A system and method for improving rescue carbohydrate recommendations in diabetes management.

Abstract

This invention provides a control device, method, and program for determining the recommended amount of carbohydrates (RC). [Solution] In the system (10), the control device (30) includes an acquisition unit (32) configured to acquire user data, an outlier unit that determines whether or not the user data is considered an outlier, an RC unit (36) that estimates an RC recommendation using a mathematical function of the user data, and a transmission unit (38) that transmits the RC recommendation. The transmission unit (38) transmits the RC recommendation only if the last user data is not determined to be an outlier.

Description

[Technical Field]

[0001] This invention relates to the field of diabetes management, and more specifically to a system and method for improving the accuracy and relevance of carbohydrate rescue recommendations. [Background technology]

[0002] In the field of diabetes management, the use of automated insulin delivery devices and associated software applications is becoming increasingly widespread. These systems often rely on continuous glucose monitoring (CGM), providing real-time data on blood glucose levels, which is then used to calculate and administer the appropriate insulin dose. A particular challenge in this field is the management of hypoglycemia, a condition characterized by abnormally low blood glucose levels. To prevent hypoglycemia, individuals with diabetes may be advised to consume additional carbohydrates, often called rescue carbohydrates (RCs). RC recommendations are typically based on current and predicted blood glucose levels, as well as other factors such as insulin onboard (IOB), which refers to the amount of insulin administered and still available in the user's body.

[0003] However, the accuracy and relevance of RC recommendations can be affected by various factors, including the reliability of the CGM data and the algorithms used to interpret this data. For example, a CGM sensor can be affected by outliers, which are data points that deviate significantly from other observations in the dataset. Outliers may arise due to variability in measurements or may indicate errors. Outliers can occur for many reasons, including changes in instrument behavior or environmental factors.

[0004] Furthermore, blood glucose dynamics are complex and can be influenced by various physiological and lifestyle factors. For example, physical activity can increase insulin sensitivity and glucose uptake by muscles, potentially increasing the risk of exercise-induced hypoglycemia. Therefore, the ability to accurately predict future blood glucose levels and adjust RC recommendations accordingly is a challenging aspect of diabetes management.

[0005] Furthermore, RC-recommended calculations often involve the use of mathematical models and algorithms that take various parameters into account. However, the accuracy and reliability of these models can be affected by the precision of the input parameters and the inherent variability in the user's physiological response.

[0006] Despite advancements in diabetes management technologies, conventional techniques still have shortcomings in optimizing the accuracy and relevance of RC recommendations across various contexts. Existing systems often fail to effectively identify and handle outliers in CGM data, accurately predict future blood glucose levels, and dynamically adjust RC recommendations based on real-time data and individual user characteristics. Therefore, there is a substantial unmet need for systems and methods that can enhance hypoglycemia management through improved RC recommendations while maintaining a user-friendly design and compatibility with existing diabetes management infrastructure.

[0007] European Patent Application Publication No. 4268719 relates to an electronic device comprising a drug delivery unit for delivering a drug to a user, and a biological sensor for sensing one or more analytes or biological data of the user and generating biological sensor data indicating the sensed biological data. The electronic device further comprises a biosensor and / or a pressure sensor for sensing pressure on the user's biological tissue and generating pressure sensor data. However, such pressure sensors can malfunction, are heavy, and consume energy. For this reason, pressure sensors can be a burden for the user.

[0008] This system and method aim to address technical issues in the prior art by improving the relevance and accuracy of rescue carbohydrate (RC) alerts in various contexts.

[0009] Therefore, the present invention aims to address the above technical problems at least partially. [Overview of the Initiative] [Means for solving the problem]

[0010] This disclosure relates to a system and method for improving the relevance and accuracy of rescue carbohydrate (RC) alerts in the context of diabetes management. More specifically, the disclosed system and method aim to address the technical challenges associated with accurate prediction and recommendation of RC in various scenarios, such as during periods of physical activity, postprandial periods, and during sleep when compression artifacts may occur. The system and method disclosed herein determines RC recommendations using a combination of user data, including insulin infusion volume, ingested carbohydrates, and physiological values ​​such as measured blood glucose or interstitial blood glucose levels. The system further incorporates an outlier unit to identify and exclude abnormal blood glucose measurements, thereby improving the accuracy of RC recommendations. The system's RC unit estimates RC recommendations using a mathematical function, such as the slope, of measured blood or interstitial blood glucose levels, providing a dynamic assessment of the user's glucose tendency. This allows the system to predict potential hypoglycemic events and recommend timely and appropriate interventions. The system's ability to accurately estimate and recommend the appropriate amount of RC is a preventative measure to prevent severe hypoglycemia and ensure user safety. The ability of a system to transmit RC recommendations to another device enables users or healthcare providers to receive alerts in a timely manner and take appropriate action, thereby improving user safety and the effectiveness of diabetes management. Therefore, the disclosed system and method provide a comprehensive and personalized approach to managing RC alerts, addressing technical challenges in prior art, and improving the safety and effectiveness of diabetes management.

[0011] Therefore, the present invention is a control device for determining the amount recommended by RC, and the control device is An acquisition unit, configured to acquire user data, each piece of user data having a timestamp, the user data relating to a specific user, and the user data being at least, The amount of insulin injected into a specific user, The amount of carbohydrates consumed by a specific user, A search unit comprising: multiple physiological values ​​of a specific user, including at least measured blood glucose levels; An outlier unit configured to determine whether one of the measured blood glucose values ​​is considered an outlier, An RC unit, wherein the RC unit is configured to estimate RC recommendations using a mathematical function of the measured blood glucose level, A transmitting unit configured to send RC recommendations, including, The transmitting unit is configured to send an RC recommendation only if the last measured blood glucose level among the measured levels is not determined to be an outlier.

[0012] According to the present invention, "measured blood glucose level" means blood glucose level measured in the blood or blood glucose level measured in interstitial tissue.

[0013] The ability of the outlier unit to determine whether a measured blood glucose level is an outlier allows the system to ignore anomalous readings that could reduce the accuracy of RC recommendations.

[0014] The RC unit's ability to estimate RC recommendations using the slope of measured blood glucose levels enables a dynamic assessment of the user's glucose tendencies. This allows the system to predict potential hypoglycemic events and recommend timely and appropriate interventions, potentially reducing the risk of severe hypoglycemia, for example.

[0015] The conditional operation of the transmission unit that transmits the RC recommendation only when the last measured blood glucose value is not considered an outlier ensures that the recommendations provided to the user are based on accurate and reliable data. This selective transmission functions as a safeguard against providing potentially harmful advice based on incorrect readings.

[0016] In the context of the present disclosure, "carbohydrate" (RC) refers to the amount of carbohydrate intake recommended to the user to counteract predicted or occurring hypoglycemic events. The RC recommendation value is calculated to raise the user's blood glucose level to a safe range, thereby providing a rapid source of glucose that can be easily absorbed and utilized by the body. RC recommendations are particularly useful for individuals with diabetes who are at risk of hypoglycemia due to insulin therapy or other factors that affect blood glucose levels. The function of accurately estimating and recommending the appropriate amount of RC is a preventive measure to prevent severe hypoglycemia and ensure the safety of the user.

[0017] In the context of the present disclosure, an "outlier" refers to a data point that significantly deviates from other observed values within a dataset. Outliers may occur due to the variability of the measured values or may indicate errors, and the latter may be excluded from the dataset. Outliers can occur for many reasons, such as changes in device behavior, such as CGM compression, or environmental factors, when the measured blood glucose value is measured by a continuous glucose monitor (CGM). The ability to identify and process outliers is particularly useful in systems that provide recommendations or warnings based on accurate data, as it helps to ensure that such recommendations or warnings are based on representative and reliable data.

[0018] In certain embodiments, the control device comprises an alert unit configured to create an alert based on the RC recommendation. With such a configuration, the control device can warn the user, and thus reduce the health risk of hypoglycemia. The alert must be of any kind, for example, a visual alert, an audible alert, or a tactile alert.

[0019] In some aspects, the transmitting unit may be configured to transmit the RC recommendation to another device such as a mobile phone, a smartwatch, a dedicated receiver, or a healthcare provider's monitoring system. This transmission can be carried out via various communication protocols including, but not limited to, Bluetooth®, Wi-Fi, cellular networks, or Near Field Communication (NFC). The ability to transmit the RC recommendation to another device enables the user or the healthcare provider to receive the alert in a timely manner and take appropriate actions, thereby improving the user's safety and the effectiveness of diabetes management.

[0020] In some embodiments, the transmitting unit may be configured to send an alert to the user or a designated recipient when the RC recommendation is transmitted. The alert functions to notify the user or the recipient of the RC recommendation and may prompt the user or the recipient to take an action based on the provided recommendation. The alert can be in the form of a visual notification, an audible alarm, a vibration, or any combination thereof, depending on the user's preference and the capabilities of the receiving device. This feature enhances the responsiveness of the user or caregiver to potential hypoglycemic events and contributes to improved safety and preventive diabetes management.

