IT201800005918A1 - SYSTEM FOR THE DETECTION OF MALFUNCTIONS IN DEVICES FOR THE ADMINISTRATION OF INSULIN - Google Patents

Description

Description of the industrial invention entitled:

“SYSTEM FOR DETECTION OF MALFUNCTIONS IN

DEVICES FOR THE ADMINISTRATION OF INSULIN "

TECHNICAL FIELD

The present invention relates to the sector of systems for detecting malfunctions in medical devices. In particular, the invention refers to a system for detecting malfunctions in devices for the administration of insulin, such as devices comprising an insulin pump and a glucose sensor. The invention also refers to a device of this type incorporating such a detection system.

STATE OF THE ART

Type I diabetes is a disease in which the patient no longer produces insulin and therefore needs to be artificially administered using a special medical device, such as an insulin pump. Often this insulin pump is integrated with a glucose sensor and in some cases the delivery is automated by a control algorithm (artificial pancreas).

While the insulin pump periodically infuses a predetermined dose of insulin through a subcutaneous needle connected to the pump via a catheter, the artificial pancreas is a more advanced system that works in a closed loop.

In addition to the simple insulin pump, the artificial pancreas also includes a sensor capable of detecting the patient's blood glucose value. The sensor periodically measures, for example every five minutes, the patient's blood glucose and sends the information to a control unit that controls the pump and decides how much insulin to administer in order to keep the patient's blood glucose values under control.

A problem with this type of medical device is that failures or malfunctions can occur which are potentially very dangerous for the patient's health. Both the sensor and the pump could break or malfunction, with the risk of delivering the wrong amounts of insulin. However, even in the absence of pump or sensor breakages, there may be other problems that affect the correct functioning of the medical device, for example the catheter that connects the pump to the needle could come off and the insulin cannot be administered.

Given the importance of proper insulin administration, manufacturers of artificial pancreases are using systems to detect the malfunction of the medical device and warn the patient of the malfunction.

Some manufacturers measure the force exerted by the pump motor in infusing the insulin and determine if there are problems in delivering the insulin.

Other systems for detecting malfunctioning of the artificial pancreas use a mathematical model of the patient to make predictions on the trend of blood sugar over time. If the sensor measurements deviate from the predictions provided by the model beyond a certain tolerance, then this is interpreted as a malfunction of the medical device and an alarm is signaled to the user. However, this model-based approach has the limit that it is difficult to identify a model that is suitable for each patient.

An alternative to model-based systems is constituted by model-free systems that are based on the analysis of historical data. Basically, these systems perform an analysis of the collected data and learn from them what are the "normal" operating modes and "anomalous" behaviors, according to a classification criterion that depends on the method. For example, one of the simpler methods assumes that if a value is very distant from the others observed it is probably an anomaly (outlier). These methods can be divided into supervised (supervised learning) and unsupervised (unsupervised learning) depending on whether or not they require a preliminary procedure of labeling historical data as "failed" / "not failed".

An example of model-free methods, applied to the detection of malfunctions of medical devices for the administration of insulin is described in Rojas et al. “Multivariate statistical analysis to detect insulin infusion set failure,” in A merican Control Conference (A CC), 2011. IEEE, 2011, pp. 1952– 1957. In this article, Rojas et al. use supervised learning methods to train a classifier to detect insulin pump failures.

The main disadvantage of supervised model-free systems is that they require a set of data labeled as "failed" / "not failed" sufficient to allow learning. However, such data is not readily available because during the data collection period, the patient finds himself having to use a medical device whose functioning is not known and could be faulty. The data must therefore be retrospectively labeled by an operator with the disadvantage of requiring a very complex, time-consuming and error-prone operation. such a procedure can only be done on a small group of patients who can be monitored, and then use the large-scale dataset in the medical devices that are sold. In doing so, however, the historical data collected is not the data. history of the specific patient and, in part, the advantage of model-based systems is lost, which is to be adapted and customized ti on the individual patient.

