WO2009136372A1

WO2009136372A1 - Method and device for processing glycemia level data by means of self-adaptive filtering, predicting the future glycemia level and generating alerts

Info

Publication number: WO2009136372A1
Application number: PCT/IB2009/051870
Authority: WO
Inventors: Giovanni Sparacino; Claudio Cobelli; Andrea Facchinetti
Original assignee: Universita' Degli Studi Padova
Priority date: 2008-05-08
Filing date: 2009-05-07
Publication date: 2009-11-12
Also published as: ITMI20080837A1

Abstract

It is described a method and device for processing glycemia level data, by means of self-adaptive filtering, future glycemia level prediction and alert generation. The device is portable and small-sized, and used in combination with Continuous Glucose Monitoring (CGM) devices, thus allowing to integrate and improve the information provided thereby. The device incorporates three innovative modules, based on statistic base algorithms for: filtering data in order to reduce the entity of the noise present therein, adapting to the sensor, to the individual and to the SNR evolution during the monitoring step, preferably by means of a "self-tuning" procedure of the filter parameters in real time and on statistic basis, with the determination of the confidence interval on the filtered glycemia level; predicting the future glycemia level with its confidence interval; generating alerts for preventing hypo-/hyperglycemia episodes based on the previously acquired statistic information.

Description

"Method and device for processing glycemia level data by means of self- adaptive filtering, predicting the future glycemia level and generating alerts" DESCRIPTION State of the art Diabetes - It is estimated that diabetes affects approximately 250 million people in the world today. The number is expected to rise to 350 million in less than 20 years. The rapid, constant increase of diabetic patients make this disease one of the social-health emergencies of the third millennium. Most diabetics follow a metabolic monitoring therapy based on a combination of insulin injections and/or drugs, diet and physical exercise. The therapy is determined by the physician on the basis of glycemia level measurements that the patient measures by him or herself in capillary blood 3 or 4 times a day (self-monitoring). This approach presents inevitable shortcomings due to the low amount of glycemia data available related to the high glycemia range during the day. Due to the shortcomings of the monitoring system, glycemia may exceed normal limits. Hyperglycemia (a situation in which the concentration of glucose in blood is higher than 180 mg/dl) causes various long-term complications (cardiovascular disease, hypertension, retinopathies, etc.), while on the short-term, hypoglycemia (glucose concentration lower than 70 mg/dl) may even be more dangerous (e.g. it may lead to diabetic coma) also because it may be difficult for the patient to recognize, particularly at night.

Known Continuous Glucose Monitoring (CGM) devices. New minimally invasive or in some cases not at all invasive devices for continuously monitoring glucose levels, known as Continuous Glucose Monitoring (CGM) devices, have been developed and marketed during the past years. Such devices provide blood glycemia values at a very high frequency (from every 1 to 5 minutes, according to the sensor) and continuously stabile operation for several days (from 3 to 7 days, according to the technology). Most of the devices known heretofore are of the "minimally invasive type", because each determination requires the "collection" of a very small amount of interstitial fluid by the sensor, in which glycemia is measured by observing a chemical reaction. The need to eliminate this limited source of invasiveness has spurred, even more recently, to start studies on non-invasive measurement techniques. Various entirely non-invasive devices (e.g. based on ionophoresis, optic technologies, etc.) are being studied, and imaginably their use will be possible in the forthcoming years. CGM devices are unanimously considered the most interesting due to the high glycemia monitoring frequency (e.g. from 1 minute to 5 minutes), which allows to continuously monitor glycemia levels during the entire day. Retrospective glycemia level analysis (glycemia level data can normally be downloaded by means of an appropriate software present in the device) may for example identify hypo- /hyperglycemia episodes which are not revealed by 3-4 daily measurements determined in capillary blood, and thus significantly help reviewing the specific patient's therapy.

One of the features of CGM sensors is their capacity to estimate the current glucose level in real time. This makes them particularly interesting for recognizing potentially dangerous episodes in real time, such as exceeding of the aforesaid hypo-/hyperglycemia thresholds. For this reason, some CGM sensors are provided with alert generation methods, which warn the patient when the estimated glycemia value exceeds predetermined thresholds. The currently developed Continuous Glucose Monitoring (CGM) devices implement different technologies. Some examples of known CGM devices are briefly outlined below.

- System Gold™ Continuous Glucose Monitoring System (CGMS)®, made by Medtronic MiniMed, Northridge, CA. This device consists in a thin measuring sensor to be inserted in the abdomen and which communicates (by means of a specific wire) with a portable device (monitor) which stores the glycemia values every 3 minutes for a maximum total of 72 hours. The device does not allow to display data in real-time; data may only be retrospectively displayed after having been downloaded to a PC.

