictive value, a sensitivity, and a specificity can be determined.

[0112] The positive predictive value specifies the percentage of values that have been correctly classified as problematic among all results that have been classified as problematic (corresponding to the second row of a four-field table).

[0113] Accordingly, the negative predictive value specifies the percentage of values that have been correctly classified as error-free among all results that have been classified as error-free (corresponding to the second row of a four-field table).

[0114] The sensitivity specifies the percentage of objects that have been correctly classified as positive in actual positive measurements.

[0115] The specificity specifies the percentage of objects that have been correctly classified as negative in measurements that are indeed negative.

[0116] In an embodiment, a binary model with variables A2, D, and I90 and the predictions of an overall model can be illustrated by a four-field table. In the embodiment shown in Table 7, the binary model has the greatest difficulty in the region of the false negative rate, which is reflected in a sensitivity of ≈0.7857.

[0117] Table 7:

[0118] In an alternative embodiment, after discussing Naive Bayes in the context of a binary problem, then all error types and variables can be highlighted in a second stage. This implementation can be based on all available data. If the behavior of the accuracy and the Kappa value is similar in these two model versions, then the argument that Naive Bayes with less data can already yield good results can be strengthened.

[0119] Logistic Regression Logistic regression can be implemented as is known per se (Multivariate Analysemethoden: Eine anwendungsorientierte Einführung by Backhaus et al. (Multivariate Analysis Methods: An Application-Oriented Introduction), Springer, Berlin Heidelberg, 2015). Logistic regression can be used to determine the connection between the performance of independent variables and dependent variables. Generally, the binary dependent variable Y is coded as 0 or 1, i.e., 1 for the presence of an error and 0 for the absence of an error. A possible application of logistic regression in the CGM context is to determine whether current values, splines, and sensitivities are related to the performance of an error.

[0120] In an embodiment, a generalized linear model can be used to implement logistic regression (see, for example, Dobson's An Introduction to Generalized Linear Models, 2nd Edition, a statistical science textbook by Chapman and Hall / CRC, Taylor and Francis, 2010). This may be advantageous since linear models are easy to interpret.

[0121] Evaluation: Table 8 shows a comparison of a simplified model using variables I90, A2, and D for one embodiment with a model using all variables. In this embodiment, the accuracy of the model using all variables is approximately 7% higher than the accuracy of the simplified model, indicating that the simplified model does not use variables that are closely related to classification.

[0122] Table 8:

[0123] "Backward elimination" (Sheather's A Modern Approach to Regression with R, Springer Science & Business Media, 2009) and the Akaike Information Criterion (Aho Ket al.'s Model selection for ecologists: the worldviews of AIC and BIC, Ecology, 95:631–636, 2014) can be used to identify relevant parameters. These parameters can be examined with respect to the prediction error of the logistic regression. Figure 10 Shows the distribution density of variables and the location of mispredicted values for one embodiment. Since there are mispredicted values at the edges of the distribution and in the measurement regions without error, it is not possible in this embodiment to correctly predict all problematic measurements with simple association rules.

[0124] In an embodiment, a Receiver Operating Characteristic curve (ROC) can be used to determine sensitivity and specificity. In this case, the ideal curve rises vertically at the start, meaning an error rate of 0%, while the false positive rate only rises later. A curve along the diagonal implies a random process. Figure 11 Shows the ROC for logistic regression for an exemplary embodiment.

[0125] Polynomial Logistic Regression: In polynomial logistic regression, the dependent variable X may have more than two different values, making binary logistic regression a special case of polynomial logistic regression.

[0126] Random Forest Random forests follow the bagging principle, which states that a combination of multiple classification methods can improve the classification accuracy by training several classifiers with different samples of the data. In an embodiment, a random forest algorithm known per se can be used (Random Forests by Breiman, Mach. Learn. 45.1, S. 5–32. DOI: 10.1023 / A:1010933404324, 2001).

[0127] In such an embodiment, when a new element is fed to the decision tree, each tree determines a class as the result. In the next step, the result class is determined based on the class proposed by the majority of the trees. Figure 12 A tree of an exemplary embodiment is shown.

[0128] The random forest can be optimized using, for example, the number of trees and / or the number of nodes in the trees. In Figure 13 it, an example of the error of a random forest for an embodiment is shown, where the error probability regarding the maximum current error oscillates between 50% and 100%. In this example, all "other errors" are misclassified, as can be seen from the top line. This may be due to the low occurrence of the maximum current error and other errors.