The need is therefore felt for systems for detecting failures in medical devices for the administration of insulin which are efficient and adaptable to the individual patient.

PURPOSE AND SUMMARY OF THE INVENTION

An object of the present invention is to solve the problems of medical devices for administering insulin, and in particular of systems for detecting malfunctions in this type of device.

In particular, an object of the present invention is to present a system for detecting malfunctions in devices for the administration of insulin, such as devices with insulin pump and glucose sensor, such as e.g. the artificial pancreas. In particular, it is an object of the present invention to present a system for detecting malfunctions of the model-free type with unsupervised learning.

It is also an object of the present invention to present a system for detecting malfunctions in a device for the administration of insulin of the model-free type with unsupervised learning that can be customized on the patient to whom the medical device is applied.

These and other objects of the present invention are achieved by means of a system for detecting malfunctions in devices for the administration of insulin of the type comprising at least one sensor for monitoring glucose present in the blood of a person and an insulin infusion pump, incorporating the characteristics of the appended claims, which form an integral part of the present invention.

In one embodiment, the system comprises:

- a data processing unit adapted to receive glucose concentration measurements from said glucose monitoring sensor and further adapted to receive measurements of the amount of insulin infused by said infusion pump;

- a data entry unit operationally connected to the data processing unit and capable of receiving estimates from a user of the amount of carbohydrates consumed;

- a storage unit operationally connected to the data processing unit and suitable for storing glucose concentration measurements, measurements of the amount of insulin infused and estimates of the amount of carbohydrates consumed.

From the measurements stored in the storage unit, the data processing unit calculates, at a given instant of time, an estimate of the residual insulin present in the blood (IOB), and an estimate of the residual glucose present in the blood (COB ).

The data processing unit calculates the value of a DCOB variable (t) which represents the change in current glucose weighted by the residue of ingested carbohydrates and estimated to still be present in the blood:

where the variable that describes the derivative of the measurements of glucose concentration over time, COB (t) is the variable that describes, over time, the estimate of residual carbohydrates present in the blood, α and β are two positive constants. The variable DCOB (t) will be indicated in the following "weighted glucose variation" or "glucose variation weighted on residual carbohydrates".

In addition, the data processing unit is further able to calculate an ICOB variable (t) which represents an estimate of the residual insulin weighted by the residue of ingested carbohydrates and estimated to still be present in the blood:

where IOB (t) is the variable that describes, over time, the estimate of residual insulin present in the blood, γ and δ are two positive constants. In the following, this variable will be called "weighted residual insulin" or "residual carbohydrate weighted residual insulin".

In addition, the data processing unit is capable of executing at least one anomalous value detection algorithm to detect anomalous values within a set of patient data. Each of the patient data includes at least one triplet of values consisting of: a value of the variable glucose change (DCOB) in a given instant of time, a value of the weighted residual insulin (ICOB) in the instant of time, and a measurement of the glucose concentration carried out by the sensor in the instant of time.

The data processing unit is able to generate an alert signal if the algorithm detects at least one anomalous value within the patient data set.

The proposed triad constitutes a particular non-linear transformation of the signals acquired by the data processing unit, and is a replacement (or supplementary) of these signals due to the fact that it significantly increases the possibility of highlighting anomalies.

Preferably, the outlier detection algorithm is of the unsupervised learning type. Furthermore, more preferably, this algorithm is the Local Outlier Factor (LOF) and / or the Connectivity-based Outlier Factor (COF) and / or the Isolation Forest (iForest).

In this way it is possible to determine the anomalies through unsupervised learning carried out on the patient data on which the medical device is applied and without the need for “faulty / not faulty” labeling to be provided to the device.

