HK1177127B

HK1177127B - Methods and systems for processing glucose data measured from a person having diabetes

Info

Publication number: HK1177127B
Application number: HK13104531.3A
Authority: HK
Inventors: David L. Duke; Abhishek S. Soni; Stefan Weinert
Original assignee: F. Hoffmann-La Roche Ag
Priority date: 2010-01-26
Filing date: 2011-01-24
Publication date: 2015-12-11

Description

Method and system for processing glucose data measured from a person with diabetes

Technical Field

The present invention relates generally to processing glucose data measured from a person with diabetes, and in particular to estimating an actual glucose level of the person in the presence of glucose sensor noise and/or glucose sensor failure.

Background

By way of background, people suffer from type I or type II diabetes in which the sugar level in the blood is not properly regulated by the body. Many of these people can use Continuous Glucose Monitoring (CGM) to monitor their glucose levels without interruption. To perform CGM, a glucose sensor may be placed under the skin, which is capable of measuring the glucose level of a person in interstitial fluid. The glucose sensor may periodically measure the person's glucose level at known time intervals, such as every minute, and communicate the results of the glucose measurement to an insulin pump, blood glucose meter, smart phone, or other electronic monitor.

In some cases, the measured glucose results (from the glucose sensor) may contain sensor "noise" that causes it to deviate from the actual glucose level of the person. Sensor noise may be caused, for example, by physical movement of the glucose sensor relative to the skin or due to electrical noise that may be inherent to the sensor itself. Furthermore, the glucose sensor may occasionally fail, such that the measured glucose results (from the glucose sensor) may differ substantially from the actual glucose level of the person. Glucose sensors may fail in this manner due to, for example, failure of the sensor electronics or battery, or due to sensor "loss of signal". Sensor signal loss may occur due to physiological problems of attachment of the glucose sensor to the person, such as movement of the sensor relative to the person. Sensor signal loss may cause the measured glucose result to "drop" to near zero, although the actual glucose level of a person may be much higher.

As a result, embodiments of the present disclosure may process measured glucose results from a person so that the actual glucose level of the person may be estimated even in the presence of sensor noise and/or sensor failure. Additionally, based on the estimated glucose level, a future glucose level of the person may be predicted.

Disclosure of Invention

In one embodiment, a method for estimating glucose levels of a person with diabetes comprises: receiving a plurality of measured glucose results into a computing device from a glucose sensor coupled to a person; configured to determine based on the plurality of measured glucose resultsDetermining the probability P of glucose sensor accuracy_AUsing a computing device to analyze the plurality of measured glucose results; and based on the probability P of being accurate with a glucose sensor_AWeighting the plurality of measured glucose results, estimating, using the computing device, a glucose level of the person with a recursive filter configured to estimate the glucose level.

In another embodiment, a computer-readable medium having computer-executable instructions for performing a method for estimating glucose levels of a person with diabetes is disclosed, wherein the method comprises: receiving a plurality of measured glucose results into a computer from a glucose sensor coupled to a person; based on the plurality of measured glucose results, with a probability P configured to determine that the glucose sensor is accurate_AUsing a computer to analyze the plurality of measured glucose results; and based on the probability P of being accurate with a glucose sensor_AWeighting the plurality of measured glucose results, estimating, using a computer, a glucose level of the person with a recursive filter configured to estimate the glucose level.

In another embodiment, an apparatus for estimating glucose levels of a person with diabetes comprises: microcontroller, input device and display, wherein: the microcontroller is electrically coupled to an input device configured to receive a plurality of measured glucose results from a glucose sensor coupled to a person, wherein the microcontroller is configured to receive the plurality of measured glucose results from the input device; the microcontroller is configured to determine a probability P of glucose sensor accuracy based on the plurality of measured glucose results_ATo analyze the plurality of measured glucose results; the microcontroller is configured to determine the accuracy of the glucose sensor based on the probability P_AWeighting the plurality of measured glucose results to estimate a glucose level of the person with a recursive filter configured to estimate the glucose level; and the microcontroller is electrically coupled to the displaySuch that the microcontroller is configured to transmit information regarding the estimate of the person's glucose level to the display.

Drawings

The embodiments set forth in the drawings are illustrative and exemplary in nature and are not intended to limit the invention defined by the claims. The following detailed description of the illustrative embodiments can be understood when read in conjunction with the following drawings, where like structure is indicated with like reference numerals and in which:

FIG. 1 depicts a Continuous Glucose Monitoring (CGM) system according to one or more embodiments shown and described herein;

FIG. 2 depicts a glucose monitor according to one or more embodiments shown and described herein;

FIG. 3 depicts a graph of measured glucose results and actual glucose levels of a person according to one or more embodiments shown and described herein;

FIG. 4 depicts a probabilistic analysis tool and a recursive filter according to one or more embodiments shown and described herein;

FIG. 5 depicts state transitions for a hidden Markov model in accordance with one or more embodiments shown and described herein;

FIG. 6 depicts the operation of a hidden Markov model and a Kalman filter in accordance with one or more embodiments shown and described herein;

FIG. 7 depicts the operation of a prediction algorithm according to one or more embodiments shown and described herein; and

fig. 8 depicts a method of predicting a glucose level of a person using a probabilistic analysis tool and a recursive filter according to one or more embodiments shown and described herein.

Detailed Description

Embodiments described herein relate generally to methods and systems for processing glucose data measured from a person with diabetes and, in particular, for estimating an actual glucose level of the person in the presence of sensor noise and/or sensor failure. For the purposes of defining the present disclosure, a "measured glucose result" is a glucose level of a person measured by a glucose sensor; "actual glucose level" is the actual glucose level of a person; and "estimated glucose level" is an estimated glucose level of the person, which may be based on the measured glucose result.

Fig. 1 depicts a Continuous Glucose Monitoring (CGM) system 10 that may be used to continuously measure a person's glucose level. The CGM system 10 may include a glucose sensor 16 having a needle 18, the needle 18 may be inserted under the skin 12 of a person with diabetes. The tip of the needle 18 may be located in the interstitial fluid 14 such that the measurements taken by the glucose sensor 16 are based on the level of glucose in the interstitial fluid 14. The glucose sensor 16 may be placed on the abdomen of a person or other suitable location and may be secured with tape or adhesive (not shown). Further, the glucose sensor 16 may be periodically calibrated to improve its accuracy. This periodic calibration can help correct for sensor drift due to sensor degradation and variations in physiological conditions at the sensor insertion site. The glucose sensor 16 may also include other components including, but not limited to, a wireless transmitter 20 and an antenna 22. Although depicted in fig. 1 as having a rectangular shape, it is contemplated that the glucose sensor 16 may take on other geometries as well. While the glucose sensor 16 may use the needle 18 to access a person's blood, other suitable devices may also be used to access a person's blood or other fluids in order to obtain glucose measurements, including those that are still to be found.

The glucose sensor 16, when acquiring the measurements, may transmit the measured glucose results to the glucose monitor 26 via the communication link 24. The communication link 24 may beTo be wireless, such as radio frequency or "RF", wherein the measured glucose results are transmitted by electromagnetic waves. For example, "Bluetooth"is a wireless RF communication system that uses frequencies of about 2.4 gigahertz (GHz). Another wireless communication scheme may use infrared light, such as the Infrared data Association(Infrared Data AssociationIrDA) A supported system. Other types of wireless communication are also contemplated, including current technology and technology yet to be developed. The communication link 24 may be unidirectional (i.e., data may be communicated only from the glucose sensor 16 to the glucose monitor 26), or it may be bidirectional (i.e., data may be communicated between the glucose sensor 16 and the glucose monitor 26 in either direction). Further, the communication link 24 may allow communication between two or more devices (e.g., a glucose sensor, a glucose monitor, an insulin pump, etc.). Although fig. 1 shows the communication link 24 as being wireless, it may alternatively be a wired link, such as ethernet. Other public or private wired or wireless links may also be used.