[0021] According to the present invention, the measured blood glucose level is the blood glucose level measured for a specific user. The blood glucose level can be measured by any means, for example, a Continuous Glucose Monitor (CGM) or a Blood Glucose Monitor (BGM).

[0022] According to the present invention, the amount of carbohydrates ingested by a specific user corresponds, for example, to the amount of sugar ingested in a meal.

[0023] In some embodiments, the outlier unit may be configured to determine outliers using arbitrary Bayesian methods. Bayesian methods provide a probabilistic framework for updating beliefs about the state of the system based on new observations. This approach allows the outlier unit to incorporate prior knowledge about the expected behavior of blood glucose levels and continuously improve its understanding as new data becomes available. By utilizing Bayesian techniques, the outlier unit may be able to more accurately identify anomalous readings while taking into account the inherent variability of physiological measurements. This can lead to improved robustness in outlier detection across a wide range of scenarios and patient profiles.

[0024] According to one embodiment, the outlier unit is configured to determine outliers using a Kalman filter (KF).

[0025] An outlier unit configured to use KF to identify outliers allows the outlier unit to improve its accuracy in identifying inaccurately measured blood glucose levels by effectively filtering out statistical noise and inaccuracies inherent in the data. This results in a more reliable dataset for generating RC recommendations, which is particularly beneficial for users who require accurate and timely interventions to manage their blood glucose levels. The predictive power of KF allows for a continuous improvement in the system's understanding of the user's physiological state, thereby improving the overall accuracy and reliability of the RC recommendations provided by the control device.

[0026] In the context of this disclosure, a “Kalman filter” (KF) is an algorithm that uses a series of measurements observed over time, including statistical noise and other imperfections, to generate estimates of unknown variables that tend to be more accurate than those based on a single measurement or a single model alone.

[0027] According to one embodiment, KF operates in a two-step process. Firstly, a prediction step estimates the current state of the system, and secondly, an update step refines the estimate by incorporating the most recent measurements. KF is particularly advantageous in control devices for determining outliers in measured blood glucose levels because it can effectively remove noise, and more significantly, the analysis of KF residuals, which represent the prediction error relative to the observations, provides an indicator of agreement between the measurements and the model output. This analysis is particularly useful for detecting sensor failures because it enables the identification of discrepancies that may indicate outliers, thereby increasing the reliability of the dataset used to generate RC recommendations. For example, KF can predict the system state by integrating a personalized insulin-vs-glucose model, and then refine this prediction by comparing it with actual measured blood glucose levels, thereby identifying any discrepancies that may indicate outliers.

[0028] In some embodiments, the outlier unit may employ various derivations of the Kalman filter (KF) for detecting candidate outliers in measured blood glucose levels. These derivations may include, but are not limited to, the extended KF, unflavored KF, cubature KF, square root KF, information KF, ensemble KF, and particle filter. Each of these KF derivations provides a different approach to processing and analyzing data, some being more computationally intensive than others. For example, the particle filter may require the propagation of many particles through model equations to estimate a posterior state distribution that may be non-Gaussian, but requires fewer assumptions about the linearity of the system and the statistical properties of noise.

[0029] The outlier unit can further evaluate discrepancies between model predictions and actual measured blood glucose levels (i.e., CGM measurements) using various models, whether linear or nonlinear, physiological or empirical (black box, gray box). This discrepancy evaluation is a core function of the outlier unit, as it helps identify when measured blood glucose levels deviate from the expected pattern and can indicate faults or outliers in measured blood glucose levels. Depending on the hardware capabilities of the diabetes management system and specific requirements, more economical versions of the KF, such as a base KF coupled with a linear model, may be preferred. This approach focuses on the ability to capture sudden fluctuations in the discrepancy between the model and the measured values, which is of paramount importance for sensor fault detection. In contrast, more accurate state estimation, which may be desirable for model-based control applications such as calculating insulin from state predictions, potentially requires more complex KF derivations. Therefore, the selection of the KF derivation and model type is guided by a trade-off between computational efficiency and the desired level of accuracy and robustness in outlier detection.

[0030] According to one embodiment, to improve the accuracy of RC recommendations, the control unit is configured to perform a three-step procedure for managing outliers in measured blood glucose levels. First, the outlier unit is configured to analyze the residuals of blood glucose measurement levels using α(KF) to identify candidate outliers that deviate from the predicted values. These candidates are then subjected to a series of data-driven confirmation rules, which may include assessing the temporal context of the readings, the physiological validity of the glucose trend, and the consistency of the data with known user activity or events. Once an outlier is confirmed, the control strategy of the RC unit is improved accordingly. This includes adjusting the RC recommendation to take the identified outlier into account, ensuring that the system's response is based on the user's true blood glucose state. By systematically determining, confirming, and responding to outliers, the control unit provides a robust and reliable method for managing RC alerts, thereby improving the safety and effectiveness of diabetes management.

[0031] According to one embodiment, KF is configured to predict the system state by integrating a personalized sixth-order insulin-vs-glucose linear model from user settings, the model parameters including at least the following: Diet ratio (MR), and / or Insulin sensitivity factor (ISF), and / or Diffusion time constant (τ IOB The insulin diffusion rate represented by ) and / or Digestion time constant (τ D The rate of food digestion, expressed by ).

[0032] These features allow the control system to provide a tailored approach to diabetes management by incorporating individual user settings. This personalization ensures that RC recommendations are based on a model that closely reflects the user's specific physiological responses to insulin and carbohydrate intake, leading to more accurate and effective management of blood glucose levels. The use of a sixth-order linear model allows for a nuanced representation of insulin-glucose dynamics, which can explain complex interactions over time, thereby improving the system's predictive accuracy. This can be particularly advantageous in preventing hypoglycemic events, as the system can more accurately predict blood glucose trajectories and propose timely and appropriate RC interventions.

[0033] In the context of this disclosure, the “sixth-order insulin-vs-glucose linear model” refers to a mathematical representation that characterizes the relationship between insulin administration and blood glucose levels over time using a linear state-space model that approximates actual system dynamics. This model is described by a set of differential equations that express the dynamics of insulin action and glucose absorption up to the sixth-order derivative with respect to time. The sixth-order model aims to capture the first-order dynamics of the insulin-vs-glucose interaction and provides a balance between model complexity and computational efficiency, which may be particularly useful for real-time applications in insulin delivery systems.

[0034] As mentioned above, the use of a sixth-order linear model allows for a nuanced representation of the dynamics of insulin versus glucose, which can explain complex interactions over time and thereby improve the predictive accuracy of the system.

[0035] In the context of this disclosure, “Diet Ratio” (MR) refers to a parameter used to personalize an insulin-versus-glucose model, representing the ratio of the amount of carbohydrates in a diet to the amount of insulin administered to metabolize those carbohydrates. MR is a factor in determining the insulin dose relative to a given carbohydrate intake and is used to adjust model parameters to reflect the user’s individual dietary habits and insulin sensitivity. MR can be set by the user, the healthcare provider, or determined based on historical data and may be adjusted over time to optimize glycemic control.

[0036] In the context of this disclosure, “insulin sensitivity factor” (ISF) refers to a parameter that quantifies the relationship between insulin administration and its effect on lowering blood glucose levels. ISFs may be used to adjust insulin dosages to achieve a desired glycemic target and are typically personalized based on the user’s insulin sensitivity and metabolic response. ISFs may be determined through clinical assessment or derived from historical data of the user’s insulin use and corresponding glycemic response.

[0037] In the context of this disclosure, "diffusion time constant" (τ IOB (τ) refers to a parameter that represents the rate at which insulin diffuses into the bloodstream and begins to lower blood glucose levels after administration. This constant is an integral part of the insulin-vs-glucose model because it helps predict the time-dependent effect of insulin on blood glucose concentration. The diffusion time constant can be personalized based on the user's physiological characteristics and can vary depending on various factors. IOBBy incorporating

[0038] In the context of the present disclosure, the "digestion time constant" (τ D ) refers to a parameter that characterizes the rate at which ingested carbohydrates are digested, absorbed into the bloodstream as glucose, and thus appear in the measured blood glucose level. This constant is a component of the insulin-to-glucose model that affects the timing and magnitude of the postprandial glucose response. The digestion time constant can be personalized to reflect an individual's digestive dynamics, which can be affected by factors such as diet composition, gastrointestinal motility, and metabolic health. By incorporating τ D into the model, the control device can more accurately predict the rise in postprandial blood glucose level, which is particularly useful for determining the timing and amount of insulin administration or the RC recommendation to maintain blood glucose control.