In one embodiment, the data processing unit is adapted to employ at least one outlier detection algorithm of the unsupervised learning type selected from the Local Outlier Factor (LOF), the Connectivity based Outlier Factor (COF) and the Isolation Forest (iForest) to Weighted Glucose Change (DCOB) data, Weighted Residual Insulin (ICOB) data, and Glucose Concentration Measurement data. The data processing unit is able to generate an alert signal if the anomalous value detection algorithm generates a result outside a predefined range of threshold values.

Therefore, the system preferably operates using one or more algorithms on a set of three data defined by the weighted glucose variation (DCOB), the weighted residual insulin (ICOB) and the measurement of the glucose concentration.

In a further embodiment, the data processing unit is adapted to employ at least one outlier detection algorithm of the unsupervised learning type selected from the Local Outlier Factor (LOF), the Connectivity-based Outlier Factor (COF) and the Isolation Forest (iForest) to a working portion of the weighted glucose change (DCOB), weighted residual insulin (ICOB) data, and glucose concentration measurement data. The work portion is defined by excluding a predefined number of data in earlier time instants and a predefined number of data in later time instants with respect to a predetermined time instant.

This allows, in the algorithm evaluation, not to consider a set of strongly correlated data and therefore further decrease the density around anomalous data, thus increasing performance.

These and other objects of the present invention are further achieved by means of an insulin delivery device comprising an insulin pump, a glucose sensor and a malfunction detection system incorporating the features of the appended claims, which form an integral part of the present invention. invention.

BRIEF DESCRIPTION OF THE DRAWINGS

These and further characteristics and advantages of the present invention will become evident from the description of the preferred embodiment, illustrated by way of non-limiting example in the attached figure, in which:

Figure 1 is a schematic view of an insulin delivery device incorporating the system for detecting malfunctions in insulin delivery devices, according to the present invention.

DETAILED DESCRIPTION OF THE INVENTION

While the invention is susceptible to various modifications and alternative constructions, some preferred embodiments will be described below in detail. It must be understood, however, that there is no intention of limiting the invention to the specific embodiment described, but, on the contrary, the invention intends to cover all modifications, alternative constructions, and equivalents that fall within the scope of the invention as defined in the claims.

The use of "for example", "etc." , “Or” means non-exclusive alternatives without limitation unless otherwise indicated. The use of "includes" and "includes" means "includes or includes, but not limited to" unless otherwise indicated.

In the following, a preferred embodiment will be described, as illustrated in Figure 1, of the system 11 for detecting malfunctions in devices for the administration of insulin of the type comprising at least one sensor for monitoring glucose present in the blood of a person and an infusion pump. of insulin. Further, a preferred embodiment will be described, as illustrated in Figure 1, of an artificial pancreas 1 comprising a sensor 21 for monitoring glucose present in the blood of a person and an insulin infusion pump 31 and integrating the system 11 for detecting malfunctions.

According to a further embodiment (not shown), the system for detecting malfunctions in devices for the administration of insulin could be realized by means of a centralized architecture able to detect malfunctions in one or more devices for the administration of insulin connected to the aforementioned centralized architecture . This architecture, preferably of the cloud type, could allow the detection of malfunctions remotely. In the present embodiment, the insulin infusion pump 31 is a device that allows the continuous infusion, 24 hours a day, of insulin into the subcutaneous tissue of a patient, helping to achieve the best possible glycemic control. The Infusion Pump 31 is capable of delivering insulin into the subcutaneous tissue by continuous basal infusion, to maintain normal fasting blood glucose values, and by rapid bolus infusion, to adjust blood glucose values in relation to intake. a meal or too high blood glucose values. A peculiar characteristic of infusion pumps is that of continuously infusing insulin with the possibility of varying the infusion rate during the day according to the needs of the individual, thus reproducing in a more precise way the physiological presence of the hormone in the body which varies in the span of the day.

In particular, with the term "continuous infusion" it is meant, in the present invention, that the infusion pump 31, or the relative control device, is adapted to measure the quantity of insulin infused to the patient at predetermined instants of time.