FIG. 2 illustrates one embodiment of a glucose monitor 26, which may include a display 28, a microcontroller 32, and an input device 34. Examples of glucose monitor 26 include, but are not limited to, a blood glucose meter, an insulin pump, a cellular telephone, a smart phone, a personal digital assistant, a personal computer, or a computer server. The microcontroller 32 may be electrically coupled to an input device 34, which may be configured to receive a plurality of measured glucose results from a glucose sensor coupled to the person. The microcontroller 32 may be configured to receive the plurality of measured glucose results from the input device 34. The glucose monitor 26 may also be configured to store a plurality of measured glucose results received from the glucose sensor 16 in a memory (not shown) over a period of time. The microcontroller 32 may also be configured to analyze the plurality of measured glucose results with a probability analysis tool configured to determine a probability of failure for the glucose sensor. Further, the microcontroller 32 may be configured to estimate the glucose level of the person using a recursive filter configured to weight the plurality of measured glucose results with a probability of glucose sensor accuracy. Finally, the microcontroller 32 may be electrically coupled to the display 28 such that the microcontroller is configured to communicate information about the estimated glucose level of the person to the display 28. As discussed herein, the displayed information may include an estimated glucose level of the person and/or a predicted glucose level of the person at some time in the future. In addition, the display may also include an estimate of the quality or uncertainty of the estimated glucose level. Further, the displayed information may include warnings, alarms, etc. regarding whether the person's estimated or predicted glucose level is hypoglycemic or will become hypoglycemic at some time in the future. This may occur, for example, if a person's glucose level falls (or is predicted to fall) below a predetermined hypoglycemic threshold, such as 50 milligrams of glucose per deciliter of blood (mg/dl).

Microcontroller 32 may be an 8-bit device, a 16-bit device, or any other suitable device, and may have an on-chip peripheral device to facilitate its operation. For example, the microcontroller 32 may have internal memory for storing computer programs, internal memory for data storage, and internal memory for non-volatile storage of parameters used by the microcontroller during its operation. Further, the microcontroller 32 may have a timer, a serial communication port, an interrupt controller, and the like. The microcontroller 32 may also be part of an Application Specific Integrated Circuit (ASIC) that may include other circuitry to facilitate operation of the glucose monitor 26. The input device 34 may be a wireless communication module configured to wirelessly receive measured glucose results from a glucose sensor (as shown in fig. 1). As such, the antenna 30 may be used to improve the robustness of the wireless connection. Alternatively, the input device 34 may be a wired communication module that receives measured glucose results over a wired connection, such as over ethernet or similar protocol. The display 28 may include a Liquid Crystal Display (LCD) or other suitable technology. The display 28 may also be configured to convey information to the person in a tactile manner, such as by vibration.

FIG. 3 depicts an example of a two-dimensional plot of measured glucose results 40 from a glucose sensor coupled to a person with diabetes. The horizontal axis represents time in hours, while the vertical axis represents measured and actual glucose levels of a person in milligrams of glucose per deciliter of blood (mg/dl). The measurements may be taken automatically by the glucose sensor at a predetermined rate, such as every minute. The measured glucose result 40 may generally correspond to an actual glucose level 42 of the person. However, the glucose sensor may occasionally fail, as shown during time period 44 in FIG. 3. The failure may be due to a loss of sensor signal caused by a physiological problem near the needle of the glucose sensor, or it may be due to an electrical problem of the glucose sensor itself. During the fault period 44, the measured glucose result 40 may be well below the actual glucose level 42 of the person. At the end of the fault period 44, the measured glucose result 40 may be restored such that it again corresponds to the actual glucose level 42 of the person.

Still referring to fig. 3, the glucose sensor may also show noise from time to time, as shown during time period 46. This noise may be due to physical movement of the needle of the glucose sensor relative to the skin, or it may be due to electrical noise inherent to the glucose sensor itself. During the noise time period 46, the measured glucose results 40 may fluctuate such that some results are above and some results are below the actual glucose level 42. The noise may even appear to oscillate around the actual glucose level 42. At the end of the noise period 46, the measured glucose result 40 may be restored such that it again closely corresponds to the actual glucose level 42 of the person. While shown as occurring at different times, faults and noise may occur simultaneously as well as in succession. Further, the duration of the fault and/or noise may be shorter or longer than depicted in FIG. 3.

Against the above background, embodiments in accordance with the present disclosure are provided that estimate an actual glucose level of a person in the presence of sensor noise and/or sensor failure. Fig. 4 depicts a system 50 configured to estimate a glucose level of a person. The system 50 may receive the measured glucose results 40 from a glucose sensor 16 that may be coupled to a person with diabetes (not shown). The glucose sensor 16 may be coupled to the person as previously described herein or in any suitable manner. The glucose sensor 16 may be configured to periodically measure a person's glucose level, e.g., every minute, and communicate the measured glucose results 40 to the system 50 (e.g., over a communication link as described above).

A system 50 for estimating glucose levels of a person with diabetes may include a probability analysis tool 54 and a recursive filter 52. The probability analysis tool 54 may be configured to receive the measured glucose results 40 and calculate a probability P that the glucose sensor is accurate_A58, i.e., the probability that the glucose sensor 16 is functioning properly (i.e., not malfunctioning). Probability P of glucose sensor accuracy_A58 may be based solely on observable data, such as the measured glucose result 40 and/or changes thereto. Accordingly, the probabilistic analysis tool 54 may be used to distinguish between sensor noise that may have a normal distribution and sensor faults that may not be. Each type of uncertainty may be treated differently due to differences in their uncertainty distributions.

The probability analysis tool 54 may include any number of mathematical algorithms capable of analyzing the measured glucose results 40 and/or changes thereto and calculating the probability P that the glucose sensor is accurate_A58. The probability analysis tool 54 may also be configured to receive other types of dataData of (2), probability of glucose sensor accuracy P_A58 may be based on such data as when the patient eats, when the patient exercises, and when to deliver insulin to the patient. In addition, P_A58 may be based on one or more measurements from an impedance measurement device coupled to the person and configured to measure impedance in the person's body. Such an impedance measurement device may indicate whether the glucose sensor 16 is properly coupled to a person, and thus whether the measured glucose result 40 from the glucose sensor 16 is accurate. Other types of data may also be used.

The probabilistic analysis tool 54 may take many different forms, such as a state machine, a Bayesian model, or other algorithm. In one embodiment, the probability analysis tool 54 may take the form of a simple state machine in which the probability P that the glucose sensor is accurate_AMay always be in the set 0, 1 (i.e., P)_A58 is either 0% or 100%, depending on the state of the state machine). In this example, if Δ CG (i.e., the change in the current measured glucose result compared to the previous measured glucose result) is less than some negative threshold, τ₁The system will transition to a state T where the sensor is inaccurate_A→IAnd if Δ CG is greater than some positive threshold, τ₂Or if the sensor CG value (i.e., the current measured glucose result) is at a physiologically possible glucose value (g)₀And g_max) Within and a certain amount of time deltat has elapsed since the transition to the state in which the sensor is inaccurate_A→I＞τ₃The system transitions back to the sensor accurate state, T_I→A. It can be mathematically expressed as:

T_A→Iif Δ CG < τ₁

T_I→AIf Δ CG > τ₂Or (g)₀＜CG＜g_maxAnd Δ t_A→I＞τ₃)

If neither of these transition conditions is met, the state machine may remain in its current state. This is merely one example of a probability analysis tool 54 in the form of a state machine. The probability analysis tool 54 may take other forms as well.