[0039] In the context of the present disclosure, model parameters including at least MR and / or ISF and / or τ IOB and / or τ D mean that the model includes the parameter or the derivative of the parameter.

[0040] The meal ratio (MR), insulin sensitivity factor (ISF), insulin diffusion rate (τ IOB ), and meal digestion rate (τ DThe synergistic integration of these parameters facilitates comprehensive and individualized management of blood glucose levels. By combining these parameters, the control system can more accurately simulate the complex interactions between carbohydrate intake, insulin administration, and their respective effects and absorption rates. This synergistic effect allows for improved models that reflect the user's specific metabolic patterns, increasing the accuracy of RC recommendations. The combined effect of these parameters ensures that RC recommendations are not only based on static measurements but are also dynamically adjusted to the user's individualized physiological responses, leading to a more effective and preventive approach to preventing hypoglycemic events.

[0041] According to one embodiment, the outlier unit is defined as having a distance (d) including the innovation matrix and KF residuals that is a dynamic threshold (d THR If the value is smaller than ), the system is further configured to consider the last measured blood glucose value among the measured values ​​as an outlier.

[0042] Such characteristics allow the control device to dynamically adjust the sensitivity of outlier detection, which is particularly beneficial in scenarios where blood glucose levels can fluctuate rapidly, such as during physical activity or after meal intake. A dynamic threshold (d) adapts to the user's current physiological state. THR By setting the appropriate parameters, the system can more accurately distinguish between true outliers and legitimate rapid changes in blood glucose levels. This results in a more robust and reliable RC recommendation process, minimizing the risk of false alarms and ensuring that interventions are based on the user's true blood glucose state. The ability to fine-tune outlier detection in real time enhances the system's usefulness across a wide range of situations and provides users with greater confidence in the RC recommendations they receive.

[0043] In the context of this disclosure, “d” refers to a metric used to quantify the discrepancy between the predicted state of the system and the actual observed measurement. In other words, d is a measure of how the KF prediction error deviates from the expected prediction error. Thus, the innovation matrix is ​​used to account for the normal variability of the prediction error. A significant distance d means that the KF is abnormally wrong in its prediction, which may be an indication of sensor failure. Specifically, the Mahalanobis distance is a multivariate distance measure that describes the correlation between variables and scales the distance calculation according to the variability of each variable. This distance is calculated from the innovation matrix, which represents the variability of the (KF) residual between the predicted and observed measurement, and the KF residual, which is the discrepancy between the actual measurement and the KF prediction. The calculated distance (d) is compared with a dynamically determined threshold (d). THR If the value is smaller than ), the last measured blood glucose level is considered an outlier. This approach enables a more sophisticated outlier detection mechanism that is sensitive to the inherent variability and correlation structure of the data, resulting in a more accurate and reliable blood glucose assessment for the purpose of RC recommendations.

[0044] In some embodiments, a dynamic threshold (d) is used for outlier detection. THR The threshold can be calculated based on the moving average or median of d signal values ​​acquired over a past time window. This approach allows the threshold to adapt to changes in the user's physiological state over time and provides a baseline that reflects the user's recent blood glucose trends. To further improve the accuracy of outlier detection, the moving average or median calculation can be trimmed to discard extreme values ​​such as peaks that may contain the sought outliers, thereby capturing a more representative baseline of d.

[0045] Additional embodiments include d THR This may include applying an additive offset, which can provide a safety margin to account for measurement uncertainty or expected physiological variability. Alternatively, the multiplicative scaling factor may be d THRThis may also be applied to allow the threshold to scale proportionally to the variability observed at d. This scaling is particularly useful in situations where measured blood glucose levels show greater variability, such as during periods of intense physical activity or stress, and ensures that false positives are avoided while outlier detection remains sensitive to true outliers.

[0046] In some embodiments, the outlier unit can use a statistical chi-squared test to assess the goodness of fit between the KF residuals and a theoretical distribution, such as a Gaussian distribution. This test is particularly useful in determining whether the observed residuals from KF deviate in a statistically significant way from what would be expected if the system correctly models the underlying physiological processes. A substantial deviation, such as one indicated by the chi-squared test result exceeding a predetermined threshold, may suggest that the measured blood glucose level is an outlier. This deviation is defined as reading a value larger than the expected value of the residual distribution, or the statistical mean, than would typically be observed in the absence of an outlier in the measured blood glucose level. Such a finding indicates a potential flaw in the measured blood glucose level or an anomaly in the glucose measurement process, prompting the outlier unit to potentially exclude the affected data point from the RC recommendation process.

[0047] In the context of this disclosure, “statistical chi-squared test” is a statistical method used to compare observed data with data that would be expected to be obtained according to a particular hypothesis.

[0048] This test calculates a chi-squared statistic, which is a measure of the discrepancy between the observed frequency and the expected frequency in one or more categories. This statistic is compared to a chi-squared distribution to determine the likelihood of a randomly occurring observed distribution. If the calculated chi-squared statistic exceeds the value of the chi-squared distribution for a given level of statistical confidence, it suggests that the observed data does not fit the expected distribution and indicates the presence of outliers or deviations from the assumed pattern.

[0049] In some embodiments, machine learning techniques can be used to enhance outlier detection based on residuals obtained from Kalman filters. These techniques involve training a machine learning model on historical data to classify normal and abnormal behavior. Algorithms such as support vector machines (SVMs), random forests, or neural networks can be used for this purpose. SVMs offer effectiveness in high-dimensional spaces and the ability to model nonlinear boundaries with kernel trick features. Random forests, consisting of multiple decision trees, offer robustness against overfitting and provide a measure of feature importance. Neural networks, particularly deep learning models, can learn complex patterns and relationships in data through their hierarchical structures.

[0050] The use of these machine learning techniques in the context of classification and clustering enables comprehensive analysis of data, with each technique potentially capturing different forms or patterns that indicate outliers. By combining these techniques, or by running them simultaneously, the system can leverage a consensus approach to increase the reliability of outlier detection. For example, a voting mechanism can be implemented, and if the majority of machine learning algorithms determine the presence of an outlier, the system can consider bypassing RC recommendations.

[0051] Alternatively, a safer mode might require all algorithms to converge on outlier detection before taking action. This multi-algorithmic approach can improve the system's robustness against false positives and false negatives, ensuring that RC recommendations are based on accurate classification of blood glucose measurements.

[0052] According to one embodiment, the outlier unit is further configured to identify outliers, particularly during the postprandial period.

[0053] These characteristics allow the control device to specifically target the postprandial period, when blood glucose levels are particularly likely to fluctuate upward following meal intake. By focusing on this period, the outlier unit can more effectively identify readings (i.e., measured blood glucose levels) that do not fit the expected postprandial blood glucose pattern, thereby improving the accuracy of RC recommendations. This targeted approach to outlier detection is particularly beneficial for users who may experience unpredictable postprandial blood glucose responses, as it provides a more refined analysis of blood glucose data during the time most relevant for accurate RC recommendations to prevent postprandial hypoglycemia.

[0054] In some cases, during the post-meal period, particularly around the time of meal reporting, the control device may adjust the predictive rules for forecasting RC recommendations. This adjustment is based on the expectation that blood glucose will rise due to the meal declared as consumed. Therefore, the system may relax the normal predictive rules to avoid raising RC alerts, which in some cases may unnecessarily confuse the user. This relaxation ensures that an RC alert is issued only when the user's blood glucose (i.e., measured blood glucose level) is maintained and truly guaranteed to fall below the level desired for the RC trigger. This approach helps prevent overcorrection of blood glucose levels, which can be particularly destructive after a meal when the body is naturally working to assimilate ingested carbohydrates and stabilize blood glucose.

[0055] According to one embodiment, the RC unit is further configured to estimate an RC recommendation if the most recent measured blood glucose level among the measured blood glucose levels indicates a slowdown.

[0056] Such characteristics enable the control device to provide RC recommendations that not only respond to current blood glucose levels but also predict future trends. By analyzing the deceleration of blood glucose levels, the RC unit can identify when the rate of decrease is slowing down, which can indicate an impending stabilization or a potential reversal of a downward trend. This proactive function can be particularly advantageous in preventing hypoglycemia because it allows for early intervention by the RC while blood glucose levels are still within a safe range, rather than waiting for blood glucose levels to reach a level that already indicates hypoglycemia. This proactive approach can improve the user's quality of life by reducing the frequency and severity of hypoglycemic episodes and can also minimize the user's cognitive burden by reducing the number of decisions the user has to make regarding glucose management.

[0057] According to one embodiment, the control device includes a prediction unit, which is configured to calculate at least a predicted blood glucose level based at least on user data.