[0129] Figure 14 A comparison of the accuracies of exemplary learning algorithms of alternative embodiments is shown: polynomial logistic regression, naive Bayes, and random forest. The confidence intervals of the accuracies are shown on the left. The Kappa values of each model are shown on the right.

[0130] For this embodiment, the Kappa value allows the assumption of a trend according to which the accuracy of the polynomial logistic regression is less significant compared to the other models.

[0131] This assumption is confirmed by the prediction of the trained model on the test dataset of this embodiment, which is illustrated in the four-field table summarized in Table 9. Measurements of the test dataset are randomly selected to simulate actual data input. Although the maximum current error does not occur in the test dataset, the polynomial logistic regression incorrectly predicts this error type. However, the biggest problem with this model is the fluidics error, and none of the cases correctly classifies this error.

[0132] Table 9:

[0133] Thus, for this embodiment, the polynomial logistic regression corresponds to an accuracy of 66%, thus being lower than the Naive Bayes with 80% correctly classified cases and the Random Forest with 88% correctly classified cases. The first possible reason for such a situation may be the correlation between the parameters, which may lead to estimation distortion and an increase in the standard deviation. However, the Naive Bayes also requires the parameters to be uncorrelated, and for the shown embodiment, this model yields significantly better results. The reason may be that the Naive Bayes can already achieve high accuracy with a very small amount of data. By training the model with a higher amount of data, regardless of the correlation of the parameters, the accuracy of the Naive Bayes will increase significantly. However, the second assumption of the polynomial logistic regression, i.e., the "independence of irrelevant alternatives", may also be violated. This assumption specifies that the odds ratio of two error types is independent of all other response categories. For example, it can be assumed that the choice of the outcome categories "fluidics error" or "no error" is not affected by the presence of "other errors".

[0134] In the embodiment, the Random Forest provides the highest ratio of 86% correctly classified cases, whereby, even in the presence of a fluidics error, multiple misclassified cases are predicted as "no error". The reason why the Random Forest represents the most successful model regarding the prediction in this embodiment may be that, on the one hand, the tree structure enables the arrangement of the parameters regarding their interaction. On the other hand, due to the number of trees, the Random Forest can be optimized with little effort compared to the polynomial logistic regression and the Naive Bayes in R. This can be achieved with the help of a graph of the error versus the number of decision trees, which shows the number of decision trees at which the error converges.

[0135] As an alternative to the compressed data, uncompressed data can be used. For data showing time resolution, a neural network (such as a recurrent network) can be used to achieve the prediction. The advantage of the recurrent neural network is that no assumptions need to be made before creating the model.

Claims

1. A method for detecting an operating state of a continuous glucose measurement sensor, the method comprising in a state machine - receiving continuous monitoring data related to the operation of the sensor and including compressed monitoring data; - providing a trained learning algorithm for detecting the operating state of the sensor indicative of sensor functionality, wherein the learning algorithm is trained according to a training data set including historical data and compressed training data; - detecting the operating state of the sensor by analyzing the continuous monitoring data with the trained learning algorithm; - providing output data indicative of the detected operating state of the sensor; and - sending the output data to one or more output devices, wherein the one or more output devices include an alarm generator for generating an alarm in response to the output data, wherein - the historical data consists of data collected, detected, and measured at least one of before the operating state is detected; and - the compressed monitoring data and the compressed training data are determined by at least one of a linear regression method and a smoothing method and are the result of reducing the size of the monitoring data and the training data respectively, wherein at different stages of compression, the monitoring data and the training data respectively include data per second, data per minute, and / or statistical data including eigenvalue, and wherein the eigenvalue is a sensor parameter, variance, noise, and / or rate of change.

2. The method according to claim 1, wherein, the detection includes at least one of the following - detecting a manufacturing defect state of the sensor, which indicates a defect in the manufacturing process of the sensor; - detecting a failure state of the sensor, which indicates a failure of the sensor; - detecting an abnormal state of the sensor, which indicates an abnormality in sensor operation; - detecting a blood glucose indication state of the sensor, which indicates the blood glucose index of the patient for whom the continuous monitoring data is provided; and - detecting a memory indication state of the sensor, which indicates the memory patient state of the patient for whom the continuous monitoring data is provided.