In another example, the probability analysis tool 54 may include a hidden markov model with the following two states for a glucose sensor: 1) in which the glucose sensor is in the correct state, using "S_A"and 2) a state in which the sensor is inaccurate, with" S_I"is used for representing. The hidden Markov model may provide a state transition function that defines a slave state S_ATransition to state S_ISuch as the following function:

where "CG" is the current measured glucose result, "Δ CG" is the change from the previous measured glucose result to the current measured glucose result, and α₁To alpha₄Is a constant that depends on the characteristics of the glucose sensor. The output values for this function range from zero to one, where zero represents a 0% probability of sensor accuracy and one represents a 100% probability of sensor accuracy. The "min" function takes the minimum of the mathematical expression and the number one (i.e., 100%). This transfer function may be based on the current CG and Δ CG values. Furthermore, the conversion function may be a sigmoid function (sigmoid), wherein the parameter α₁And alpha₃Controlling the position of the S-shaped function transformation, and a parameter alpha₂And alpha₄Controlling the slope of the sigmoid function. These functions may be tuned for a particular patient and/or sensor batch.

Continuing with the example of hidden Markov models, it remains in state S_I(when the current state is S_ITime) may be

And only the value of Δ CG and being in or transitioning to state S_IPrevious probability ofAs a function of (c). The output values for this function range from zero to one, where zero represents 0% probability and one represents 100% probability. The "max" function takes the maximum value of the mathematical expression and the number zero (i.e., 0%). The parameter "γ" is less than one and is designed to stay at S if there is no hold from CG and Δ CG values_ITo gradually change the state of the hidden Markov model back to S_AThe attenuation term of (2). The parameter γ may be a constant and may remain at S when Δ CG is relatively normal_IIs related to the probability of (c). For example, γ may be selected such that the hidden Markov model remains at S when Δ CG is relatively normal_IFor about 10 minutes. The probability function further includes detecting and detecting S_AReturns a rapidly rising sigmoid function in the associated CG signal. Parameter alpha₅Controlling the position of the S-shaped function transformation, and a parameter alpha₆Controlling the slope of the sigmoid function. Both parameters are tuned for a particular patient and/or sensor batch.

Conversion to S_ICurrent probability P of_IIs P_A→IOr P_I→IDepending on the current state being S_AOr S_I. Glucose sensor inaccuracy (i.e., is S)_I) Current probability P of_IMay be (S)_A×P_A→I)+(S_I×P_I→I). Note, status (S)_AOr S_I) When in this state is "1" and otherwise is "0". This includes switching to S_IProbability (P) of_A→I) (given being at S)_AProbability of) and remains at S), and_Iis multiplied by the probability of being currently at S_IThe probability of (c). P_I→IIs equal to 1-P_I→AAnd the probability of sensor accuracy is simply P_A＝1-P_I. Thus, for this example, the probability that the glucose sensor is accurate may be P_A＝1-[(S_A×P_A→I)+(S_I×P_I→I)]。

FIG. 5 depicts two transfer functions P within a histogram of Δ CG_A→IAnd P_I→A(i.e., 1-P)_I→IWhen the current state is S_IHour from S_IConversion to S_AProbability of (d) of the image. The histogram includes a gaussian shape component 68 centered at zero with two tails associated with transitions leading to sensor failure and off-sensor failure. The two transfer functions are plotted on the histogram to indicate that they can be tuned to trigger on the tail of the histogram. The gaussian shape component 68 may represent a range of Δ CG values that may occur during normal operation of the glucose sensor. The Δ CG values located within gaussian shaped component 68 may be due to, for example, sensor noise. The Δ CG values located outside and to the left of the Gaussian shape component 68 may be due to the sensor slave S_AConversion to S_IAnd is caused by this. The shape of this distribution can be used to characterize a batch of glucose sensors after production and to encode the sensors. That is, a particular batch of glucose sensors can be targeted (by adjusting α)₁To alpha₆Parameter) to the transfer function (P)_A→IAnd P_I→I) Adjusted to correspond to gaussian shape component 68. Thus, based solely on the measured glucose results and their variations, hidden Markov models can be used to determine the sensorAccurate probability P_A。

Figure 6 illustrates a diagram showing an example of the operation of a hidden markov model during a glucose sensor fault and in the presence of glucose sensor noise. The graph includes a measured glucose result 40 and a probability P of glucose sensor accuracy_A58. During the time period 70, the glucose sensor may malfunction, thus causing the measured glucose results 40 to become inaccurate; at the same time, P_A58 (determined by hidden markov models) may be reduced from about 100% (prior to time period 70) to approximately 0% during time period 70. This may be due to the detection of a hidden markov model that measures a rapid drop in the value of the glucose result 40 at the beginning of the time period 70 (i.e., when the fault first occurred). At the end of time period 70, the glucose sensor may begin to operate normally (i.e., measured glucose result 40 becomes accurate again) and P_A58 may again increase back to about 100%. As previously described, this may be due to the detection of a hidden markov model of a rapid increase in the value of the measured glucose result 40 at the end of the time period 70 (i.e., when the glucose sensor returns to normal operation). P_AThe rate of change of 58 from near 0% to about 100% may depend on how quickly the glucose sensor transitions from malfunctioning (inaccurate) to normal (accurate) operation. If the transition is relatively fast, P_A58 can quickly transition from near 100% to about 0%. However, if the glucose sensor slowly transitions from malfunctioning to normal operation, P_A58 may also be slowly switched from near 0% to about 100%. If there is little or no hold at S from CG and Δ CG values_IIs then the decay term γ (in equation P)_I→IFound) can allow P_A58 gradually transition back to S_A。

Still referring to FIG. 6, depending on the severity and level of noise, the glucose sensor noise shown to occur during time period 72 may also cause P_A58 is reduced. As depicted in FIG. 6, glucose sensor noise during time period 72 may cause P_A58 are slightly reduced. Of course, both the glucose sensor fault and the sensor noise may have varying amplitude and/or duration levels. Further, glucose sensor faults and sensor noise may partially or completely overlap in time. Hidden Markov models can be configured to determine P under any of these conditions_A58. As will be discussed below, P may be used in the recursive filter_A58 to minimize the effects of glucose sensor failure and thereby provide an accurate estimate of the person's actual glucose level in the presence of glucose sensor failure and/or sensor noise.

Referring again to FIG. 4, a system 50 for estimating glucose levels of a person with diabetes may include a recursive filter 52 that may be used to estimate glucose based on a probability P of accuracy with glucose_A58 to estimate the glucose level of the person from the plurality of measured glucose results. Examples of recursive filters that may be used include kalman filters and Extended Kalman Filters (EKFs). Of course, many other types of recursive filters may be used.