[0058] These features allow the control device to enhance diabetes management by predicting the user's future blood glucose levels. The ability of the prediction unit to calculate predicted blood glucose based on user data enables the system to predict potential hypoglycemic events before they occur, allowing for proactive measures. This predictive capability provides users and healthcare providers with valuable information for making decisions based on carbohydrate intake and insulin medication, potentially leading to improved blood glucose control. Furthermore, the use of user data by the prediction unit for predictive purposes ensures that predictions are personalized and tailored to individual metabolic responses, further improving the accuracy and relevance of RC recommendations. This can result in a reduction in hypoglycemic episodes, improved patient safety, and an overall improvement in the quality of life for individuals with diabetes.

[0059] In the context of this disclosure, “Predictive Unit” refers to a component within a control device responsible for calculating future physiological states, such as predicted blood glucose levels, based on current and historical user data. The Predictive Unit utilizes various mathematical models, algorithms, and data processing techniques to analyze trends and patterns in the user’s physiological data, thereby enabling the system to predict future changes in blood glucose levels. For example, the Predictive Unit may predict blood glucose levels using a linear regression model based on recent measurements and known rates of glucose absorption and insulin action. In another embodiment, the Predictive Unit may use machine learning algorithms, such as neural networks or support vector machines, to analyze complex datasets and identify subtle correlations that can improve the accuracy of its predictions. Furthermore, the Predictive Unit may incorporate time-series analysis to detect cyclical patterns in blood glucose fluctuations that may be influenced by daily activities, meal times, and sleep cycles. These examples demonstrate the versatility of the Predictive Unit in adapting to the individualized characteristics of the user’s metabolic responses and lifestyle factors, thereby enhancing the personalized management of diabetes. In some embodiments, the Predictive Unit may also employ a Model Predictive Control (MPC) strategy to enhance the prediction of future physiological states. A predictive control (MPC) is a form of control algorithm that involves creating a model of the system to predict future outcomes and adjusting control inputs accordingly. In the context of blood glucose management, MPCs can be particularly useful for taking into account the delayed effects of insulin delivery and carbohydrate absorption on blood glucose levels. By considering the range of possible future scenarios and the possible effects of different insulin administration strategies, a predictive unit can optimize blood glucose control over a specified predictive range, thereby contributing to more stable and accurate diabetes management.

[0060] According to one embodiment, the RC unit is configured to estimate an RC recommendation when the predicted blood glucose is lower than the predicted blood glucose threshold (θ1), the current blood glucose is lower than the current blood glucose threshold (θ2), and the insulin onboard (IOB) is higher than the IOB threshold (θ3).

[0061] These features allow the control device to provide a more nuanced and responsive approach to managing blood glucose levels, particularly in the context of insulin therapy. By incorporating multiple thresholds that take into account predicted blood glucose, current blood glucose, and IOB, the RC unit can make more informed decisions about when to recommend carbohydrate intake. This multifaceted analysis ensures that RC recommendations are made at the appropriate time and reduces the risk of over- or under-treatment of hypoglycemia. The integration of these thresholds enables more individualized and precise management of diabetes, which can lead to better overall glycemic control and a reduced incidence of hypoglycemic events. This proactive strategy can improve patient outcomes and reduce long-term complications associated with diabetes.

[0062] According to one embodiment, the RC unit is configured to dynamically adjust the predicted blood glucose threshold (θ1), the current blood glucose threshold (θ2), and the IOB threshold (θ3) based on real-time data and the user's current physiological state. For example, the thresholds may be adjusted to take into account factors such as physical activity, stress, illness, or changes in medication. This dynamic adjustment allows the control device to provide RC recommendations tailored to the user's current situation, thereby improving the safety and effectiveness of insulin therapy.

[0063] Furthermore, the RC unit may incorporate safety features that prevent RC recommendations from being issued when they are not due to safety or user interference reasons. For example, if the predicted blood glucose level exceeds the predicted blood glucose threshold (θ1) and the current blood glucose level exceeds the current blood glucose threshold (θ2), the RC unit may withhold an RC recommendation even if the IOB is greater than the IOB threshold (θ3). This safety mechanism ensures that RC recommendations are made cautiously, minimizing the risk of unnecessary carbohydrate intake and potential subsequent hyperglycemia, in accordance with the user's actual blood glucose status.

[0064] In the context of this disclosure, “Insulin Onboard” (IOB) refers to the amount of insulin administered that is still available in the user’s body. IOB is a dynamic value that reflects insulin that has not yet exerted its full glucose-lowering effect. Estimating IOB involves calculating the remaining available insulin from a previous dose, taking into account the time since administration and the known pharmacokinetics of the insulin preparation used. IOB can be influenced by factors such as the type of insulin, the dosage, and the individual’s insulin sensitivity. For example, rapid-acting insulin has a different IOB profile compared to long-acting insulin. IOB is an integral part of the decision-making process in insulin management because it helps prevent insulin buildup and the risk of hypoglycemia. In some embodiments, the control device is configured to estimate IOB by utilizing a decay function that models the progressive absorption effect of insulin over time. This estimation allows the RC unit to adjust RC recommendations by taking into account the residual effect of previously administered insulin on current and predicted blood glucose levels.

[0065] According to one embodiment, one of the measured blood glucose values ​​is determined to be an outlier if it is determined that it is affected by compression artifacts, as determined by an outlier unit.

[0066] Such features allow the control device to improve the reliability of RC recommendations by identifying and excluding glucose measurements that may be distorted due to physical interference with the glucose monitoring device, such as a CGM. When the sensor is subjected to pressure, compression artifacts can occur, leading to erroneous low glucose measurements that may trigger unnecessary RC recommendations. By recognizing these artifacts, the control device can prevent inappropriate responses to these inaccurate readings, thereby maintaining the accuracy of RC recommendations and avoiding potential overtreatment of hypoglycemia. This feature is particularly beneficial for users, for example, who are active or sleeping on the monitoring device, as it ensures that RC recommendations are based on true glucose values ​​rather than artifacts from the measurement process.

[0067] In the context of this disclosure, “compression artifact” refers to a type of error in glucose readings that occurs when an external force is applied to a glucose monitoring device, typically a continuous glucose monitor (CGM), resulting in a transient deformation of the sensor or surrounding tissue. This deformation can cause interstitial fluid dynamics or disruption of sensor function, potentially leading to transient, artificial changes in measured blood glucose levels that do not accurately reflect the true blood glucose concentration. Compression artifacts are often characterized by a sudden, non-physiological drop in recorded blood glucose levels, followed by a return to a more typical reading after the pressure is released. These artifacts can be particularly problematic during periods of rest or sleep when the user is lying on the sensor, or during strenuous physical activity where the device may be compressed against the body. The ability to accurately identify and exclude measurements affected by compression artifacts is an essential part of ensuring the reliability of glucose monitoring systems and user safety, as it helps prevent inappropriate treatment decisions based on erroneous data.

[0068] In some embodiments, the control device may utilize confirmation rules to verify that a candidate KF outlier is indeed a non-physiological compression artifact. The confirmation process may include nocturnal conditions, which focus on detecting nocturnal compression artifacts to reduce the risk of false positives that may be increased by daily activity. The outlier unit may further be configured to take into account the time distance to past declared, estimated, or measured physical activity and carbohydrate intake, as these factors may influence blood glucose levels and the likelihood of compression artifacts.

[0069] The downward residual condition may apply if the negative residual value is a negative value that confirms the lowest point trend during the out-of-range period. This condition is particularly relevant at night as it enhances detection safety by reducing the possibility of false positives. However, this condition may be relaxed in other applications where nighttime detection is not appropriate.

[0070] According to the present invention, "lowest point trend" refers to the pattern or trajectory of measured blood glucose levels reaching those lowest points over a given period of time.

[0071] The outlier unit may be further configured to evaluate the first derivative condition, where the current slope of the blood glucose trend is compared to a predetermined negative threshold. If the slope is smaller than this threshold, a substantial and abnormal decrease in blood glucose is observed, indicating a potential compression artifact. Furthermore, if a sudden discontinuity in the slope compared to the predetermined threshold demonstrates a non-physiological change in the trend, the second derivative condition may be evaluated.

[0072] An alternative embodiment of the second derivative condition may include calculating the ratio of two consecutive first derivative samples and comparing it to a predetermined threshold. This approach allows for the detection of rapid changes in the rate of blood glucose reduction, which is characteristic of compression artifacts.

[0073] Furthermore, a net IOB condition may be implemented, and the net IOB value is expected to be less than zero to enhance the safety of the detection process. This ensures that a rapid drop in measured blood glucose levels cannot be attributed to the effects of past insulin infusions. By incorporating these confirmation rules, the control device can more accurately identify and eliminate compression artifacts, thereby improving the reliability of RC recommendations and ensuring that interventions are based on accurate blood glucose measurements.