- Guardian® Real Time Continuous Glucose Monitoring System, made by Medtronic MiniMed, Northridge, CA. This device consists in a thin measuring sensor to be inserted in the abdomen, a transmitter and a portable device (monitor). Transmitter and monitor reciprocally communicate by means of wireless technology. The monitor displays the following in real time: the measured glycemia value; a trend graph, which plots the glycemia during the last period of time (e.g. 3 hours); a trend arrow, which indicates whether the glycemia value is increasing, decreasing or stationary. The monitor is provided with an alert generator. The alerts are generated (i) when the current glycemia value exceeds certain thresholds, (ii) when the glycemia level changes (based on trend arrow) and (iii) when the prediction (which may be set from 5 to 30 minutes forward) of the glycemia level exceeds a given value.

- GlucoWatch® G2 Biographer, made by Cygnus Inc, Redwood City, CA. The monitoring system looks like a wristwatch. Glucose is measured by using a reversed ionophoresis technique through the skin using a disposable component which is stuck by means of an adhesive onto the skin, thus putting it into contact with a small electric current. The electric charges take the glucose onto the skin surface, where it is measured using a traditional glucometer. The device generates an alert in presence of a rapid change in glucose level, excessive sweating and for each measurement over or under a given threshold.

- GlucoDay® S, made by Menarini Diagnostics, Florence, Italy. This device consists in a micro-dialysis tube for collecting a sample of interstitial fluid by means of a pump and conveying it into the device itself, where glycemia is measured. The device is provided with a readout which displays the measured value in real-time.

- STS-7™ Continuous Glucose Monitoring System, made by Dexcom, San Diego, CA. The device consists of three parts: a needle sensor which makes the measurement, a wireless transmitter and a receiver. The receiver is a portable wireless device provided with a readout, which allows to display the current glycemia value in real time, and a trend graph of the previously measured values. The device can generate alerts when the measured value exceeds given thresholds.

- FreeStyle Navigator™ Continuous Glucose Monitor, made by Abbott Laboratories, Alameda, CA. The device consists of three parts: a needle sensor which makes the measurement, a wireless transmitter and a receiver. The receiver is a wireless portable device provided with a readout for displaying the current glycemia value in real time and a trend arrow which indicates if the glycemia value is increasing, decreasing or stationary. The device can generate alerts when the measured value exceeds given thresholds.

Problems of the known devices. However, at least with regards to the real-time generation of hypo-/hyperglycemia alerts, the performance of these known CGM devices are currently disappointing, because many of the generated alerts (nearly 50%) are false. Therefore, the low reliability of such alert systems makes them useless in practice, and even potentially dangerous, because the patient cannot distinguish an alert to be taken into consideration from a false one. One of the problems responsible for the poor hypo-hyperglycemia alert generating performance of CGM devices is constituted by the presence of "noise" in the measurements. In other words, the sensor measurement presents an error which may cause the generation of a false alert or the non-recognition of an actually occurring episode (missed alert), thus exposing the patient to risks or incorrect therapy choices.

Furthermore, it would be obviously preferable to prevent hypo-/hyper glycemia episodes, instead of simply generating alerts when such an episode occur. For example, an alert generated to indicate that without intervention a hypoglycemia situation will occur within 20 minutes would allow the patient to ingest and absorb sugar in time.

The above-mentioned problems will now be described in detail. - SIGNAL NOISE MAKES THE MEASUREMENT UNRELIABLE. The glycemia value measured by the CGM device presents an interference component which causes an error in the glycemia level determination. In the specific case, by altering the real glucose concentration value, such an error may generate false alerts or the non-recognition of an actually occurring episode (missed alert), thus exposing the patient to risks of incorrect therapy choices. Some CGM devices contain therein algorithms for reducing this interference component in the signal, but such algorithms are evidently too "crude". For example, the filtering algorithm implemented in the CGM device is set and "blocked" once and for all. Consequently, the algorithm cannot adapt to the different intensities of the noise component actually found in different monitoring steps, nor can it adapt to the variation of such intensity during the use of a device in a single monitoring step, variation for example due to a deterioration of the measuring sensor performance. In other terms, by maintaining the same sensor, the ratio between real signal and noise component (Signal-Noise Ratio, SNR), varies by intensity not only from individual to individual (inter-variability), but also during the monitoring step in a single individual (intra- variability).

Finally, none of the CGM devices known heretofore adopt strategies, specifically on statistic basis, for taking into consideration that the filtering needs to be adapted, in real time, to each individual's different SNR entity. Where implemented in the device, the currently used filtering methods are insensitive to SNR variations from individual to individual (inter-individual variability) and to the SNR variations during the same monitoring step (intra-individual variability). - GLYCEMIA PREDICTION FOR PREVENTIVE PURPOSE. In the scope of preventing the consequences of diabetes, it is obviously preferable to prevent hypo-hyperglycemia episodes instead of simply generating alerts when such episodes occur. For example, an alert generated to indicate that without intervention a hypoglycemia episode will occur within 20 minutes would allow the patient to ingest and absorb sugar in time. The noise component described in the previous paragraph prevents the correct application of predictive algorithms to the obtained measurements, thus negatively affecting predictive system performance. Some articles related to techniques for predicting short-term glycemia levels (e.g. 30 - 45 - 60 minutes) are present in scientific literature.