The continuous glucose monitoring sensor 21 is also inserted into the patient's subcutaneous tissue to measure the glucose concentration at predefined instants of time, avoiding the patient from making the same measurement by pricking the finger.

The artificial pancreas 1 can be equipped with both the infusion pump 31 and the monitoring sensor 21, or be operationally connected to the latter. Furthermore, the artificial pancreas 1 according to the present invention is further provided with the system 11 to detect malfunctions in devices for the administration of insulin according to the present invention or can be operatively coupled to the latter.

The system 11 for detecting malfunctions according to the present invention therefore allows to detect the malfunctions that can afflict the infusion pump 31 with regard to mechanical defects, occlusions or detachments of the pump 31 itself from the catheter. The aforementioned defects can in fact lead to infusing a reduced amount of insulin which can lead to hyperglycemia, or high blood glucose values.

The aforementioned system 11 includes a data processing unit 111, a data entry unit 311 and a storage unit 211.

The data processing unit 111 comprises one or more microprocessors programmed to carry out a plurality of operations as defined in the operation described below. In particular, the data processing unit 111 according to the present invention is operatively connected to the aforementioned glucose monitoring sensor 21 and to the aforementioned insulin infusion pump 31. In this regard, the data processing unit 111 is designed to receive measurements of a patient's glucose concentration from the continuous glucose monitoring sensor 21, called "g (t)". In addition, the data processing unit 111 is adapted to receive measurements of the amount of insulin infused to the patient by the infusion pump 31, called "i (t)". The latter value i (t) includes both the administration of a continuous basal infusion and the administration of a rapid bolus.

The data entry unit 311 is, for example, made up of a data entry interface, or user interface, provided directly on the artificial pancreas 1 device or by an external system operatively connected to it.

By way of example, this external system can comprise a wireless connection device, for example a smartphone or similar devices. In particular, the data input unit 311 is operatively connected to the data processing unit and is able to receive from the user the estimate of the quantity of carbohydrates consumed in one or more of said instants of time called "m (t)", which are known to be associated with increased postprandial glucose.

Finally, the storage unit 211 is preferably of the non-volatile type, for example a storage unit 211 arranged in the artificial pancreas device 1 or in an external system that is operationally similar to that described above for the data entry unit 311. In particular, the storage unit 211 is operatively connected to the data processing unit 111 and suitable for storing the measurement of the glucose concentration g (t), the measurement of the quantity of insulin infused i (t) and the estimate of the quantity of carbohydrates m (t) taken in in said instants of time. This storage allows, therefore, to keep the data relating to the performance of the aforementioned signals for the calculation of the subsequent variables, as described below.

The user can be modeled as a dynamic system whose output to be controlled, at each instant of time, is the measure of the glucose concentration g (t). This output is influenced by the measurement of the quantity of insulin infused i (t) and by the estimate of the quantity of carbohydrates m (t) taken in the same instants of time.

Therefore, a set of potential variables that describe the state of the system at each instant of time, according to the known art, is the following: g (t), ġ (t), i (t) and m (t), where ġ (t) is the variable describing the derivative of the measurements of glucose concentration over time g (t).

However, the insulin infused in a given instant of time does not have a direct impact on the behavior of the system at the same instant of time but on the behavior of the same system in the following 1-8 hours. This is due to the slow absorption dynamics and persistence of past insulin, whose concentration decreases exponentially over a time range of 4-8 hours after the insulin infusion. Therefore, on the basis of the aforementioned parameters, or starting from the measurements stored in the storage unit 211, the data processing unit 111 is able to calculate, at a given instant of time, the estimate of the residual insulin present in the blood ( IOB) as a function of the measure of the amount of insulin taken at each instant of time i (t). In particular, the IOB value at a given time instant is preferably calculated as a convolution of the measure i (t) with an exponential decay function, as described and incorporated by reference in /. Ellingsen, E. Dassau, H. Z isser, B. Grosman, M. W. Percival, L. Jovanovioc, and F. J. Doyle III, “Safety constraints in an artificial pancreatic _ cell: an implementation of model predictive control with insulin on board,” Journal of diabetes science and technology, vol. 3, no.3, pp. 536-544, 2009.