[0074] In the context of this disclosure, “Net Insulin Onboard” or “Net IOB” refers to the amount of insulin administered and still available in the user’s body, adjusted for the basal insulin rate. Net IOB is calculated by subtracting from total IOB the amount of insulin that would have been delivered as part of the user’s basal rate. This calculation takes into account insulin injected to neutralize ingested carbohydrates and insulin administered to correct high blood glucose levels. Since Net IOB takes into account insulin that effectively exceeds the basal insulin requirement, it provides a more accurate reflection of the potential effect of insulin on lowering blood glucose levels at any given moment.

[0075] In some embodiments, the outlier unit of the control device is configured to analyze the relationships between various features representing a feature space, such as measured blood glucose derivatives, net IOB, carbohydrate onboard (COB), KF residuals, and additional sensor data such as accelerometer readings or heart rate, using a support vector machine (SVM). The SVM allows for the creation of a hyperplane that effectively separates the feature space into two distinct categories: out-of-range periods and physiological periods. This separation helps distinguish between normal physiological variability and potential outliers that may indicate defects in measured blood glucose levels or non-physiological events affecting glucose measurements.

[0076] In the context of this disclosure, “hyperplane” is a geometric concept that generalizes the concept of a plane to higher dimensions. Specifically, in an n-dimensional feature space, the hyperplane is a flat affine subspace of dimension n-1, which means it has one less dimension than the feature space itself. For example, in a 3-dimensional space, the hyperplane is a 2-dimensional plane. In the context of machine learning, particularly in the use of support vector machines (SVMs), hyperplanes are used to separate different classes of data points by finding a plane that maximizes the margin between classes. A hyperplane is defined by a set of weights and biases, which are optimized during SVM training to create a decision boundary that can classify new data points based on their position relative to the hyperplane.

[0077] In another embodiment, the outlier unit may utilize a trained random forest method to define a threshold surface in the feature space. The random forest method, consisting of an ensemble of decision trees, can classify measured blood glucose levels as either outliers or non-outliers based on the aforementioned features. This method is particularly advantageous due to its ability to handle high-dimensional data and its robustness against overfitting, which is beneficial for the accurate detection of outliers in measured blood glucose levels.

[0078] Furthermore, in some cases, the outlier unit may implement a neural network, including a convolutional neural network (CNN) or a recurrent neural network (RNN), for outlier classification. These types of neural networks are adept at capturing complex patterns and time dependencies in the data, which can be particularly useful for identifying subtle or transient anomalies in measured blood glucose levels that may not be readily apparent through other means.

[0079] Furthermore, the outlier unit can also incorporate Gaussian process regression (GPR) to model the relationship between input features and output classifications. GPR provides not only predictions for classification but also uncertainty estimates, which are of great value when defining threshold surfaces with confidence intervals. These uncertainty estimates can enhance the decision-making process by providing a probabilistic assessment of the data, thereby enabling more informed and careful interventions in blood glucose management.

[0080] According to one embodiment, if one of the measured blood glucose levels is considered an outlier, the transmitting unit is configured not to transmit an RC recommendation until a bypass period has elapsed. This bypass period is maintained until at least one of the measured blood glucose levels recovers and is no longer considered an outlier. The bypass period ends based on two conditions: either the time has reached or exceeded the time obtained by adding a predetermined rise time to the lowest point time, or the measured blood glucose level has risen above a certain level.

[0081] According to one embodiment, the bypass period is designed with safety exceptions to ensure the well-being of a specific user. If the most recently measured blood glucose level falls below a safety threshold, such as 55 mg / dL, or if the bypass period extends beyond its maximum duration, the transmitting unit is configured to resend the RC recommendation.

[0082] This configuration allows the control device to prevent the risk of measured blood glucose levels becoming outliers and failing to detect actual hypoglycemia.

[0083] According to one embodiment, the control device is configured to evaluate the blood glucose prediction up to a predetermined prediction range before the bypass period. If this blood glucose prediction falls below a predetermined hypoglycemic safety threshold, the bypass period is shortened to zero. This is because it indicates a true risk of hypoglycemia. With such a configuration, the control device can prevent the risk of actually experiencing hypoglycemia. For the purpose of performing these evaluations, the control device may be configured to use any suitable prediction means, such as a linear model, a nonlinear model, or a machine learning algorithm. The selection of the prediction means can be tailored to the specific requirements of the diabetes management system, taking into account factors such as computational efficiency, prediction accuracy, and the individual physiological response patterns of the user.

[0084] According to one embodiment, the outlier unit is configured to determine that the measured blood glucose level is affected by compression artifacts, based at least on deviations from the predicted blood glucose level and the measured blood glucose level.

[0085] This configuration allows the control device to distinguish between true fluctuations in blood glucose levels and fluctuations affected by compression artifacts. By comparing the deviation between predicted blood glucose and actually measured blood glucose levels, the outlier unit can effectively identify when readings are likely to be impaired. This ensures that RC recommendations are based on accurate and reliable data, especially in situations where the CGM sensor may be under pressure, such as during sleep or physical activity. The ability to filter and remove these specific types of outliers prevents unnecessary or inaccurate recommendations for carbohydrate intake, thereby optimizing the user's blood glucose management and reducing the risk of hypoglycemia. This feature is particularly advantageous for maintaining the integrity of the RC recommendation process and ensuring that the user receives warnings and interventions that accurately reflect their physiological state.

[0086] According to this disclosure, the outlier unit of the control device is configured to utilize a predictive model that takes into account both predicted and actual measured blood glucose levels. The predictive model can predict expected blood glucose levels based on a variety of factors, including but not limited to historical blood glucose data, the user's recent blood glucose trends, and the user's physiological parameters, using a variety of algorithms and techniques. When the outlier unit detects a deviation between the predicted blood glucose and the actually measured blood glucose level that exceeds a predetermined threshold, it determines this to be an indicator of a potential pressure artifact. This threshold can be dynamically adjusted based on the user's current state and the variability of the user's blood glucose measurements. By setting this threshold, the control device can distinguish between normal physiological fluctuations in blood glucose levels and readings that are likely to be affected by external pressure on the glucose monitor. This feature is particularly useful in scenarios where the user is engaged in activities that may exert pressure on the sensor, such as sleep or exercise. The ability to accurately identify and ignore these affected measurements helps maintain the integrity of the RC recommendation process and ensures that the user is provided with reliable and actionable information for managing their diabetes.

[0087] According to one embodiment, the acquisition unit is further configured to acquire user data, including the user's physical activity data, and the RC unit is configured to adjust RC recommendations based on the physical activity data.

[0088] Such a configuration allows the control device to consider the impact of physical activity on blood glucose levels, a dynamic factor in diabetes management. By including physical activity data in the RC recommendation process, the system can provide more personalized, context-aware recommendations. By tailoring RC recommendations based on the user's physical activity, the system can better predict the physiological effects of exercise, which often include increased insulin sensitivity and glucose uptake by muscles, potentially increasing the risk of exercise-induced hypoglycemia. This feature enhances the system's ability to tailor RC recommendations to the user's lifestyle, promoting more effective and safer diabetes management. Furthermore, it helps users maintain an active lifestyle by providing them with confidence that their blood glucose levels are being monitored and managed in consideration of their physical activity.

[0089] According to one embodiment, the RC unit is configured to calculate an RC recommendation based on the distance between a blood glucose target and a linearly predicted blood glucose level for a given predicted horizon, the distance being converted to carbohydrate units via a glycation ratio.

[0090] Such a configuration allows the control device to provide accurate and personalized RC recommendations by quantifying the precise amount of carbohydrates most likely to bring the user's blood glucose levels back into the desired target range. By calculating the distance between the blood glucose target and the linearly predicted blood glucose, the RC unit can determine the appropriate size of intervention for a given situation. By converting this distance into carbohydrate units via the glycation ratio, the process is simplified for the user, and complex data is transformed into readily usable and understandable guidance. This approach not only helps in the immediate correction of hypoglycemia but also contributes to the long-term stability of blood glucose levels by providing tailored recommendations that are directly aligned with the user's physiological requirements. The ability to predict blood glucose and translate it into tangible RC recommendations allows the user to manage their condition with greater confidence and accuracy, potentially reducing the frequency of hypoglycemic episodes and improving overall quality of life.

[0091] In the context of this disclosure, the “glycation ratio” is calculated based on the relationship between the amount of carbohydrates (CHO) expected to raise blood glucose levels to a target range and the expected increase in blood glucose levels over a given predicted range. The glycation ratio is estimated using a formula that takes into account the user’s body weight (BW) and a standard absorption coefficient typically derived from empirical data. For example, the glycation ratio can be estimated from the formula sugRatio=20BW / (0.680) in [gCHO / g / L], where BW is body weight in kilograms. This ratio is then used to convert the distance between predicted blood glucose and the blood glucose target into carbohydrate units, which represents the recommended amount of carbohydrates to be consumed to correct or prevent hypoglycemia. Thus, the glycation ratio functions as a conversion factor that translates expected blood glucose fluctuations into equivalent amounts of carbohydrates, facilitating the calculation of personalized RC recommendations.