It is worth noting that the term "prediction" is often used with various meanings also in clinical context only. Herein, the term will be related to a short-term forecast (e.g. 45 minutes) of future glycemia level useful for preventing hypo- /hyperglycemia, when obtained in real time.

For this purpose, the article by Sparacino et al., listed in literature under [1], illustrates predictive algorithms based on polynomial and self-regressive models. The articles by Palerm et al., listed in literature under [2], [3], on the identification and prediction of hypoglycemia episodes, illustrate a prediction algorithm based on Kalman Filter (KF), which however cannot adapt to the features of the noise present on the glycemia signal and in which SNR is identified on subjective basis. In short, none of the CGM devices of those known heretofore adopt strategies, specifically on statistic basis, for predicting future glycemia by using self-adaptive signal models (i.e. customized for the monitored individual and with time-variant parameters to adapt to the variability of the glycemia signal dynamics also within the same individual).

- ALERT GENERATION. All the known CGM devices which contemplate alert generation to warn the patient about the danger of the current situation, simply base such a generation on the currently measured glycemia value. The alert is generated if the value exceeds given thresholds (in positive or negative direction). As previously mentioned, the signal noise may cause the generation of a false alert, or worse a missed alert. A high percentage of false alerts thwarts all the advantages that a CGM device introduces. In CGM devices, the alert generation system is simply based on the value measured by the sensor. If such a value exceeds predetermined thresholds, then the alert is generated. Such an alert is not very effective in presence of high signal noise, which is put into a poor accuracy of the sensor.

Finally, none of the known CGM devices adopt strategies, specifically on statistic basis, to generate hypo-/hyperglycemia alerts on the basis of filtered and/or predictive glycemia measurements on statistic and self-adaptive base. Summary of the invention.

It is thus the object of the present invention to solve the aforesaid problems and indicate a method and device for processing glycemia level data, specifically data from Continuous Glucose Monitoring devices, capable of:

- filtering the noise which is present in said data in real time, preferably by means of a filter, the parameters of which can adapt to the variability of the noise found from individual to individual (inter-variability), as well as within the same individual (intra-variability), by means of a "self-tuning" procedure of the parameters of the filter itself;

- short term predicting the future glycemia level in real time, on the basis of the signal with SNR improved by filtering;

- generating alerts in real time for warning the patient when the glycemia value reaches a level such to imply risks for health, by exploiting the filtered and/or predicted signal and the corresponding confidence level calculated by virtue of the statistic base of the methods.

It is the object of the present invention a small-sized portable device for use in combination with Continuous Glucose Monitoring (CGM) devices, which allows to integrate and improve the information provided thereby. Specifically, the device incorporates three innovative modules, based on original statistic based algorithms, for: a) filtering data in order to reduce the entity of the noise present therein, adapting itself to the sensor, to the individual and to the SNR evolution during monitoring, preferably by means of a "self-tuning" procedure of the filter parameters in real time and on statistic basis, with determination of the filtered glycemia level confidence interval; b) predicting the future glycemia level with its confidence interval; c) generating alerts for preventing hypo-/hyperglycemia episodes, based on the previously acquired statistic information.

It is also the object of the present invention a method for processing glycemia level data, said glycemia level data being generated by a device of the Continuous Glucose Monitoring (CGM) type, characterized in that it comprises the steps of:

- filtering the noise from said glycemia level data, thus obtaining filtered values x_tjt (1) and measuring the error variance P_t|_t(1 ,1 ) thereof (on the basis of which the so-called standard confidence interval is constructed);

- calculating the probability of hypoglycemia Phypo and hyperglycemia Phyper, using said filtered values x_t|t(l) and error variance P_t|t(1 ,1 ), hypoglycemia thresholds hypoT and hyperglycemia hyperT, and comparing such probabilities with respectively correlated reference thresholds Rhypo and Rhyper;

- generating a hypo- or hyperglycaemia "SEVERE ALERT" condition if Phypo is higher than Rhypo, or if Phyper is higher than Rhyper.

Furthermore, the method optionally contemplates obtaining said noise filtering of glycemia level data by means of a filter the parameters of which are estimated using a "self-tuning" procedure so as to adapt the filter to the noise variability. Furthermore, the method optionally comprises the further steps of: - calculating a predictive algorithm on said filtered values x_t|t(l) and with estimated error variance Pt|t(1 ,1 ), thus obtaining a predicted glycemia value *_t|t+PH (l) referred to a given number (PH) of instants forward in time, and the corresponding error variance Pt+_PH|t(1_>1) (from the root of which the prediction confidence interval is constructed);

- calculating the probability of hypoglycemia Phypo and hyperglycemia Phyper on said predicted glycemia value and corresponding confidence interval;

- generating a hypo- or hyperglycemia "PREDICTIVE ALERT" condition if Phypo is higher than Rhypo, or if Phyper is higher than Rhyper. It is important to note that the method and the device presented here are independent from the CGM sensor technology, which may be either minimally invasive (needle, micro-dialysis, etc.) or non-invasive (optical, ionoforesis, etc.).