Similarly, meals ingested at a given instant of time have no impact on the behavior of the system at that instant, as meals are absorbed in 15-45 minutes and gradually dissolve over the next 4-6 hours. Therefore, on the basis of the aforementioned parameters, or starting from the measurements stored in the storage unit 211, the data processing unit 111 is able to calculate, in a given instant of time, the estimate of the residual carbohydrates present in the blood (COB ) as a function of the estimate of the quantity of carbohydrates consumed in each instant of time m (t). In particular, the COB value at a given time is preferably calculated as a convolution of the measure m (t) with an exponential decay function, as described and incorporated by reference in M. Schiavon, C. Dalla Man, Y. C. Kudva, A. Basu, and C. Cobelli, “Quantitative estimation of insulin sensitivity in type 1 diabetic subjects wearing a sensor-augmented insulin pump,” Diabetes care, vol. 37, no. 5, pp. 1216–1223, 2014.

Therefore, a set of potential variables that better describe the state of the system at each instant of time, according to the known art, is the following: g (t), ġ (t), IOB (t) and COB (t ).

To improve the possibility of detecting the fault with respect to the known art, the system 11 according to the present invention determines any anomalies on the basis of a modified and improved set of variables defined by a non-linear transformation of the acquired signals (g (t), m (t) and i (t)) and / or processed (ġ (t), IOB (t), COB (t)) by the data processing unit 111, and stored in the storage unit 211, as described in detail below .

In this regard, the data processing unit 111 is able to calculate, in a given instant of time, a variable called "weighted glucose variation" or "residual carbohydrate weighted glucose variation" (DCOB) as a function of the first derivative the measurement of the glucose concentration in the blood weighted for the estimate of the carbohydrates ingested and still present in the blood COB. Preferably, the variable weighted glucose variation (DCOB) is described by the following formula

in which

the variable describing the first derivative of the measure of

blood glucose concentration,

the variable that describes, over time, the estimate of carbohydrates

residues present in the blood (COB), e

they are two constants, preferably positive and preferably predefined.

In detail: in non-abnormal conditions, blood glucose values are expected to rise in response to a meal. Hence, elevated COB values are, even under conditions of proper functioning, associated with elevated blood glucose variations. In this condition, however, the DCOB ratio is small. On the contrary, during an occlusion of the infusion pump 31, the blood glucose increases persistently even not in response to a meal and therefore also in correspondence with low or zero COB values. When COB (t) is small or equal to zero (i.e. distant from meals), DCOB (t) is strongly influenced by the change in glucose ġ (t). Hence, large positive DCOB (t) values occur only when glucose rises disproportionately not at meals and are symptoms of infusion pump 31 failure.

From the description it can be seen how the weighted glucose variation DCOB (t) decreases the importance of the glucose variation ġ (t) when the COB value is large (meals) and amplifies it when COB is small (fasting).

The constants are predefined constants, preferably positive, which allow you to increase the impact of one signal on the other.

In addition, the data processing unit 111 is able to calculate, for each of the instants of time, an ICOB variable called "residual weighted insulin" or "residual insulin weighted on residual carbohydrates". Preferably, the weighted residual insulin variable (ICOB) is described by the following formula

in which

the variable that describes, over time, the estimate of residual insulin

present in the blood (IOB),

the variable that describes, over time, the estimate of carbohydrates

residues present in the blood (COB), e

they are two constants, preferably positive and preferably

predefined.