[0092] According to one embodiment, the RC unit is further configured to adjust the RC recommendation based on at least IOB such that the adjusted RC recommendation is an increasing function of IOB.

[0093] Such a configuration allows the control unit to dynamically adjust RC recommendations according to the amount of insulin still available in the user's body, known as insulin onboard (IOB). By taking IOB into account, the RC unit can more accurately determine the amount of carbohydrates the user may need to prevent or treat hypoglycemia. This is particularly advantageous as it accounts for the ongoing glucose-lowering effect of insulin, which can vary depending on factors such as the type of insulin used, the time elapsed since administration, and the user's insulin sensitivity. The ability to adjust RC recommendations as an increasing function of IOB ensures that the user is not advised to consume more carbohydrates than they actually need based on their current physiological state, thereby avoiding potential hyperglycemia. This feature increases the accuracy of RC recommendations, leading to more effective and safer glycemic management, and ultimately contributing to the user's overall health and metabolic stability.

[0094] At a minimum, IOB-based RC recommendations correspond to net IOB-based recommendations and ensure that RC units consider dynamic insulin levels when determining the appropriate amount of rescue carbohydrates recommended to prevent or treat hypoglycemia. This approach enables more personalized and accurate RC recommendations, as it aligns the recommendations with the user's current physiological insulin profile.

[0095] In the context of this disclosure, “basal insulin rate” refers to the insulin administered to a user to maintain normal blood glucose levels in the absence of food. This rate is typically set based on the user’s basal metabolic insulin requirement, which is the fasting or intermeal insulin requirement. The basal insulin rate is designed to mimic the steady-state release of insulin by a healthy pancreas and is often delivered by an insulin pump or through long-acting insulin injections. The basal insulin rate can be adjusted based on various factors such as the user’s daily life, physical activity, stress levels, or changes in other physiological conditions that may affect insulin sensitivity. It is a fundamental component of insulin therapy for individuals with diabetes, particularly those receiving, for example, intensive insulin therapy. The basal insulin rate is distinct from bolus insulin administered to manage the rise in blood glucose levels due to carbohydrate intake from food.

[0096] The present invention further relates to a method for determining the amount recommended by RC, the method being carried out by the control device described above, The steps include acquiring user data and determining RC recommendations using at least a mathematical function of the measured blood glucose level, This includes the step of submitting an RC recommendation, The step of sending an RC recommendation is performed only if the last measured blood glucose level among the measured levels is not determined to be an outlier.

[0097] The embodiments, technical effects, and definitions disclosed herein with respect to the control device are also applicable to the method described herein. This method encompasses steps that fully utilize the functions and features of the control device described herein. Therefore, all embodiments, technical effects, and definitions relating to the device, including but not limited to changes in ISF over time and methods for calculating ISF, are equally applicable to the method. This ensures a comprehensive and unified understanding of both the embodiments of the control device and the method of the present invention, facilitating the implementation and use of the disclosed technology across a variety of applications.

[0098] The present invention also relates to a computer program that includes instructions for causing the control device described above to perform the steps of the method described above.

[0099] Furthermore, the disclosed systems and methods may be adapted for use in other applications within the medical industry where accurate and reliable prediction of physiological events and appropriate response recommendations are desired.

[0100] The present invention further relates to a control system comprising an insulin infusion device, a continuous glucose monitor (CGM), and a control device equipped with a recommendation unit and an acquisition unit.

[0101] While exemplary embodiments of the present invention have been described, those skilled in the art will understand that various modifications, omissions, and / or additions can be made without departing from the spirit and scope of the invention, and that elements can be replaced with equivalents. Furthermore, many modifications can be made without departing from the scope of the invention to adapt specific situations or materials to the teachings of the invention. Thus, the present invention is not limited to the specific embodiments disclosed for the purpose of carrying out the invention, and the present invention is intended to include all embodiments that fall within the scope of the appended claims. Furthermore, unless otherwise specified, the use of terms such as first, second, etc. does not indicate any order or importance, but rather the terms first, second, etc. are used to distinguish one element from another. Embodiments of the present invention are described below with reference to the drawings and are briefly described below. [Brief explanation of the drawing]

[0102] [Figure 1] The system shown is an embodiment of one of the following. [Figure 2] A method for determining the amount of rescue carbohydrates according to one embodiment of the present invention is shown. [Modes for carrying out the invention]

[0103] Figure 1 shows a system 10 comprising an insulin infusion device 20, a continuous glucose monitor (CGM) 12, and a control device 30.

[0104] The control device 30 is configured to determine the recommended amount of carbohydrates (RC) for a specific user and has an acquisition unit 32. The acquisition unit 32 is configured to acquire user data, each data point having a timestamp, and the user data is associated with a specific user. The user data includes the amount of insulin injected into the specific user, the amount of carbohydrates ingested by the specific user, and several physiological values ​​of the specific user, the several physiological values ​​of the specific user including at least measured blood glucose levels. Measured blood glucose levels are measured by a CGM 12.

[0105] In the context of this disclosure, “carbohydrate” (RC) refers to the amount of carbohydrate intake recommended to the user to counteract anticipated or occurring hypoglycemic events. RC may also be called a rescue curve. The RC recommendation is calculated to raise the user’s blood glucose level to a safe range, thereby providing a rapid source of glucose that can be readily absorbed and utilized by the body. RC recommendations are particularly useful for individuals with diabetes who are at risk of hypoglycemia due to insulin therapy or other factors affecting blood glucose levels. The ability of the control device 30 to accurately estimate and recommend an appropriate amount of RC is a preventative measure to prevent severe hypoglycemia and ensure user safety.

[0106] According to the present invention, the measured blood glucose level is the blood glucose level measured for a specific user. The blood glucose level can be measured by any means, such as a continuous glucose monitor (CGM) or a blood glucose monitor (BGM) as shown in Figure 1. The amount of carbohydrates ingested by the specific user corresponds, for example, to the amount of sugar ingested in a meal.

[0107] The acquisition unit 32 is further configured to acquire user data, including the user's physical activity data.

[0108] The control device 30 also includes an outlier unit 34. The outlier unit 34 is configured to determine whether one of the measured blood glucose values ​​is considered an outlier. The ability of the outlier unit 34 to determine whether a measured blood glucose value is an outlier allows the control device 30 to ignore anomalous measurements that may reduce the accuracy of RC recommendations. In the context of this disclosure, “outlier” refers to a data point that deviates significantly from other observations in a dataset. Outliers may result from variability in measurements or may indicate errors, the latter of which may be excluded from the dataset. Outliers can occur for many reasons, such as changes in device behavior or environmental factors, such as compression of the CGM, when measured blood glucose values ​​are measured by a CGM. The ability to identify and handle outliers is particularly useful in systems that provide recommendations or warnings based on accurate data, because it helps ensure that such recommendations or warnings are based on reliable and representative data.

[0109] The outlier unit 34 is configured to determine outliers using a Kalman filter (KF). In the context of this disclosure, a “Kalman filter” (KF) is an algorithm that uses a series of measurements observed over time, including statistical noise and other inaccuracies, to generate estimates of unknown variables that tend to be more accurate than those based on a single measurement or a single model alone. The outlier unit 34, configured to determine outliers using a KF, allows the outlier unit 34 to improve its ability to identify inaccurately measured blood glucose levels by effectively filtering out statistical noise and inaccuracies inherent in the data. This results in a more reliable dataset for generating RC recommendations, which is particularly beneficial for users who require accurate and timely interventions to manage their blood glucose levels. The predictive power of the KF allows for a continuous improvement in the system’s understanding of the user’s physiological state, thereby improving the overall accuracy and reliability of the RC recommendations provided by the control device 30.

[0110] As shown in Figure 2, KF operates in a two-step process. Firstly, the prediction step 52 estimates the current state of the system, and secondly, the update step 54 refines the estimate by incorporating the latest measurements. KF is particularly advantageous in the control unit 30 for determining outliers in measured blood glucose levels because it can effectively remove noise, and more notably, the analysis of KF residuals, which represent the prediction error for the observed values, provides an indicator of agreement between the measured values ​​and the model output. This analysis is particularly useful for detecting sensor failures because it enables the identification of discrepancies that may indicate outliers, thereby increasing the reliability of the dataset used to generate RC recommendations. For example, KF can predict the system state by integrating a personalized insulin-vs-glucose model, and then refine this prediction by comparing it with actual measured blood glucose levels, thereby identifying any discrepancies that may indicate outliers.