3. The method according to claim 1, wherein, the detection includes at least one of the following - detecting a manufacturing defect state of the sensor, which indicates a defect in the manufacturing process of the sensor; - detecting a failure state of the sensor, which indicates a failure of the sensor; - detecting an abnormal state of the sensor, which indicates an abnormality in sensor operation; and - detecting a memory indication state of the sensor, which indicates the memory patient state of the patient for whom the continuous monitoring data is provided.

4. The method according to any one of claims 1 to 3, wherein, providing a trained learning algorithm includes providing at least one learning algorithm selected from the group consisting of - K-Nearest Neighbor; - Support Vector Machine; - Naive Bayes; - Logistic Regression; - Neural Network; - Decision Tree; and - Bayesian Network.

5. The method according to claim 4, wherein, the decision tree is a random forest.

6. The method according to claim 4, wherein, the logistic regression is polynomial logistic regression.

7. The method according to any one of claims 1 to 6 further comprises training a learning algorithm according to a training data set including historical data.

8. The method according to claim 7, wherein, the training comprises training the learning algorithm according to a training data set, the training data set including at least one of in vivo historical training data and in vitro historical training data.

9. The method according to claim 7 or 8, wherein, the training comprises training the learning algorithm according to a training data set including continuous monitoring historical data.

10. The method according to any one of claims 7 to 9, wherein, the training comprises training the learning algorithm according to a training data set including test data from the following groups: manufacturing test data, patient test data, personalized patient test data, population test data including a plurality of patient data sets.

11. The method according to any one of claims 7 to 10, wherein, the training comprises training the learning algorithm according to a training data set, the training data set including training data indicating one or more sensor-related parameters from the following groups: the current value of the sensor; the voltage value of the sensor, or the voltage value between the reference electrode and the working electrode; the temperature of the sensor environment during measurement; the sensitivity of the sensor; the offset of the sensor; and the calibration state of the sensor.

12. The method according to claim 11, wherein, the current value of the sensor is the current value of the working electrode of the sensor.

13. The method according to claim 11 or 12, wherein, the one or more sensor-related parameters include at least one of the following - non-cross-correlated sensor-related parameters; and - cross-correlated sensor-related parameters.

14. The method according to any one of claims 1 to 13 as long as claim 2 is cited further comprises validating the trained learning algorithm according to a validation data set, the validation data set including measured continuous monitoring data and / or simulated continuous monitoring data, which indicate at least one of the following for the sensor: manufacturing defect state, failure state, blood glucose indication state, and memory indication state.

15. The method according to any one of claims 1 to 13 as long as claim 2 is cited further comprises validating the trained learning algorithm according to a validation data set, the validation data set including measured continuous monitoring data and / or simulated continuous monitoring data, which indicate at least one of the following for the sensor: manufacturing defect state, failure state, and memory indication state.

16. The method according to any one of claims 1 to 15 further comprises - validating the trained learning algorithm according to a validation data set including compressed validation data; wherein, the compressed validation data is determined by at least one of a linear regression method and a smoothing method.

17. The method according to any one of claims 1 to 16, wherein the operating state includes fluidics error and / or maximum current error.

18. A state machine system having one or more processors configured to process data and execute a method for detecting an operating state of a continuous glucose monitoring sensor, the method comprising - receiving continuous monitoring data related to the operation of the sensor and including compressed monitoring data; - providing a trained learning algorithm for detecting the operating state of the sensor indicative of sensor functionality, wherein the learning algorithm is trained according to a training data set including historical data and compressed training data; - detecting the operating state of the sensor by analyzing the continuous monitoring data with the trained learning algorithm; - providing output data indicative of the detected operating state of the sensor; and - sending the output data to one or more output devices, wherein the one or more output devices include an alarm generator for generating an alarm in response to the output data, wherein - the historical data consists of data collected, detected, and measured at least one of before the operating state is detected; and - the compressed monitoring data and the compressed training data are determined by at least one of a linear regression method and a smoothing method and are the result of reducing the size of the monitoring data and the training data respectively, wherein at different stages of compression, the monitoring data and the training data respectively include data per second, data per minute, and / or statistical data including eigenvalue, wherein the eigenvalue is a sensor parameter, variance, noise, and / or rate of change.

Images