It is the particular object of the present invention a method and device for processing glycemia level data by means of real-time, self-adaptive filtering future glycemia level prediction, as better described in the claims, which form an integral part of the present description.

Brief description of the figures.

Further objects and advantages of the present invention will be apparent from the detailed description that follows of an example of embodiment of the same (and its vatiants) and the accompanying drawings given by way of non-limitative example, in which:

FIG. 1 shows a flow chart of the main operations carried out by the device object of the invention;

FIGS. 2.1 , 2.2, 2.3 show flow charts of the data filtering operation in FIG. 1 ; FIG. 3 shows flow charts of the glycemia level prediction operation in FIG. 1 ;

FIG. 4 shows a block diagram of the device object of the present invention.

The same numbers and reference letters in the figures refer to the same elements or components.

Detailed description. The data processing method according to the invention is described below.

The operations shown in the flow chart in FIG. 1 are carried out when the data are received from the CGM device, as illustrated in detail below.

Table 1 below contains a list of the symbols used in the flow charts in the figures

(with corresponding meanings).

Table 1

With reference to FIG. 1 , filtering and prediction operations are carried out in real time on the glycemia level measured by a CGM device.

In detail, the device according to the invention receives the glycemia value (in mg/dl or mg/mmol) measured by the CGM sensor in combination with which is it used as an input (step 1.0) at time instant t. The value is filtered ( x_tjt(l) ) according to the operations which will be described below in detail with reference to FIG. 2 (step 1.1). The newly obtained filtered value of x_t|t(l) is provided (step 1.2) with

Pt|t(1 ,1 )_> which corresponds to the obtained data estimation error variance, and which is in direct relationship with the confidence interval of the filtered glycemia measurement.

The next step consists in retrieving the hypo- and hyperglycemia threshold values (hypoT and hyperT, respectively) (step 1.3) and the correlated reference probability thresholds (Rhypo and Rhyper) from the system memory. Assuming that the filtered glycemia punctually has a probability plot describable with a Gaussian distribution with average x_t|t(l) and variance Pt|t(1 ,1 ). the hypo- and hyperglycemia probabilities (Phypo and Phyper) are easily calculated (step 1.4) by using the hypoT and hyperT values as a reference. A hypo- and hyperglycemia "SEVERE ALERT" is generated (step 1.5) if Phypo is higher than Rhypo,, or if Phyper is higher than Rhyper, otherwise no alert is generated. The filtered data (along with the confidence interval) and the possible alert are shown on a readout. A predictive algorithm is then applied (step 1.6) (an embodiment of which is shown in FIG. 3 and shown in detail below) and the glycemia value predicted for a certain number (PH) of sampling times ahead in time is obtained. The estimate error variance, and thus its confidence interval, is also available for the predicted value. Assuming again a Gaussian distribution, the hypo- and hyperglycemia probabilities (Phypo and Phyper) are calculated (step 1.8). A hypo- and hyperglycemia "PREVENTIVE ALERT" is generated (step 1.9) if Phypo is higher than Rhypo, or if Phyper is higher than Rhyper, otherwise no alert is generated. The predicted data (along with the confidence interval) and the possible alert are shown on the display (step 1.10).

With reference to FIG. 2.1 , 2.2 and 2.3, the filtering operation of the signal fed back by the CGM device exploits models (named a priori), which make bland, yet realistic assumptions concerning the two contributions in play, i.e. "measurement noise" and "true CGM signal":

- the measurement noise is additive and may be assimilated to a white noise process, with zero average and unknown variance σ² (such a variance may be variable in time);

- the "true" CGM signal (i.e. without noise) can be assimilated to the creation of a so-called random-walk (or an integrated random-walk) process driven by white noise of zero average and unknown variance λ² (such a variance may be variable in time). The parameters σ² and λ² determine the filtering action because their ratio is related to the data signal-noise ratio (SNR). However, in each single acquisition by means of the sensor, their value is unknown, and may even vary in time. According to the invention, the filtering algorithm uses a statistic criterion for estimating the values of σ² and λ² , from the data only, by means of a "self-tuning" procedure, shown in the steps going from 2-20 to 2-29. Such an estimate may be carried out in real time and may even be updated whenever a new sensor measurement is received, so as to adapt to SNR variability in time. The estimation process of σ² and λ² is as follows. The glycemia value (in mg/dl or mg/mmol) measured by the CGM sensor in combination with which is it used is received as an input at time instant t (step 2.11 ). The data is stored (step 2.12) in the device. The last received unfiltered values n are loaded (step 2.14) in vector y (dimension n x 1). The vector y is given by: y = Gu + v (1 ) where u is a column vector (n x 1 ) which represents the vector of the noise-free data, G is an identity matrix (dimension n x n, step 2.15), and the vector v represents the measurement noise (zero average and variance σ²). The vector u has an a priori covariance matrix given by λ² (F^TF)^"1. Specifically, F is an n x n matrix (step 2.15) obtained from a Toeplitz lower triangle matrix, with first column given by [1 -1 0 ... 0]^τ , multiplied by itself a number of times (ord) equal to the number ord of integrators used to describe the signal regularity expectations (e.g., ord=1 if the process u is described as a random-walk, ord=2 if the process u is described as an integrated random-walk, etc.). Matrix F will thus again be a Toeplitz lower triangle with first column given for example by [1 -1 0 ... 0]^τ if the process u is described as a random-walk, by [1 -2 1 0 ... 0]^τ in the case of a priori integrated random-walk model. The algorithm thus determines the minimum error variance linear estimate. Numerically, the method proceeds by seeking the vector 0 which minimizes a cost function constituted by two terms: the first (step 2.25) relates to data adherence (specifically given by WRSS=(y-Gϋ)^τ(y-Gϋ), quadratic sum of the weighed residues), while the second term (step 2.26) provides the signal regularity (WESS=ύ^τF^τFύ, quadratic sum of weighed estimates). The second term is weighed by the parameter Y = σ²/λ² which, as mentioned above, is unknown. If Y were known, the vector u would be found by means of the formula (step 2.24): ύ = (G^τG+γF^τF)-¹G^τy (2)