In conditions of correct functioning of the device 1, in order to keep the subject in a good metabolic control, the new carbohydrates ingested are compensated by an extra administration of insulin, that is the rapid bolus. So after a meal IOB will tend to grow a lot even in case of correct functioning, but in this condition the ICOB value remains small because COB also grows at the same time. On the contrary, in the presence of a fault, the pump 31 tries to neutralize the glucose g (t) by increasing the insulin infusion. However, the attempt is unsuccessful, since insulin i (t) is not actually administered due to the fault. In this case the IOB value increases significantly compared to the COB value. For example, if the fault occurs far from the meal, the IOB increases while the COB remains small and the ICOB is very high.

From the description it is clear that the weighted residual insulin variable decreases the importance of IOB when COB is large since large simultaneous values of this quantity are expected in diabetic patients. Conversely, the denominator does not penalize IOB when COB (t) is small, so large positive ICOB values are symptomatic of a great effort to lower glucose g (t).

The constants, ome, are predefined constants, preferably positive, which allow you to increase the impact of one signal on the other.

Finally, the data processing unit 111 is adapted to execute at least one anomalous value detection algorithm for detecting anomalous values within a patient data set, in which each of the patient data comprises at least one set of three values, preferably consisting of from a value of the weighted glucose variation variable (DCOB) at a given instant of time, a value of the weighted residual insulin variable (ICOB) at said instant of time, and a measurement of the glucose concentration carried out by the sensor 21 at said instant of time time.

Therefore, the data processing unit 111 is capable of generating an alert signal if the algorithm detects at least one anomalous value within the patient data set. Preferably, the data processing unit 111 generates an alert signal if the algorithm detects a result outside a predefined range of threshold values.

Therefore, the set of preferred characteristics which describe the state of the system at a given instant of time, according to the present invention, is the following: g (t), ICOB (t) and DCOB (t). Preferably, moreover, the constants

they are all equal and equal to 10. Different values can still be

employees.

Other embodiments of the invention can expand the triad including successive derivatives of the signals g (t), IOB (t), COB (t), ICOB (t) and DCOB (t) and their combinations, possibly non-linear.

The anomalous value detection algorithm used by the processing unit according to the present invention is preferably of the unsupervised and model-free learning type.

In this way it is possible to determine the anomalies through unsupervised learning carried out on the patient data on which the medical device is applied and without the need for "faulty / not faulty" labeling to be provided to the device and without the need to determine a mathematical model for the patient's normal response (model-free).

The algorithms that will be discussed below are the Local Outlier Factor (LOF), the Connectivity-based Outlier Factor (COF) and the Isolation Forest (iForest). Further algorithms can however be used and fall within the inventive concept according to the present invention.

As previously described, the system 11 preferably operates by employing one or more algorithms, in particular selected from LOF, COF and iF, on a set of three data defined by the variable glucose variation DCOB (t), from the weighted residual insulin ICOB (t) and by measuring the glucose concentration g (t) and able to fully describe the system.

The LOF and COF algorithms belong to the family of density-based methods. These approaches are based on the study of neighborhood values: an observation in a dense region is considered an inlier, or a non-anomalous value, while a given point in a low-density region is considered an outlier, or an anomalous value.

The basic idea of the LOF algorithm lies in comparing the density of a point with respect to neighboring points instead of considering all the points in the data set. By doing so, it is possible to identify regions with similar densities and points with a significantly lower density than its neighbors, the latter identified as outliers. The data processing unit 111 is adapted to use at least the above-mentioned algorithm for detecting abnormal values of the unsupervised learning type LOF to the entire set of preferred characteristics which describe the state of the system, that is to the data of the variable weighted glucose variation. DCOB (t), ICOB weighted residual insulin (t) and measurement of glucose concentration. Therefore, the data processing unit 111 is able to generate an alert signal if the algorithm detects an anomalous value within the patient data set, or a result outside a predefined range of threshold values.

The use of this algorithm will not be described in detail below, as it is similar to what has already been described and incorporated by reference in M. M. Breunig, H.-P. Kriegel, R. T. Ng, and J. Sander, “Lof: identifying density-based local outliers,” in A CM sigmod record, vol. 29, no.2. A CM, 2000, pp. 93-104.