In one embodiment, recursive filter 52 may be a Kalman filter (the references to "Kalman filter" hereinafter also apply to "extended Kalman filter") configured to process measured glucose results 40 (i.e., raw glucose sensor data) in a second order linear system, as embodied in the following equations. The kalman filter may in particular comprise a state vector representing the estimated state of the estimated variable, which in the present example is the glucose level of the person. The Kalman filter may comprise a prediction step in which the prior state and covariance are predicted and in which the A posteriori Kalman gain (K) is updated_k) And measuring the state vector and the covariance. The state vector may be updated each time a new input is received (i.e., recursively). In the present disclosure, the variables in the state vector x may represent an estimate of the actual glucose level of the person based on the measured glucose result 40. The estimated glucose level vector x may represent an estimated glucose level g of the person; first derivative thereofAnd its second derivativeThe measured glucose result vector z may include current CG and Δ CG values. Other dynamic models may also be used. The vectors x and z can be represented asAnd z_k＝[CG ΔCG]^TWhere k denotes the kth sample. The glucose level vector, x, can be estimated using the following equation:where k denotes the k-th sample,K_kis the Kalman gain, and P_AAnd 58 is the probability that the glucose sensor is accurate (from the probability analysis tool). In this way, the probability P of sensor accuracy can be used_A58 to weight the measured glucose results, in matrix z_kIs also disclosed. The supporting equations and matrices for the kalman filter may be as follows:

C_k-i＝CG_k-i-Hx_k-iand an

The parameter β in the matrix A can be expressed₁And beta₂Set slightly less than one so that the estimated glucose level is damped when a sensor failure occurs. The matrix Q may represent the process noise covariance, and K_kThe kalman filter gain may be expressed. Initial estimates for these parameters may be determined as is known in the art.

In Extended Kalman Filters (EKFs), a non-linear model may be usedTo represent the system and also with a non-linear model z_k＝h(x_k) To represent the measurement results. This non-linear model may include inputs u from other sources_kThe other sources may include meals, insulin, exercise, or other inputs that may affect the glucose model. The non-linear model can be derived from a proprietary physiological model of glucose. The prediction step is accomplished by evaluating a non-linear model, and using the Jacobian (Jacobian), F, of the model_kThe state vector is used to calculate the uncertainty of the prediction. This results in a localized linear model for the current system state. The following equation may be used by EKF:

after the predicting step, the junction may be measured in a correcting step using the current glucose sensorFruit CG_k. The correction step may also include a previously calculated probability P of glucose sensor accuracy_A58 (from a probability analysis tool). The kalman filter may be configured to weight the currently measured glucose results with a probability of glucose sensor accuracy. For example, when P_A58 low, the effect of the currently measured glucose results on the kalman filter may be close to zero; on the contrary, when P is_AAt 58 f, the effect of the currently measured glucose results may be higher. Using P in this manner_A58 may be a logical modification to the operation of the kalman filter, as the currently measured glucose results may provide little or no useful information about the actual glucose level of the person when a sensor failure occurs.

Distinguishing sensor faults from sensor noise may facilitate estimating a person's glucose level, and as such, the kalman filter may process them differently. For normally distributed sensor noise, the kalman filter may be configured to average out such noise. This may be due to the fact that the sensor noise may be characterized for each type and/or batch of glucose sensors, including but not limited to the frequency range of the noise and the corresponding range of amplitude variations of the measured glucose results. Some or all of the parameters of the kalman filter may be measured (e.g., byOr) To embody these noise characteristics such that the kalman filter is configured to filter out the noise and provide a relatively accurate estimate of the person's glucose level, even in the presence of sensor noise. On the other hand, sensor fault errors are generally not normally distributed and therefore should be handled differently within the framework of the kalman filter. In one embodiment of the Kalman filter, P_A58 (determined by the probability analysis tool) may be used by the kalman filter to perform the measured glucose resultsWeighted such that the measured glucose results are largely ignored when sensor failures occur.

An example of the operation of the kalman filter is shown in fig. 6, fig. 6 depicting a measured glucose result 40 and an estimated glucose level 60 of a person. As discussed previously, the probability P that the glucose sensor is accurate is also shown_A58. Normally, the estimated glucose level 60 of the person may generally follow the measured glucose result 40. However, during time period 70, the sensor may fail; at the same time, P_A58 may be reduced to near 0% (as determined by the operation of the probabilistic analysis tool) indicating a low probability of the glucose sensor being accurate. Thus, the Kalman filter may be P_A58 are taken into account to mitigate the importance of the glucose results measured during the time period 70 of sensor failure in estimating the glucose level of the person.

With continued reference to fig. 6, the measured glucose results 40 may contain noise during the time period 72. The kalman filter may filter this noise to produce a relatively smooth estimated glucose level 60 during this time period 72. Although the measured glucose results may contain noise during time period 72, P_A58 may remain relatively high (e.g., near 100%) during this time because the probabilistic analysis tool may be able to distinguish between sensor noise and sensor failure. As such, the Kalman filter may continue to place a relatively high importance on the measured glucose results during time period 72 (as represented by P_A58 is relatively high during time period 72).

Uncertainty R of measurement result of glucose sensor_kGenerally not constant. It can currently be taken as the most recent sensor measurement z; probability P of glucose sensor accuracy_A(ii) a Maximum uncertainty of measurement resultsAnd normal uncertainty associated with continuous glucose measurementsIs estimated. Can be expressed as_maxCalculated as the maximum physiological change in glucose in a person with poorly controlled diabetes. Which can be estimated from samples of CGM data. Similarly, σ_CGIs the minimum uncertainty for the glucose sensor when working properly. It may be the best case performance for the sensor and may be estimated from the change in measured glucose results compared to finger stick data when the sensor is operating ideally. Other methods for estimating measurement uncertainty may exist, including using higher frequency glucose sensor data. This may be interpreted as a change in the difference between the most recent past CG measurements and the estimated kalman filter state.

The estimated glucose level of the person determined by the recursive filter may be used to predict the glucose level of the person at some time in the future. These estimates can also be used to analyze human behavior and glucose patterns. Referring back to FIG. 4, a prediction algorithm 62 may be used to predict whether and/or when a person is likely to become hypoglycemic and may provide an associated alarm or alert. The prediction algorithm 62 may receive the estimated glucose level 60 of the person from the recursive filter 52 and may also receive an uncertainty of the estimated glucose level. However, the predictive algorithm 62 may be supplemented with other input data, including meal time, carbohydrates, medications, exercise, insulin dosage, and the like. The prediction algorithm 62 may also further receive information from other data sources, such as measured glucose results (i.e., raw glucose sensor data) or processed glucose sensor data. The prediction algorithm 62 may use gaussian process regression to learn the patient specific prediction model indicated by the training model 64 in fig. 4. The predictive model 62 may also estimate a predicted uncertainty, which may allow the alarm threshold to be adjusted for sensitivity. The alarm threshold may also be adjusted based on the current activity of the patient; for example, the sensitivity may be increased when the patient is sleeping.

As an examplePrediction of hypoglycemia can be accomplished using a system model of a kalman filter or an extended kalman filter. In this example, the prediction step is repeated for the desired prediction timeOrAnd compares the predicted value with a particular threshold value. For example, if the kalman filter is updated every minute, the predicting step may repeat the kalman filter forty-five times in order to predict the person's glucose level from now to forty-five minutes into the future. The predictive model may include additional predictive inputs, such as expected meals, insulin, exercise, or other expected future inputs.

In another example, the estimated glucose value g estimated by the recursive filter and the rate of change of the glucose value are usedTo define a linear prediction that is compared to a hypoglycemic threshold. By multiplying the derivative by the desired prediction times t_ptTo calculate a predicted glucose valueThe prediction is accomplished using the following equation.