A statistic base criterion is used in order to determine y, described for example in the article listed in literature under [4]. In practice, a trial and error type of procedure is used, in logarithmic scale in order to speed up the operation. Initially, a tentative value of Y = 10exp((logio(γmax)+logio(γmin))/2) is employed (step 2.21), where γrτiax and ymin are the extreme values that Y may assume. After having calculated ϋ with tentative Y value, (eq. 2), the consequent WRSS and WESS and q(γ)= trace(G(G^τG+γF^τF)^'1G^τ) are calculated (degrees of freedom, Cf. [4]). If the following inequality (step 2.28) is fulfilled: WRSS(γ) _ WESS(r)

< P (3) n-q(γ) Φ)

where p is a predetermined threshold value, then the newly tested tentative value of Y is the final value, and the measurement noise variance is determined (step 2.29) as:

₂ WRSS σ² = """" (4) n -q(γ)

and subsequently that of the signal as λ² = σ²/ v. If (3) is not fulfilled, the method evaluates (step 2.20) which of the two terms in the module is higher and consequently modifies either ymax or vmin. The entire process described heretofore is thus repeated, recalculating both y and u until (3) is satisfied or until the maximum predetermined number of iterations is reached. Such values of λ² and σ² are thus individualized, i.e. different not only from sensor to sensor but also from individual to individual. Because the vector y can be updated at each new sampling instant t (as if the measurements n contained therein refer to a sliding window), the parameters λ² and σ² may be periodically updated, so as to adapt to the possible SNR variations which occur during the various monitoring days due to the prolonged immersion of part of the sensor in the biological fluids (e.g. 5-7 days).

After having estimated the values of σ² and λ² (step 2.30), the a priori information are translated in the corresponding state model:

where x_t represents the state vector (the first component of which precisely coincides with the filtered signal), A and H are transition matrixes, and using the notation used until here, y_t represents the scalar glycemia value measured by the CGM device at instant t. Vectors v_t and w_t are respectively the vectors of the noises related to measurement and state. The method assumes that v_t and w_t are reciprocally independent, with covariance matrixes R and Q. R is scalar and equal to σ², Q is scalar and equal to λ² in the case of a priori type random-walk model, and is of 2x2 dimension instead in case of a priori integrated random-walk model, with Q(1 ,1)= λ² and all other elements zero. The state can be estimated from observations by employing Bayes estimate, implemented in static manner or recursive manner by means of the known Kalman Filter (KF). In our case, we chose to implement the recursive approach of this second alternative. The equations are typical KF equations, described for example in the book listed in literature under [5]. In the equations shown in the flow chart in Figure 2, the matrix P_t|t represents (step 2.35) the state estimate covariance matrix error at instant t given the measurements made until the instant t included. The method related to the glycemia level prediction for sampling periods PH ahead is shown in Figure 3.

The part which comprises the predictive algorithm may be implemented in different manners. An example of implementation based on the cyclic use of the KF predictive step only is shown below. At instant t, x_t|, and Pm are available, i.e. the estimate of the state vector (containing the filtered glycemia value in first position) with corresponding uncertainty (step 3.1 ). The KF predictive step consists of the one-step state prediction equation:

and of the equation which describes the relative uncertainty:

P₁₊₁₁₁ = AP₁₁₄A⁷ + Q (7)

Having indicated with PH the number of sampling periods ahead for which the glycemia value will be predicted, the equations (6-7) (steps from 3.2 to 3.6) will be iterated PH times, replacing from the second iteration onwards, x_t+,_|t+, with x_t+,_|t and P_t+i|_t+i with P_t+ϊ|t, (i being the index representing the iteration cycle) in order to calculate the predicted value ( X_t+PHI_t(I ) ) and its error variance (Pt+PH|t(1 Λ )) ■