The COF algorithm is a modified version of LOF, more effective in the case of complex and low-dimensional data structures.

Also in this case the data processing unit 111 is adapted to use at least the above-mentioned algorithm for detecting anomalous values of the unsupervised learning type COF to the entire set of preferred characteristics which describe the state of the system, i.e. the variable variation of DCOB weighted glucose (t), ICOB weighted residual insulin variable (t) and measure of glucose concentration. Therefore, the data processing unit 111 is able to generate an alert signal if the algorithm detects an anomalous value within the patient data set, or a result outside a predefined range of threshold values.

The use of this algorithm will not be described in detail below, as it is similar to what has already been described and incorporated by reference in J. Tang, Z. Chen, A. W.-c. Fu, and D. Cheung, “A robust outlier detection scheme for large data sets,” in 6th Pacific-A is Conf. On K nowledge Discovery and Data Mining. Citeseer, 2001 and in J. Tang, Z. Chen, A. Fu, and D. Cheung, “Enhancing effectiveness of outlier detections for low density patterns,” A dvances in K nowledge Discovery and Data Mining, pp. 535-548, 2002.

The iF algorithm is based on the hypothesis that anomalous points must be few and isolated. The main idea of this method is based on the concept of space partitioning: an isolated point, or an anomaly, requires on average fewer iterations of space partitions to be isolated through partitioning than a non-anomalous value.

Also in this case the data processing unit 111 is able to use at least the above-mentioned algorithm for detecting anomalous values of the unsupervised learning type iF to the entire set of preferred characteristics which describe the state of the system, i.e. the variable variation of DCOB weighted glucose (t), ICOB weighted residual insulin (t) and glucose concentration. Therefore, the data processing unit 111 is able to generate an alert signal if the algorithm detects an anomalous value within the patient data set, or a result outside a predefined range of threshold values.

The use of this algorithm will not be described in detail below, as it is similar to what has already been described and incorporated by reference in F. T. L iu, K. M. Ting, and Z.-H. Zhou, "Isolation forest," in Data Mining, 2008. ICDM'08. Eighth IEEE International Conference on. IEEE, 2008, pp. 413-422.

Algorithms, especially LOF and COF, are methods that effectively capture isolated points. However, in the problem of this invention the points are strongly correlated temporally and never completely isolated, because the patient, a few moments before the current instant, is always in a condition not too dissimilar from the current condition. This causes anomalies to be generally more difficult to detect than in other problems where LOF and COF methods are used as prior art. Before applying one or more of the aforementioned algorithms, it is possible to modify the data set, that is the working portion of the aforementioned algorithms, excluding a predefined number of data in earlier temporal instants and a predefined number of data in later temporal instants with respect to a predetermined instant of time. In this way, the data normal points will be even more embedded in a clustered region, while the anomalous points will be more isolated and possibly further away from the defined cluster region. In other words, this procedure allows the algorithm not to consider a set of strongly correlated data and therefore further decrease the density around anomalous data, thus increasing performance.

This selection is, in particular, possible thanks to the storage of data in the storage unit 211 during their measurement. The selection of the number of front and rear time instants to be excluded is a function of the design choices relating to the use of the related algorithms and can be carried out in a system setup phase 11.

The anomaly detection system 11, and the relative artificial pancreas 1, according to the present invention is therefore efficient and adaptable to the individual patient. In particular, the system 11 for the detection of malfunctions is of the model-free and unsupervised type, therefore it is possible to determine the anomalies on the patient data on which the medical device is applied and without the need for "faulty / not faulty" labeling ”To be supplied to the device and without the need to determine a mathematical model for the patient's normal response.

In particular, by means of the set of preferred variables that describe the state of the system at each instant of time, in accordance with the present invention (g (t), ICOB (t) and DCOB (t)) and which constitute a non-linear transformation of the signals acquired by the data processing unit 111, it is possible to describe the system more effectively for identifying the faults than is possible with the same signals in a single instant in time. Doing so greatly increases the possibility of highlighting anomalies with any type of algorithm and, in particular, with unsupervised model-free algorithms, for example the LOF, COF and iF algorithms described above.