[0111] In some embodiments, the outlier unit 34 can use various derivations of KF for detecting candidate outliers in measured blood glucose levels. These derivations may be, but are not limited to, extended KF, unflavored KF, cubature KF, square root KF, information KF, ensemble KF, and particle filter. Each of these KF derivations provides a different approach to processing and analyzing data, some being more computationally intensive than others. For example, the particle filter may require the propagation of many particles through model equations to estimate a posterior state distribution that may be non-Gaussian, but requires fewer assumptions about the linearity of the system and the statistical nature of noise. The outlier unit 34 can further utilize various models, whether linear or nonlinear, physiological or empirical (black box, gray box), to evaluate discrepancies between model predictions and actual measured blood glucose levels (i.e., CGM measurements). This discrepancy evaluation is a core function of the outlier unit 34, as it helps identify when measured blood glucose levels deviate from the expected pattern and can indicate abnormalities or outliers in measured blood glucose levels. Depending on the hardware capabilities and specific requirements of the diabetes management system, more economical versions of the KF, such as a base KF coupled with a linear model, may be preferred. This approach focuses on the ability to capture sudden fluctuations in the discrepancy between the model and the measured values, which is of paramount importance for sensor failure detection. In contrast, more accurate state estimation, which may be desirable for model-based control applications such as calculating insulin from state predictions, potentially requires more complex KF derivations. Thus, the selection of the KF derivation and model type is guided by a trade-off between computational efficiency and the desired level of accuracy and robustness in outlier detection. According to a preferred embodiment, the KF is configured to predict the system state by integrating a personalized sixth-order insulin-vs-glucose linear model from a user setting, the model parameters including at least: Diet ratio (MR), and / or Insulin sensitivity factor (ISF), and / or The insulin diffusion rate, expressed by the diffusion time constant (τIOB), and / or The rate of food digestion, expressed by the digestion time constant (τD).

[0112] These features allow the control device 30 to provide a tailored approach to diabetes management by incorporating individual user settings. This personalization ensures that RC recommendations are based on a model that closely reflects the user's specific physiological response to insulin and carbohydrate intake, leading to more accurate and effective management of blood glucose levels. The use of a sixth-order linear model allows for a nuanced representation of insulin-to-glucose dynamics, which can explain complex interactions over time, thereby improving the system's predictive accuracy. This can be particularly advantageous in preventing hypoglycemic events, as the system can more accurately predict blood glucose trajectories and propose timely and appropriate RC interventions. It should also be noted that the synergistic integration of meal ratio (MR), insulin sensitivity factor (ISF), insulin diffusion rate (τIOB), and meal digestion rate (τD) within the sixth-order insulin-to-glucose linear model facilitates comprehensive and individualized management of blood glucose levels. By combining these parameters, the control device 30 can more accurately simulate the complex interactions between carbohydrate intake, insulin administration, and their respective effects and absorption rates. This synergistic effect allows for improvements to the model that reflects the user's specific metabolic pattern, thereby increasing the accuracy of RC recommendations. The combined effect of these parameters ensures that RC recommendations are not only based on static measurements but are also dynamically adjusted to the user's individualized physiological responses, leading to a more effective and preventive approach to preventing hypoglycemic events.

[0113] In the context of this disclosure, the “sixth-order insulin-vs-glucose linear model” refers to a mathematical representation that characterizes the relationship between insulin administration and blood glucose levels over time using a linear state-space model that approximates actual system dynamics. This model is described by a set of differential equations that express the dynamics of insulin action and glucose absorption up to the sixth-order derivative with respect to time. The sixth-order model aims to capture the first-order dynamics of the insulin-vs-glucose interaction and provides a balance between model complexity and computational efficiency, which may be particularly useful for real-time applications in insulin delivery systems.

[0114] As mentioned above, the use of a sixth-order linear model allows for a nuanced representation of the dynamics of insulin versus glucose, which can explain complex interactions over time and thereby improve the predictive accuracy of the system. However, KF, which is configured to predict the system state by integrating an insulin versus glucose linear model of any order, can also yield acceptable results.

[0115] In the context of this disclosure, “Diet Ratio” (MR) refers to a parameter used to personalize an insulin-versus-glucose model, representing the ratio of the amount of carbohydrates in a diet to the amount of insulin administered to metabolize those carbohydrates. MR is a factor in determining the insulin dose relative to a given carbohydrate intake and is used to adjust model parameters to reflect the user’s individual dietary habits and insulin sensitivity. MR can be set by the user, the healthcare provider, or determined based on historical data and may be adjusted over time to optimize glycemic control.

[0116] In the context of this disclosure, “insulin sensitivity factor” (ISF) refers to a parameter that quantifies the relationship between insulin administration and its effect on lowering blood glucose levels. ISFs may be used to adjust insulin dosages to achieve a desired glycemic target and are typically personalized based on the user’s insulin sensitivity and metabolic response. ISFs may be determined through clinical assessment or derived from historical data of the user’s insulin use and corresponding glycemic response.

[0117] In the context of this disclosure, the “diffusion time constant” (τIOB) refers to a parameter that represents the rate at which insulin diffuses into the bloodstream and begins to lower blood glucose levels after administration. This constant is an integral part of the insulin-vs-glucose model because it helps predict the time-dependent effect of insulin on blood glucose concentration. The diffusion time constant can be personalized based on the user’s physiological characteristics and can vary depending on a variety of factors. By incorporating τIOB into the model, the control device 30 can more accurately predict the effect of insulin on blood glucose levels over time, which is particularly useful for timing RC recommendations to prevent or mitigate hypoglycemia.

[0118] In the context of this disclosure, “digestion time constant” (τD) refers to a parameter that characterizes the rate at which ingested carbohydrates are digested and absorbed into the bloodstream as glucose, and thus appear in measured blood glucose levels. This constant is a component of the insulin-vs-glucose model that influences the timing and magnitude of the postprandial glucose response. The digestion time constant can be personalized to reflect the user’s individual digestive dynamics, which may be influenced by factors such as dietary composition, gastrointestinal motility, and metabolic health. By incorporating τD into the model, the control device 30 can more accurately predict the rise in postprandial blood glucose levels, which is particularly useful in determining the timing and amount of insulin administration or RC recommendations to maintain glycemic control.

[0119] In the context of this disclosure, model parameters including at least MR and / or ISF and / or τIOB and / or τD mean that the model includes parameters or derivatives of such parameters.

[0120] According to one embodiment, the outlier unit 34 is further configured to identify outliers, particularly during the postprandial period. This feature allows the control device 30 to specifically target the postprandial period, during which blood glucose levels are particularly likely to fluctuate upwards due to meal intake. By focusing on this period, the outlier unit 34 can more effectively identify readings (i.e., measured blood glucose levels) that do not fit the expected postprandial blood glucose pattern, thereby improving the accuracy of RC recommendations. This targeted approach to outlier detection is particularly beneficial for users who may experience unpredictable blood glucose responses after meals, as it provides a more refined analysis of blood glucose data during the time most relevant for accurate RC recommendations to prevent postprandial hypoglycemia. During the postprandial period, particularly around the time of meal reporting, the control device 30 can adjust the predictive rules for predicting RC recommendations. This adjustment is based on the expectation that blood glucose will rise due to the amount of carbohydrates ingested. Thus, the system can relax the normal predictive rules to avoid raising RC alerts, which in some examples may unnecessarily confuse the user. This mitigation ensures that the RC alert is triggered only when the user's blood glucose is maintained and truly guaranteed to reach a level lower than the desired RC trigger level. This approach helps prevent overcorrection of blood glucose levels, which can be particularly destructive after meals, when the body is naturally working to assimilate ingested carbohydrates and stabilize blood glucose.

[0121] According to a preferred embodiment, the outlier unit 34 has a distance (d) including the innovation matrix and KF residuals that is a dynamic threshold (d THRIf the value is smaller than ), the device is further configured to consider the last measured blood glucose value among the measured values ​​as an outlier. This characteristic allows the control device 30 to dynamically adjust the sensitivity of outlier detection, which is particularly beneficial in scenarios where blood glucose levels can fluctuate rapidly, such as during physical activity or after meal intake. Dynamic threshold d adapts to the user's current physiological condition. THR By configuring this setting, the control unit 30 / system 10 can more accurately distinguish between true outliers and legitimate rapid changes in blood glucose levels. This results in a more robust and reliable RC recommendation process, minimizing the risk of false alarms and ensuring that interventions are based on the user's true blood glucose state. The ability to fine-tune outlier detection in real time enhances the system's usefulness across a wide range of situations and provides users with greater confidence in the RC recommendations they receive.