The use of the Kalman filter KF on CGM data is already known per se, having been previously used for example in [3], where however the matrixes Q and R were fixed in equal manner for all individuals (as if there were no inter-individual variability) and once and for all (as if there were no intra-individual variability). Conversely, a statistic-based method is used for the first time in the present invention which allows to identify the unknown parameters σ² and λ² which determine the matrixes Q and R of the KF in real time and similarly to periodically update them in automated manner, even at each sampling instant. The personalised estimate, which may be adjusted in time, of the matrixes Q and R which determine the KF allows for the first time to adapt the CGM filtering action to SNR variations which may occur during the various monitoring days (e.g. due to deterioration of sensor performance). With reference to FIG. 4, there are described the features and advantages introduced by the device in accordance with the invention.

In the described embodiment, the device is configured as a portable one, and it processes the data received (via wireless connection) from any Continuous Glucose Monitoring (CGM) device and comprises the following components: - READOUT D. This is used to communicate:

- A first number: the glycemia value obtained at the end of the processing/filtering process.

- A second number: the predicted glycemia value for PH sampling periods ahead (the prediction horizontal expressed in minutes thus results by multiplying PH for the sampling time of the CGM sensor, expressed in minutes).

- WIRELESS RECEIVER RW for receiving data from CGM devices.

- PROCESSOR CPU for processing.

- MEMORY MEM for temporarily storing data.

- SOUND DEVICE DS for emitting an alert if the hypo-hyperglycemia probabilities are beyond the predetermined threshold.

- BUTTONS/INTERFACE Pl for allowing the user to interact with the device. More in detail, the most immediate manner for implementing the device object of the present invention is to use a Personal Digital Assistant (PDA) and an operating system.

The PDA presents all the necessary hardware features, the processor for processing (sufficient to run the data processing operations in satisfactory time), the RAM for temporarily storing data, a fixed memory for storage, IrDA and Bluetooth communication cards for dialoguing with the CGM devices, a Wi- Fi/LAN/802.11 card for connecting for example a personal computer (for downloading and analysing processing data off-line), a touch-screen display which acts as user interface, for displaying processing results and for entering information (button function).

The operating system comprises all the software needed to dialogue with the hardware (processor, I/O peripherals, wireless devices, etc.) and further allows to install a processing code. There are different operating systems which may be installed on a PDA, such as Palm OS, Windows Mobile and Symbian OS. The use of Symbian OS is recommended here. Briefly, the PDA - operating system pair solves the following problems:

- running a series of data processing operations which are difficult to be encoded directly in machine language to be then directly run by a microprocessor. By using such an operating system, the data processing and displaying software for implementing the above-described method may be encoded in high-level programming language, such as for example C++ (or Java).

- graphically displaying the results of the data processing operation, i.e. showing the user the current and future glycemia values and possibly also the graphic plot of the glycemia profile monitored over the past hours. Such graphic operations may be easily reproduced by means of a graphic display, which requires a graphic interface.

- wirelessly communicating with CGM device via, for example, Bluetooth technology. - storing the processing results.

- generating both graphic and acoustic critical glycemia level alerts.

The processing, data display and alert generation software is described above with reference to figures 1, 2.1, 2.2, 2.3.

With regards to this part of the software, the device contemplates setting the filtering algorithm so that:

- the parameters σ² and λ² of the filter are estimated, although on statistic basis, only once in the initial interval (named burn-in interval, and equivalent to assume that SNR remains constant);

- the estimate on statistic basis of the parameters of the filter σ² and λ² is periodically updated. For example, it is possible to have the parameters updated each hour or also more frequently or even at each new measuring step, e.g. every 60 seconds with the Navigator (Abbott Diabetes Care) sensor or every 3 minutes with the Glucoday (Menarini Diagnostics) sensors, so as to ensure the filter response as stressed as possible by the possible SNR variations occurring during monitoring. As mentioned, the various previously suggested CGM devices use alert generation for warning individuals of the risk of a critical episode. The device object of the invention however contemplates an innovative alert generation system. Indeed, the invention uses an algorithm for generating alerts which is not blindly based on the data obtained by the CGM device, but rather on the probability that a certain episode may occur. The calculation of such a probability is only possible in virtue of the previous use of stochastic algorithm for filtering and predicting, in which, in addition to the punctual value, the uncertainty (confidence interval) which characterizes the filtered and predicted glycemia, and from which the hypo-/hyper glycemia can indeed be calculated, is also obtained. The suggested device generates the alert only if such a probability exceeds a predetermined threshold.

The advantage of such an alert generation system is indeed based not only on the punctual value but also on the accuracy of the measurement (given by value uncertainty).

The above-described method with reference to Figures 1 , 2.1 , 2.2, 2.3 and 3 may be implemented with a high-level programming language, such as C++ or Java. These encoding languages also contemplate the use of libraries for interfacing with graphic devices (e.g. a readout). In this manner, it is easier to make a code which shows the processing output on screen.