Claims (9)

CLAIMS A system (11) for detecting malfunctions in insulin delivery devices of the type comprising at least one sensor (21) for monitoring a person's blood glucose and an insulin infusion pump (31), the system (11 ) including: - a data processing unit (111) adapted to receive glucose concentration measurements from said glucose monitoring sensor (21) and further adapted to receive measurements of the amount of insulin infused by said infusion pump (31); - a data entry unit (311) operationally connected to the data processing unit (111) and capable of receiving estimates from a user of the amount of carbohydrates consumed; - a storage unit (211) operationally connected to the data processing unit (111) and suitable for storing glucose concentration measurements, measurements of the amount of insulin infused, estimates of the amount of carbohydrates taken; in which, starting from the measurements stored in the storage unit (211), the data processing unit (111) is able to calculate, at a given instant of time, an estimate of the residual insulin present in the blood (IOB) , and an estimate of residual blood glucose (COB), the system (11) is characterized by the fact that the said data processing unit (111) is able to calculate the value of a DCOB variable (t) defined as: where <)> is the variable that describes the derivative of the glucose concentration over time, COB (t) is the variable that describes, over time, the estimate of the residual glucose present in the blood, α and β are two positive constants, in which the data processing unit (111) is further adapted to calculate a variable IOCB (t) defined as: where IOB (t) is the variable that describes, over time, the estimate of residual insulin present in the blood, γ and δ are two positive constants, and in which the data processing unit (111) is capable of executing at least one anomalous value detection algorithm to detect anomalous values within a set of patient data, in which each of said patient data, in an instant of time, comprises at least one triplet of values, said triplet of values being constituted by a value of the variable DCOB (t) at a given instant of time, a value of the variable ICOB (t) at said instant of time, and a measurement of the glucose concentration carried out by said sensor (21) at said instant of time, and in which said data processing unit (111) is capable of generating an alert signal if said algorithm detects at least one anomalous value within the patient data set. System (11) according to claim 1, wherein said outlier detection algorithm is of the unsupervised learning type. System (11) according to claim 2 wherein said algorithm for detecting anomalous values of the unsupervised learning type is the Local Outlier Factor (LOF). System (11) according to claim 2 wherein said algorithm for detecting anomalous values of the unsupervised learning type is the Connectivity-based Outlier Factor (COF). System (11) according to claim 2 wherein said algorithm for detecting anomalous values of the unsupervised learning type is the Isolation Forest (iForest). System (11) according to one or more of claims 2 to 5, wherein said data processing unit (111) is adapted to employ at least one anomalous value detection algorithm of the unsupervised learning type selected from Local Outlier Factor (LOF), the Connectivity-based Outlier Factor (COF) and the Isolation Forest (iForest) to the data of the said variable DCOB (t), to the data of the said ICOB variable (t) and to the data of the said measurement of the concentration of glucose, e wherein said data processing unit (111) is adapted to generate an alert signal if said anomalous value detection algorithm generates a result outside a predefined threshold value range. System (11) according to claim 6, wherein said data processing unit (111) is adapted to employ at least one outlier detection algorithm of the unsupervised learning type selected from the Local Outlier Factor (LOF), the Connectivity-based Outlier Factor (COF) and the Isolation Forest (iForest) to a working portion of the said data of the said DCOB variable (t), the data of the said ICOB variable (t) and the data of the said measure of the concentration of glucose, wherein said work portion is defined by excluding a predefined number of data in earlier time instants and a predefined number of data in later time instants with respect to a predetermined time instant. Insulin delivery device comprising an insulin infusion pump, a glucose sensor and a malfunction detection system (11) according to one or more of claims 1 to 7. 9. Device according to claim 8, characterized in that it is an artificial pancreas (1).