As an example, the particular input vector used may include three samples of estimated glucose level (CG) taken at times t-0, -15, and-30 minutes, the derivative at t-15 minutes, and the current derivative of the estimated glucose level, toAnd the time since the last meal. Meal information t_mealAnd the pill information B are optional and may also include other data. This can be expressed mathematically as

x_CG＝[CG_t＝0 CG_t＝-15 CG_t＝-30 ΔCG_t＝0…-15 ΔCG_{t＝-15…-30}]^T

x_meal＝[CG_t＝0 CG_t＝-15 CG_t＝-30 ΔCG_t＝0…-15 ΔCG_{t＝-15…-30} min(t_meal，t_max)B]^T

Gaussian process regression can be based on using (X, y) and test points (X) using the following equations^*，y^*) Training data is presented to predict a person's future glucose levels:

y^*＝k(x^*，X)(k(X，X)+μI)^-1y，

where k (x, x) is a covariance function. A gaussian covariance function may be used to generate the results, although other functions may be used herein. The gaussian covariance function that can be used is:

FIG. 7 depicts the operation of the prediction algorithm. On the left side (from time t-40 to 0) the measured glucose result 40 from the glucose sensor and the estimated glucose level 60 of the person (i.e. the output of the kalman filter) are shown. The current time is t-0. The prediction algorithm may determine the predicted glucose level 80 of the person at some time in the future (i.e., any time greater than t-0). Further, a prediction algorithm may be used to predict whether and/or when a person's glucose level may become hypoglycemic. A hypoglycemic threshold 82 may be established for the person such that an actual glucose level below this threshold means that the person has become hypoglycemic. The hypoglycemic threshold 82 may be determined uniquely for each person. The threshold for an average person may be about 50 mg/ml. Also, the hypoglycemic threshold 82 may vary for each person, such that the threshold is based on time, event, or a combination thereof. As an example, the hypoglycemic threshold 82 for a person may depend on the time of day, whether the person has taken a medication, whether and/or for how long the glucose sensor is in a loss of signal condition, and so forth. The prediction algorithm may be able to predict when the person may become hypoglycemic. In fig. 7, the prediction algorithm may predict that a person will become hypoglycemic at t-45 (i.e., 45 minutes from the current time). Of course, as time progresses, the prediction algorithm may continue to use the latest estimated glucose level (from the kalman filter) and adjust the predicted glucose level accordingly.

In addition to being able to predict future values of a person's glucose level, the prediction algorithm may be further configured to determine a probability that the prediction is accurate. For example, only one or two minutes of prediction may be highly accurate in the future, while predictions at 60 or 70 minutes in the future may be relatively inaccurate. Of course, the probability of prediction accuracy may be a continuum, starting at near 100% for the near future and decaying to near 0% as the prediction extends further into the future. This information can be used in conjunction with the actual prediction itself to provide a hypoglycemic warning system for a person. As shown in FIG. 7, the warning system may not provide an alarm 84 when the predicted glucose level 80 is sufficiently high above the hypoglycemic threshold 82; it may advise to remain cautious 86 when the predicted glucose level 80 approaches within a predetermined range of the hypoglycemic threshold 82; and it may notify the risk 88 when the predicted glucose level 80 falls below the hypoglycemic threshold 82.

The prediction algorithm as discussed in the foregoing may include a training function that learns specific characteristics of a person. The training function may produce training data that may be used in a prediction algorithm and may be weighted based on its impact on producing a prediction. The impact level of the training data may be determined by a covariance function k (x, x) used within the gaussian process regression.

The prediction algorithm may be initialized with a class set of training examples or non-training examples. When new data are measured, they may be incorporated into the prediction algorithm and/or the training function. There are many possible algorithms for including new data. These include adding data to the training set when 1) a predetermined period of time has elapsed, 2) a prediction failed for particular data, 3) no input data is represented in the training set, or 4) the patient or care provider manually includes data (including all new data, if appropriate).

When added to the training set, the new data may be included as a new vector or by re-weighting existing training vectors. The second approach includes the benefit of maintaining a constant memory requirement. The predictive algorithm may be updated on the device immediately after the additional data is added, retrospectively on a personal computer, or retrospectively at the clinic.

Referring to fig. 8, a method 100 for estimating glucose levels of a person with diabetes is shown. The method may include a number of acts, which may be performed in any suitable order. At act 102, the method 100 may receive a plurality of measured glucose results into a computing device from a glucose sensor coupled to a person. At act 104, the method 100 may analyze the plurality of measured glucose results using a computing device with a probability analysis tool configured to determine a probability of glucose sensor accuracy based on the plurality of glucose measurements. And at act 106, the method 100 may estimate the glucose level of the person using the computing device using a recursive filter configured to weight the plurality of measured glucose results with a probability of glucose sensor accuracy. The probability analysis tool and recursive filter may be established as described above.

It should now be understood that the methods and systems described herein may be used to estimate glucose levels in people with diabetes, even in the presence of noise and/or sensor inaccuracies (e.g., sensor signal loss). In addition, the methods and systems described herein may also be used to predict a person's future glucose levels. As such, they may be able to predict whether and/or when a person's glucose levels may become hypoglycemic. Upon detecting or predicting that a person is likely to become hypoglycemic, the method and system may provide corresponding information, such as a warning, to the person. The methods described herein may be stored on a computer-readable medium having computer-executable instructions for performing the methods. Such computer readable media may include compact disks, hard drives, thumb drives, random access memory, dynamic random access memory, flash memory, and the like.

It should be noted that elements of the present disclosure that are "configured" in a particular manner, are "configured" to embody a particular property or function in a particular manner, are described herein as structural statements, as opposed to statements of intended use. More specifically, references herein to the manner in which a component is "configured" denotes an existing physical condition of the component and, as such, is to be taken as a definite recitation of the structural characteristics of the component.

While particular embodiments and aspects of the present invention have been illustrated and described herein, various other changes and modifications can be made without departing from the spirit and scope of the invention. Moreover, while various inventive aspects have been described herein, such aspects need not be utilized in combination. It is therefore intended that the appended claims cover all such changes and modifications that are within the scope of this invention.

Further implementations of the invention, particularly the methods, computer-readable media, and devices as defined herein, provide for the plurality of measured glucose results to include periodic glucose measurements taken about every minute.

In the following, embodiments of the method of the invention, the computer-readable medium of the invention and the device of the invention are given.

Example 1: a method for estimating glucose levels of a person having diabetes, the method comprising:

receiving a plurality of measured glucose results into a computing device from a glucose sensor coupled to a person;

determining a probability P of glucose sensor accuracy with a sensor configured to determine a glucose result based on the plurality of measurements_AUsing a computing device to analyze the plurality of measured glucose results; and

with probability P based on accuracy with glucose sensor_AA recursive filter that weights the plurality of measured glucose results to estimate a glucose level, the glucose level of the person estimated using the computing device.

Example 2: the method of embodiment 1, wherein the computing device comprises a blood glucose meter, an insulin pump, a microprocessor coupled to a glucose sensor, a cellular telephone, a smart phone, a personal digital assistant, a personal computer, or a computer server.

Example 3: the method of embodiment 1, wherein the glucose sensor comprises a continuous glucose monitoring system physically coupled to the person with diabetes and configured to automatically measure a glucose level of the person.

Example 4: the method of embodiment 1, wherein the plurality of measured glucose results comprises periodic glucose measurements taken about every minute.