The part of the method object of the present invention may thus be advantageously implemented by means of a computer program which comprises encoding means for implementing one or more steps of the method, when this program is run on a computer. Therefore it is intended that the scope of protection is extended to said computer program and also to computer-readable means which comprise a recorded message, said computer-readable means comprising program encoding means for implementing one or more steps of the method, when said program is run on a computer. The advantages deriving from the application of the present invention are apparent.

The main strength consists in the accuracy of the supplied measurements, crucial for obtaining real-time indications, e.g. danger warnings of past or imminent hypo- /hyperglycemia episodes. The device object of the invention improves clinical performance of known CGM devices.

During prototype use on retrospectively analysed data, the filtering algorithm has proven to work very satisfactorily. Specifically, the noise removal being equal, the algorithm introduces a considerably smaller delay than that observed by implementing the known mobile average filtering techniques as described above. A fundamental advantage is due to the capacity of the device to adapt to variability between sensors to variability between individuals and also above all to intra- individual variability. Such a particularity is significantly important in clinical perspective, given the expected variability of sensor performance during monitoring, induced, for example, by the prolonged immersion by the sensor itself (e.g. by its needle) in biological fluids. Prototype results obtained from retrospectively analysed data show that by fixing confidence at 90%, it is possible to reach sensitivity in the order of 95% and specificity of over 70%, maintaining an average time gain of approximately 20 minutes, sufficient for example to compensate in time for the possible hypoglycemia episode by orally ingesting sugar.

The major feature which characterizes the originality of the device is that it can adapt to the different SNR caused by sensor technology (inter-sensor variability), it is adaptable to the SNR caused by sensor operation in the given individual (inter- individual adaptability), and it can be updated in real time to SNR variations which occur during monitoring (intra-individual variability). Indeed, the parameters σ² and λ² which characterize the action (and the performance) of the filter not only are self-adaptive by means of a "self-tuning" procedure, to the single sensor and to the single individual, but may be periodically updated, in the most extreme case, whenever the device receives new data.

The device may interface with CGM sensor of all types, e.g. both minimally invasive (subcutaneous, micro-dialysis, etc.) which are entirely non-invasive (optical, ionophoresis, etc.).

Other possible variants of the described non-limitative example are possible, without because of this departing from the scope of protection of the present invention, comprising all the equivalent implementations for a person skilled in the art. From the above description, a person skilled in the art will be capable of implementing the object of the invention without introducing further constructive details.

Literature quoted in the description [1] Sparacino G., Zanderigo F., Corazza S., Maran A., Facchinetti A., Cobelli C, "Glucose concentration can be predicted ahead in time from continuous glucose monitoring sensor time-series", IEEE Trans Biomed Eng, 2007 May;54(5):931-7. [2] Palerm C.C., Willis J. P., Desemone J., Bequette B.W., "Hypoglycemia prediction and detection using optimal estimation", Diabetes Technol Ther, 2005 Feb;7(1 ):3-14. [3] Palerm C, Bequette W., "Hypoglycemia Detection and Prediction Using Continuous Glucose Monitoring - A Study on Hypoglycemic Clamp Data", J Diabetes Sci Technol, 2007 Sep;1(5):624-9.

[4] Sparacino G., Cobelli C, "A stochastic deconvolution approach to reconstruct insulin secretion rate after a glucose stimulus", IEEE Trans Biomed Eng, 1996 May;43(5):512-29.

[5] Anderson B.D.O., Moore J., "Optimal Filtering", Dover Publications, 2005.

Claims

1. A method for processing glycemia level data, said glycemia level data being generated by a device of the^' Continuous Glucose Monitoring (CGM) type, characterized in that it comprises the steps of: - filtering the noise from said glycemia level data, obtaining filtered values x_t)t(l) and measuring the error variance Pt|t( 1 ,1 ) thereof (on the basis of which the so- called standard confidence interval is constructed);

- calculating the probability of hypoglycemia Phyper and hyperglycemia Phyper, by using said filtered values x_t)t(l) and error variance P_t|t(1 ,1), hypoglycemia thresholds hypoT and hyperglycemia hyperT, and comparing such probabilities with respectively correlated reference thresholds Rhypo and Rhyper;

- generating hypo- or hyperglycemia "SEVERE ALERT" condition if Phypo is higher than Rhypo, or if Phyper is higher than Rhyper.

2. A method for processing glycemia level data according to claim 1 , characterized in that said noise filtering of said glycemia level data is obtained by means of a filter, the parameters of which are estimated with a "self-tuning" procedure so as to adapt the filter to noise variability.

3. A method for processing glycemia level data according to claim 2, characterized in that it comprises the further steps of: - calculating a predictive algorithm on said filtered values x_t|t(l) and error variance

P_tJ_t(1 ,1), thus obtaining a predicted glycemia value for a given number of instants (PH) forward in time, and a corresponding error variance and thus a confidence interval;

- generating a hypo- or hyperglycemia "PREDICTIVE ALERT" condition if Phypo is higher than Rhypo, or if Phyper is higher than Rhyper.