Example 5: the method of embodiment 1, wherein the probability analysis tool is configured to determine the probability P that the glucose sensor is accurate further based on at least one of_A：

When the person is eating;

when the person exercises;

when to deliver insulin to the person; and

one or more measurements from an impedance measurement device coupled to the person with diabetes and configured to measure impedance within the person.

Example 6: the method of embodiment 1, wherein the probabilistic analysis tool comprises a hidden markov model, wherein:

hidden markov models have two states:

first state S_AWhich indicates that the glucose sensor is accurate, an

Second state S_IIndicating that the glucose sensor is inaccurate; and

the hidden Markov model is configured to determine a probability P of the glucose sensor being accurate based on states of the hidden Markov model and the plurality of measured glucose results_A。

Example 7: the method of embodiment 6, wherein the glucose sensor is in the second stateState S_IIs based on a most recent measured glucose result, a most recent change in the plurality of measured glucose results, or a combination thereof.

Example 8: the method of embodiment 6 wherein the hidden Markov model is derived from the first state S_ATransition to a second state S_IHas a probability of

Wherein CG is a most recently measured glucose result, Δ CG is a most recent change in the plurality of glucose measurements, and α₁、α₂、α₃And alpha₄Is a constant related to the characteristics of the glucose sensor.

Example 9: the method of embodiment 6 wherein the hidden Markov model is held in the second state S_IHas a probability of

Wherein Δ CG is a most recent change in the plurality of measured glucose results,is switched to or in the second state S_IAnd γ, α₅And alpha₆Is a constant related to the characteristics of the glucose sensor.

Example 10: the method of embodiment 6, wherein the probability P that the glucose sensor is accurate_AIs that

1-[(P_A→I×S_A)+(P_I→I×S_I)]Wherein

When the hidden Markov model is in the first state S_ATime S_A1 and otherwise S_A＝0，

When the hidden Markov model is in the second state S_ITime S_I1 and otherwise S_I＝0，

P_A→IIs from a first state S_ATransition to a second state S_IProbability of, and

P_I→Iis in the second state S_IIs maintained in the second state S_IThe probability of (c).

Example 11: the method of embodiment 1 wherein said recursive filter is a kalman filter or an extended kalman filter.

Example 12: the method of embodiment 11, further comprising predicting, using the computing device, a future glucose level of the person with a kalman filter or an extended kalman filter, wherein:

the Kalman filter or the extended Kalman filter comprises a prediction step and a measurement step; and

the predicting step is performed one or more times in order to predict the future glucose level of the person.

Example 13: the method of embodiment 11, wherein the Kalman filter or extended Kalman filter comprises a state vectorWhere k denotes the kth sample of the state vector, g denotes the estimated glucose level of the person,represents the first derivative of g, andrepresenting the second derivative of g.

Example 14: the method of embodiment 13, wherein estimating a glucose level of the person using the computing device comprises determining a state vector

Wherein the content of the first and second substances,z_k＝[CG ΔCG]^T、CG is a most recently measured glucose result, Δ CG is a most recent change in the plurality of measured glucose results, K_kIs the Kalman gain, P_AIs the probability of the glucose sensor being accurate, and beta₁And beta₂Is a constant related to the characteristics of the glucose sensor.

Example 15: the method of embodiment 14 wherein the Kalman gain K_kBased on measurement uncertainty R_kSo that the measurement uncertainty R_kIs variable and is based on the probability of sensor accuracy.

Example 16: the method of embodiment 15, wherein the measurement uncertainty R_kThe method comprises the following steps:

wherein, C_k-i＝CG_k-i-Hx_k-i、τ denotes the time history, σ_maxRepresents the maximum physiological change in glucose for use in a person with poorly controlled diabetes, and σ_CGIndicating a minimum uncertainty for the glucose sensor when accurate.

Example 17: the method of embodiment 1, wherein the recursive filter is configured to estimate the glucose level of the person further based on at least one of:

when the person is eating;

when the person exercises; and

when to deliver insulin to the person.

Example 18: the method of embodiment 1, further comprising predicting, using the computing device, the future glucose level of the person with a regression analysis tool configured to predict the future glucose level based on the estimated glucose level of the person from the recursive filter.

Example 19: the method of embodiment 18, wherein the regression analysis tool comprises a gaussian process regression analysis.

Example 20: the method of embodiment 19, wherein the gaussian process regression analysis comprises a training algorithm configured to learn one or more characteristics of the person related to a glucose level of the person.

Example 21: the method of embodiment 18, further comprising determining an uncertainty of the predicted future glucose level of the person.

Example 22: the method of embodiment 18, wherein the regression analysis tool is configured to predict future glucose levels of the person based on at least one of:

when the person is eating;

when the person exercises; and

when to deliver insulin to the person.

Example 23: a computer-readable medium having computer-executable instructions for performing a method for estimating glucose levels of a person having diabetes, the method comprising:

receiving a plurality of measured glucose results into a computer from a glucose sensor coupled to a person;

determining a probability P of glucose sensor accuracy with a sensor configured to determine a glucose result based on the plurality of measurements_AUsing a computer to analyze the plurality of measured glucose results; and

with probability P based on accuracy with glucose sensor_AA recursive filter for weighting the plurality of measured glucose results to estimate a glucose level, using a computer toThe glucose level of the person is estimated.

Example 24: the computer readable medium of embodiment 23, wherein the computer comprises a blood glucose meter, an insulin pump, a microprocessor coupled to a glucose sensor, a cellular telephone, a smart phone, a personal digital assistant, a personal computer, or a computer server.

Example 25: the computer readable medium of embodiment 23, wherein the glucose sensor comprises a continuous glucose monitoring system physically coupled to the person with diabetes and configured to automatically measure a glucose level of the person.

Example 26: the computer readable medium of embodiment 23, wherein the plurality of measured glucose results comprises periodic glucose measurements taken approximately every minute.

Example 27: the computer readable medium of embodiment 23, wherein the probability analysis tool is configured to determine the probability P that the glucose sensor is accurate further based on at least one of_A：

When the person is eating;

when the person exercises;

when to deliver insulin to the person; and

Example 28: the computer readable medium of embodiment 23, wherein the probability analysis tool comprises a hidden markov model, wherein:

the hidden Markov model has two states

First state S_AWhich indicates that the glucose sensor is accurate, an

Second state S_IIndicating that the glucose sensor is inaccurate; and

Example 29: the computer readable medium of embodiment 28, wherein the glucose sensor is at S_IIs based on a most recent measured glucose result, a most recent change in the plurality of measured glucose results, or a combination thereof.

Example 30: the computer readable medium of embodiment 28, wherein the hidden Markov model is from a first state S_ATransition to a second state S_IHas a probability of

Example 31: the computer readable medium of embodiment 28, wherein maintaining in the second state S_IHas a probability of

Example 32: the computer readable medium of embodiment 28, wherein the probability P that the glucose sensor is accurate_AIs that

1-[(P_A→I×S_A)+(P_I→I×S_I)]Wherein

Example 33: the computer readable medium of embodiment 23, wherein the recursive filter is a kalman filter or an extended kalman filter.

Example 34: the computer-readable medium of embodiment 33, wherein the method further comprises using the computer with a kalman filter or an extended kalman filter to predict the future glucose level of the person, wherein:

Example 35: the computer readable medium of embodiment 33, wherein the kalman filter or the extended kalman filter comprises a state vectorWhere k denotes the kth sample of the state vector, g denotes the estimated glucose level of the person,represents the first derivative of g, andrepresenting the second derivative of g.

Example 36: the computer readable medium of embodiment 35, wherein estimating the glucose level of the person using the computer comprises determining a state vector

Example 37: the computer readable medium of embodiment 36, wherein the Kalman gain K_kBased on measurement uncertainty R_kSo that the measurement uncertainty R_kIs variable and is based on the probability of sensor accuracy.

Example 38: the computer readable medium of embodiment 37, wherein the uncertainty R is measured_kThe method comprises the following steps:

Example 39: the computer readable medium of embodiment 23, wherein the recursive filter is configured to estimate the glucose level of the person further based on at least one of:

when the person is eating;

when the person exercises; and

when to deliver insulin to the person.

Example 40: the computer-readable medium of embodiment 23, wherein the method further comprises predicting, using the computer, the future glucose level of the person with a regression analysis tool configured to predict the future glucose level based on the estimated glucose level of the person from the recursive filter.

Example 41: the computer readable medium of embodiment 40, wherein the regression analysis tool comprises a gaussian process regression analysis.

Example 42: the computer readable medium of embodiment 41, wherein the Gaussian process regression analysis comprises a training algorithm configured to learn one or more characteristics of the person related to a glucose level of the person.

Example 43: the computer readable medium of embodiment 40, wherein the method further comprises determining an uncertainty of the predicted future glucose level of the person.

Example 44: the computer readable medium of embodiment 40, wherein the regression analysis tool is configured to predict future glucose levels of the person based on at least one of:

when the person is eating;

when the person exercises; and

when to deliver insulin to the person.

Example 45: a device for estimating a glucose level of a person having diabetes, the device comprising a microcontroller, an input device, and a display, wherein:

the microcontroller is electrically coupled to an input device configured to receive a plurality of measured glucose results from a glucose sensor coupled to the person, wherein the microcontroller is configured to receive the plurality of measured glucose results from the input device;

the microcontroller is configured to determine a probability P of glucose sensor accuracy based on the plurality of measured glucose results_ATo analyze the plurality of measured glucose results;

the microcontroller is configured to determine a probability P based on the accuracy of the sensor_AA recursive filter that weights the plurality of measured glucose results to estimate a glucose level of the person; and

the microcontroller is electrically coupled to the display such that the microcontroller is configured to communicate information relating to an estimate of a glucose level of the person to the display.

Example 46: the device of embodiment 45, wherein the device comprises a blood glucose meter, an insulin pump, a continuous glucose monitoring system, a cellular telephone, a smartphone, a personal digital assistant, a personal computer, or a computer server.

Example 47: the apparatus of embodiment 45, wherein the glucose sensor comprises a continuous glucose monitoring system physically coupled to the person with diabetes and configured to automatically measure a glucose level of the person.

Example 48: the apparatus of embodiment 45, wherein the plurality of measured glucose results comprises periodic glucose measurements taken approximately every minute.

Example 49: the apparatus of embodiment 45, wherein the probability analysis tool is configured to determine the probability P that the glucose sensor is accurate further based on at least one of_A：

When the person is eating;

when the person exercises;

when to deliver insulin to the person; and

Example 50: the apparatus of embodiment 45, wherein the probability analysis tool comprises a hidden markov model, wherein:

the hidden Markov model has two states

First state S_AWhich indicates that the glucose sensor is accurate, an

Second state S_IIndicating that the glucose sensor is inaccurate; and

Example 51: the apparatus of embodiment 50, wherein the glucose sensor is in the second state S_IIs based on a most recent measured glucose result, a most recent change in the plurality of measured glucose results, or a combination thereof.

Example 52: the apparatus of embodiment 50 wherein the hidden Markov model is from a first state S_ATransition to a second state S_IHas a probability of

Wherein CG is a most recently measured glucose result, Δ CG is a most recent change in the plurality of measured glucose results, and α₁、α₂、α₃And alpha₄Is a constant related to the characteristics of the glucose sensor.

Example 53: the apparatus of embodiment 50 wherein the hidden Markov model is held in the second state S_IHas a probability of

Example 54: the apparatus of embodiment 50 wherein the probability P that the glucose sensor is accurate_AIs that

1-[(P_A→I×S_A)+(P_I→I×S_I)]Wherein

Example 55: the apparatus of embodiment 45 wherein the recursive filter is a kalman filter or an extended kalman filter.

Example 56: the apparatus of embodiment 55, wherein the microcontroller is further configured to predict the person's future glucose level with a kalman filter or an extended kalman filter, wherein:

Example 57: the apparatus of embodiment 55, wherein the Kalman filter or extended Kalman filter comprises a state vectorWhere k denotes the kth sample of the state vector, g denotes the estimated glucose level of the person,represents the first derivative of g, andrepresenting the second derivative of g.

Example 58: the apparatus of embodiment 57, wherein the microcontroller is configured to estimate the glucose level of the person by determining the following state vectors

Wherein

Example 59: the apparatus of embodiment 58 wherein the Kalman gain K_kBased on measurement uncertainty P_kSo that the measurement uncertainty R_kIs variable and is based on the probability of sensor accuracy.

Example 60: the apparatus of embodiment 59 wherein the measurement uncertainty R_kThe method comprises the following steps:

Example 61: the apparatus of embodiment 45, wherein the recursive filter is configured to estimate the glucose level of the person further based on at least one of:

when the person is eating;

when the person exercises; and

when to deliver insulin to the person.

Example 62: the apparatus of embodiment 45, wherein the microcontroller is further configured to predict the future glucose level of the person with a regression analysis tool configured to predict the future glucose level of the person based on the estimated glucose level of the person from the recursive filter.

Example 63: the apparatus of embodiment 62, wherein the regression analysis tool comprises a gaussian process regression analysis.

Example 64: the apparatus of embodiment 63, wherein the Gaussian process regression analysis comprises a training algorithm configured to learn one or more characteristics of the person related to a glucose level of the person.

Example 65: the apparatus of embodiment 62, wherein the regression analysis tool is configured to determine an uncertainty of the predicted future glucose level of the person.

Example 66: the apparatus of embodiment 62, wherein the regression analysis tool is configured to predict future glucose levels of the person based on at least one of:

when the person is eating;

when the person exercises; and

when to deliver insulin to the person.

Claims

1. A method for estimating a glucose level (60) of a person having diabetes, the method comprising:

receiving (102) a plurality of measured glucose results into a computing device from a glucose sensor (16) coupled to a person;

determining a probability of glucose sensor accuracy with a sensor configured to determine a probability of glucose sensor accuracy based on the plurality of measured glucose results (40)(58) Using (104) a computing device to analyze the plurality of measured glucose results (40), wherein the probabilistic analysis tool (54) comprises a means for determining a probability that a glucose sensor (16) is accurate(58) The hidden markov model of (a) is,

in which the hidden markov model has two states:

first stateWhich indicates that the glucose sensor (16) is accurate, an

Second stateIndicating that the glucose sensor (16) is inaccurate; and

the hidden Markov model is configured to determine a probability of the glucose sensor being accurate based on states of the hidden Markov model and the plurality of measured glucose results (40)(58) (ii) a And

with a probability configured to be based on accuracy with a glucose sensor(58) A recursive filter (52) weighting the plurality of measured glucose results (40) to estimate a glucose level (60), estimating the person's glucose level (60) using (106) the computing device to increase or decrease an effect of a current measured glucose result of the measured glucose results in estimating the person's glucose level (60).