4. A method for processing glycemia level data, according to claim 2, characterized in that said step of noise filtering comprises processing said glycemia level data for estimating unknown variance parameters σ² and λ², said glycemia level data comprising: - an additive measurement noise, which may be assimilated to a white noise process, with zero average and known variance σ², possibly variable in time and different from individual to individual;

- a 'real" glycemia signal, without noise, which may be assimilated a priori to a process of the so-called multi-integrated white noise type, where the noise has zero average and unknown variance λ², possibly variable in time, and however different from individual to individual;

- said unknown variance parameters σ² and λ² determining the filtering action, so that their ratio is related to the data signal-interference ratio (SNR) of the glycemia level data.

5. A method for processing glycemia level data according to claim 4, characterized in that said unknown variance parameter estimate σ² and λ² is generated by means of an automated "self-tuning" procedure of the parameters which comprises: - composing a vector y (dimension n x 1 ) with the last n non-filtered glycemia level data values, the vector being y = Gu + v, where u is a column vector (n x 1 ) which represents the vector of the noise-free data, G is an identity matrix (dimension n x n), and the vector v represents the measurement noise (zero average and variance σ²); said vector u having a priori covariance matrix given by λ² (F^TF)^"1, F being a Toeplitz lower triangular matrix n x n, with a first column given by [1 -1 0 ... 0]^τ in the case of random-walk a priori model, [1 -2 1 0 ... 0]^τ in the case of a priori integrated random-walk type model;

- determining a linear estimate with minimum error variance, searching for the vector ύ which minimizes a cost function constituted by two addenda: the first related to data adherence, given by WRSS=(y-Gϋ)^τ(y-Gύ), quadratic sum of weighed residues, - the second provides the signal regularity, given by WESS=ύ^τF^τFύ, quadratic sum of weighed estimates, said second term being weighed by a parameter Y = σ²/λ² ;

- determining said vector ϋ by means of the formula

Q = (G^τG+γF^τF)-¹G^τy - determining Y by means of a statistic trial and error type criterion, in which the method initially starts with a tentative value of

Y = 10exp((logio(γmax)+logi₀(γmin))/2), where γrrιax and γrriin are the extreme values that Y may assume;

- after having calculated ύ with tentative value of Y, calculating said values WRSS, and WESS and q(γ)=trace(G(G^τG+γF^τF)-¹G^τ)

- if the following inequality is fulfilled:

WRSS(y) WESS{γ)

—r- < p where p is a predetermined threshold value, then the n-q{γ) q(r) newly tested tentative value of Y is the final value, and the measurement noise variance is determined as:

2 WRSS σ — ^■ n - q(γ) and subsequently that of the signal as λ² = σ²/ Y;

- if instead said inequality is not fulfilled, the method evaluates which of the two terms within the inequality module is higher, and consequently modifies either γmax or γmin;

- the entire process is thus repeated, by recalculating both Y and u until either the inequality is fulfilled or until the maximum predetermined number of iterations is reached.

6. A method for processing glycemia level data according to claim 5, characterized in that, after said estimation of unknown variance parameters σ² and λ² , it comprises the steps of:

- translating the a priori information into a corresponding state model:

where Xt represents the state vector, in which the first component coincides with said filtered signal, A and H are transition matrixes, y_t is the non-filtered glycemia level data at instant t, vectors v_t and w_t are respectively the vectors of the noises linked to measurement and state, assuming v_t and w_t being reciprocally independent, with covariance matrixes R and Q, containing said estimates of σ² and λ² respectively.

7. A method for processing glycemia level data according to claim 3, characterized in that it said step of calculating of a predictive algorithm comprises the steps of: - cyclically using only the predictive step of a Kalman Filter KF, having the estimate of the state vector x_t|t containing in first position the filtered glycemia value, and corresponding uncertainty Pψ available at instant t,

- said predictive step of the Kalman Filter (KF) comprising the state prediction equation:

^« Λ ^Λ

^Xt+1|t ⁼ "*t|t

and the prediction equation of its uncertainty:

P₁₊₁₁₁ = AP₄₁₁A¹ + Q

P_t+i_|t, (index i representing the iteration cycle), thus obtaining said predicted glycemia value (xt+pH|t(1 )) and its uncertainty (P_t+PH[t(1 ,1 )).

8. A device for processing glycemia level data, specifically for implementing the method in accordance with any of the preceding claims, said glycemia level data being generated by a Continuous Glucose Monitoring (CGM) device, characterized in that it comprises: - at least one displaying element of at least:

- said glycemia value obtained at the end of the processing/filtering process;

- said predicted glycemia value PH minutes ahead in time;

- at least one receiver for receiving said glycemia data;

- at least one processing unit, provided with memory; - at least one device for emitting an alert if the hypo-hyperglycemia probabilities are beyond the predetermined threshold; - devices for allowing user interaction.

9. A device according to claim 8, characterized in that said at least one processing unit comprises a Personal Digital Assistant (PDA) and an operating system.