2. The method of claim 1 wherein the computing device (32) comprises a blood glucose meter, an insulin pump, a microprocessor coupled to a glucose sensor, a cellular telephone, a smart phone, a personal digital assistant, a personal computer, or a computer server.

3. The method of claim 1 or 2, wherein the glucose sensor (16) is in the second stateBased on the most recent measured glucose result, the most recent change in the plurality of measured glucose results, or a combination thereof, and/or wherein the hidden markov model is from the first stateTransition to the second stateHas a probability of

,

Wherein the content of the first and second substances,is the result of the most recently measured glucose,is the most recent change in the plurality of measured glucose results, an、、Andis a constant related to a characteristic of the glucose sensor, and/or wherein the hidden Markov model remains in the second stateHas a probability of

,

Wherein the content of the first and second substances,is the most recent change in the plurality of measured glucose results,is switched to or in the second stateA previous probability of, and、andis a constant related to the characteristics of the glucose sensor.

4. The method of claim 1 or 2, wherein the probability of the glucose sensor being accurate(58) Is that

Wherein

When the hidden Markov model is in the first stateTime of flightAnd otherwise，

When the hidden Markov model is in the second stateTime of flightAnd otherwise，

Is from a first stateTransition to the second stateProbability of, and

is whenIn the second stateIs maintained in the second stateThe probability of (c).

5. The method of claim 1 or 2, wherein the recursive filter (52) is a kalman filter or an extended kalman filter.

6. The method of claim 5, further comprising predicting, using the computing device, the future glucose level of the person with a Kalman filter or an extended Kalman filter, wherein:

7. The method of claim 6, wherein the Kalman filter or the extended Kalman filter comprises a state vector, and wherein estimating (60) the glucose level of the person using the computing device comprises determining the state vector

,

Wherein the content of the first and second substances,、、、is the result of the most recently measured glucose,is the most recent change in the plurality of measured glucose results,is the gain of the kalman gain (in),is the probability (58) that the glucose sensor is accurate, P_IIs the probability of the glucose sensor being inaccurate, andandis a constant relating to the characteristics of the glucose sensor (16).

8. The method of claim 1 or 2, wherein the recursive filter (52) is configured to estimate the glucose level (60) of the person further based on at least one of:

when the person is eating;

when the person exercises;

when to deliver insulin to the person, and/or wherein the probability analysis tool (54) is configured to determine the probability of the glucose sensor being accurate further based on at least one of：

When the person is eating;

when the person exercises; and

when to deliver insulin to the person; and

9. The method of claim 1 or 2, further comprising predicting the future glucose level of the person using the computing device with a regression analysis tool configured to predict the future glucose level based on the estimated glucose level (60) of the person from the recursive filter (52).

10. A device for estimating a glucose level of a person having diabetes, the device comprising means for performing the steps of the method of any one of claims 1 to 9.

11. A device for estimating a glucose level of a person having diabetes, the device comprising a microcontroller (32), an input device (34), and a display (28), wherein:

the microcontroller (32) is electrically coupled to an input device configured to receive a plurality of measured glucose results (40) from a glucose sensor (16) coupled to the person, wherein the microcontroller (32) is configured to receive the plurality of measured glucose results (40) from the input device (34);

the microcontroller (32) is configured to determine a probability of glucose sensor accuracy based on the plurality of measured glucose results (40)(58) For analyzing the plurality of measured glucose results (40), wherein the probability analysis tool (54) comprises means for determining a probability that the glucose sensor (16) is accurate(58) The hidden markov model of (a) is,

in which the hidden markov model has two states:

first stateWhich indicates that the glucose sensor (16) is accurate, an

Second stateIndicating that the glucose sensor (16) is inaccurate; and

the hidden Markov model is configured to determine a probability of the glucose sensor being accurate based on states of the hidden Markov model and the plurality of measured glucose results (40)（58）；

The microcontroller (32) is configured to determine a probability of being accurate based on the glucose sensor(58) Weighting the plurality of measured glucose results (40) to estimate a glucose level recursive filter (52) to estimate a glucose level of the person (60) so as to increase or decrease an effect of a current measured glucose result of the measured glucose results (40) in estimating the glucose level (60) of the person; and

the microcontroller (32) is electrically coupled to the display (28) such that the microcontroller (32) is configured to transmit information relating to an estimate of a glucose level (60) of the person to the display.

12. The device of claim 11, wherein the device comprises a blood glucose meter, an insulin pump, a continuous glucose monitoring system (10), a cellular telephone, a smart phone, a personal digital assistant, a personal computer, or a computer server.

13. The device of claim 11 or 12, wherein the glucose sensor, the plurality of measured glucose results (40), the probabilistic analysis tool (54), the probability (58) that the glucose sensor is accurate, the probability of the hidden markov model, the recursive filter (52), the kalman filter, the extended kalman filter, the kalman gain, the measurement uncertainty and/or the regression analysis tool are each provided according to the definitions in any one of claims 1 to 10.

14. The device of claim 11 or 12, wherein the recursive filter (52) is a kalman filter or an extended kalman filter, and wherein the microcontroller (32) is further configured to predict the future glucose level of the person with the kalman filter or the extended kalman filter, wherein:

the predicting step is performed one or more times in order to predict the future glucose level of the person,

or therein

The Kalman filter or extended Kalman filter comprises a state vectorWhere k denotes the kth sample of the state vector,represents an estimated glucose level of the person,to representA first derivative of, andto representAnd the microcontroller (32) is configured to estimate the glucose level of the person by determining the following state vector:

,

wherein the content of the first and second substances,、、、is the result of the most recently measured glucose,is the most recent change in the plurality of measured glucose results,is the gain of the kalman gain (in),is the probability that the glucose sensor is accurate, andandis related to glucose sensingA constant relating to the characteristics of the device,

or therein

The microcontroller (32) is further configured to predict the future glucose level of the person with a regression analysis tool configured to predict the future glucose level of the person based on the estimated glucose level (60) of the person from the recursive filter (52).

15. The method of claim 2, wherein the glucose sensor (16) comprises a continuous glucose monitoring system (10) adapted to be physically coupled to a person having diabetes and configured to automatically measure a glucose level of the person.

16. The method of claim 6, wherein the Kalman filter or the extended Kalman filter comprises a state vectorWhere k denotes the kth sample of the state vector,representing an estimated glucose level of the person (60),to representA first derivative of, andto representThe second derivative of (a).

17. The method of claim 7, wherein the kalman gainBased on measurement uncertaintyTo make the measurement uncertainIs variable and is based on the probability of sensor accuracy.

18. The method of claim 17, wherein the uncertainty is measuredThe method comprises the following steps:

,

wherein the content of the first and second substances,、、the time history is represented by a time history,represents the maximum physiological change in glucose for use in a person with poorly controlled diabetes, andrepresents a minimum uncertainty for the glucose sensor (16) when accurate.

19. The method of claim 9, wherein the regression analysis tool is configured to predict future glucose levels of the person further based on at least one of:

when the person is eating;

when the person exercises; and

when to deliver insulin to the person.

20. The method of claim 9, wherein the regression analysis tool comprises a gaussian process regression analysis.

21. The method of claim 20, wherein the gaussian process regression analysis comprises a training algorithm configured to learn one or more characteristics of the person related to a glucose level of the person.

22. The method of claim 9, further comprising determining an uncertainty of the predicted future glucose level of the person.