[0122] In the context of this disclosure, “distance” (d) refers to a metric used to quantify the discrepancy between the predicted state of the system and the actual observed measurements. In other words, d is a measure of how the KF prediction error deviates from the expected prediction error. Thus, the innovation matrix is ​​used to account for the normal variability of the prediction error. A significant distance d means that the KF is abnormally wrong in its prediction, which may be an indication of sensor failure. Specifically, the Mahalanobis distance is a multivariate distance measure that accounts for the correlation between variables and scales the distance calculation according to the variability of each variable. This distance is calculated from the innovation matrix, which represents the variability of the KF residuals between the predicted and observed measurements, and the KF residuals, which are the discrepancy between the actual measurements and the KF predictions. The calculated distance d is a dynamically determined threshold d THR If the value is smaller than this, the last measured blood glucose level is considered an outlier. This approach enables a more sophisticated outlier detection mechanism that is sensitive to the inherent variability and correlation structure of the data, resulting in a more accurate and reliable blood glucose assessment for the purpose of RC recommendations.

[0123] Dynamic threshold d for outlier detection THRThe threshold is calculated based on the moving average or median of d signal values ​​acquired over a past time window. This approach allows the threshold to adapt to changes in the user's physiological state over time and provides a baseline that reflects the user's recent blood glucose trends. To further improve the accuracy of outlier detection, the calculation of the moving average or median can be trimmed to discard extreme values ​​such as peaks that may contain the sought outliers, thereby capturing a more representative baseline of d. THR An additive offset is applied to this, which can provide a safety margin to account for measurement uncertainty or expected physiological variability. Alternatively, d THR Applying a multiplicative scaling factor to d THR This allows for scaling in proportion to the variability observed in d. This scaling is particularly useful in situations where measured blood glucose levels show greater variability, such as during periods of intense physical activity or stress, and ensures that outlier detection remains sensitive to true outliers while avoiding false positives.

[0124] In some embodiments, the outlier unit 34 can use a statistical chi-squared test to assess the goodness of fit between the KF residuals and a theoretical distribution, such as a Gaussian distribution. This test is particularly useful in determining whether the observed residuals from KF deviate in a statistically significant way from what would be expected if the system correctly models the underlying physiological processes. A substantial deviation, such as one indicated by the chi-squared test result exceeding a predetermined threshold, may suggest that the measured blood glucose level is an outlier. This deviation is defined as reading a value larger than the expected value of the residual distribution, or the statistical mean, than would typically be observed in the absence of an outlier in the measured blood glucose level. Such a finding indicates a potential impairment of the measured blood glucose level or an anomaly in the glucose measurement process, prompting the outlier unit 34 to potentially exclude the affected data point from the RC recommendation process. In the context of this disclosure, “statistical chi-squared test” is a statistical method used to compare observed data with data that would be expected to be obtained according to a particular hypothesis. The test calculates a chi-squared statistic, i.e., a measure of the discrepancy between the observed frequency and the expected frequency in one or more categories. This statistic is compared to the chi-squared distribution to determine the likelihood of a randomly occurring observed distribution. If the calculated chi-squared statistic exceeds the value of the chi-squared distribution for a given level of statistical confidence, it suggests that the observed data does not fit the expected distribution and indicates the presence of outliers or deviations from the assumed pattern.

[0125] In some embodiments, machine learning techniques can be used to enhance outlier detection based on residuals obtained from KF. These techniques involve training a machine learning model on historical data to classify normal and abnormal behavior. Algorithms such as support vector machines (SVMs), random forests, or neural networks can be used for this purpose. SVMs offer effectiveness in high-dimensional spaces and the ability to model nonlinear boundaries with kernel trick features. Random forests, consisting of multiple decision trees, offer robustness against overfitting and provide a measure of feature importance. Neural networks, particularly deep learning models, can learn complex patterns and relationships in the data through their hierarchical structures.

[0126] The use of these machine learning techniques in the context of classification and clustering enables a comprehensive analysis of the data, and each technique can capture different forms or patterns that indicate outliers. By combining these techniques or running them simultaneously, the system can leverage a consensus approach to increase the reliability of outlier detection. For example, a voting mechanism can be implemented, and if the majority of the machine learning algorithms determine the presence of an outlier, the control device 30 may consider bypassing the RC recommendation.

[0127] Alternatively, a safer mode might require all algorithms to converge on outlier detection before taking action. This multi-algorithmic approach can improve the system's robustness against false positives and false negatives, ensuring that RC recommendations are based on accurate classification of blood glucose measurements.

[0128] According to one embodiment, to improve the accuracy of RC recommendations, the control unit 30 is configured to perform a three-step procedure for managing outliers in measured blood glucose levels. First, the outlier unit 34 is configured to analyze the residuals of blood glucose measurement levels using KF, thereby identifying candidate outliers that deviate from the predicted values. These candidates are then subjected to a series of data-driven confirmation rules, which may include assessing the temporal context of the readings, the physiological validity of the glucose trend, and the consistency of the data with known user activity or events. Once an outlier is confirmed, the control strategy of the RC unit 36 ​​is improved accordingly. This includes adjusting the RC recommendation to take the identified outlier into account, ensuring that the system's response is based on the user's true blood glucose state. By systematically determining, confirming, and responding to outliers, the control unit 30 provides a robust and reliable method for managing RC alerts, thereby improving the safety and effectiveness of diabetes management.

[0129] In one embodiment, one of the measured blood glucose values ​​is determined to be an outlier if it is determined by the outlier unit 34 to be affected by compression artifacts. Such a feature allows the control unit 30 to improve the reliability of RC recommendations by identifying and excluding glucose measurements that may be distorted due to physical interference with a glucose monitor, such as a CGM. When the sensor (i.e., CGM) is subjected to pressure, compression artifacts can occur, leading to falsely low glucose measurements that may trigger unnecessary RC recommendations. By recognizing these artifacts, the control unit 30 can prevent inappropriate responses to these inaccurate readings, thereby maintaining the accuracy of RC recommendations and avoiding potential overtreatment of hypoglycemia. This feature is particularly beneficial for users who are active or sleeping on the monitor, for example, as it ensures that RC recommendations are based on true glucose values ​​rather than artifacts from the measurement process.

[0130] In the context of this disclosure, “compression artifact” refers to a type of error in glucose readings that occurs when an external force is applied to a glucose monitoring device, typically a continuous glucose monitor (CGM), resulting in a transient deformation of the sensor or surrounding tissue. This deformation can cause interstitial fluid dynamics or disruption of sensor function, potentially leading to transient, artificial changes in measured blood glucose levels that do not accurately reflect the true blood glucose concentration. Compression artifacts are often characterized by a sudden, non-physiological drop in recorded blood glucose levels, followed by a return to a more typical reading after the pressure is released. These artifacts can be particularly problematic during periods of rest or sleep when the user is lying on the sensor, or during strenuous physical activity where the device may be compressed against the body. The ability to accurately identify and exclude measurements affected by compression artifacts is an essential part of ensuring the reliability of glucose monitoring systems and user safety, as it helps prevent inappropriate treatment decisions based on erroneous data.

[0131] In some embodiments, the control device 30 can utilize confirmation rules to verify that a candidate KF outlier is indeed a non-physiological compression artifact. The confirmation process may include a nocturnal condition, which focuses on detecting nocturnal compression artifacts to reduce the risk of false positives that may be increased by daily activity. The outlier unit 34 may further be configured to take into account the temporal distance to past declared, estimated, or measured physical activity and carbohydrate intake (i.e., the amount of carbohydrates consumed by a particular user), as these factors may influence the measured blood glucose level and the likelihood of a compression artifact. A downward residual condition may be applied if the negative residual value is a negative value that confirms a lowest point trend during the out-of-range period. This condition is particularly relevant at night as it enhances the safety of detection by reducing the likelihood of false positives. However, this condition may be relaxed in other applications where nocturnal detection is not appropriate. According to the present invention, “lowest point trend” refers to the pattern or trajectory of measured blood glucose levels reaching their lowest points over a given period.

[0132] The outlier unit 34 may be further configured to evaluate a first derivative condition in which the current slope of the measured blood glucose trend is compared to a predetermined negative threshold. If the slope is less than this threshold, a substantial and abnormal decrease in blood glucose indicating a potential compression artifact is confirmed. Furthermore, a second derivative condition may be evaluated if a sudden discontinuity in the slope compared to the predetermined threshold demonstrates a non-physiological change in the trend. An alternative embodiment of the second derivative condition may include calculating the ratio of two consecutive first derivative samples and comparing it to a predetermined threshold. This approach allows for the detection of rapid changes in the rate of blood glucose decrease, which is characteristic of compression artifacts.

[0133] Furthermore, a net IOB condition may be implemented, and the net IOB value is expected to be less than zero to enhance the safety of the detection process. This ensures that a rapid drop in measured blood glucose levels cannot be attributed to the effects of past insulin infusions. By incorporating these confirmation rules, the control device 30 can more accurately identify and eliminate compression artifacts, thereby improving the reliability of RC recommendations and ensuring that interventions are based on accurately measured blood glucose levels. In the context of this disclosure, “net insulin onboard” or “net IOB” refers to the amount of insulin administered and still available in the user’s body, adjusted for the basal insulin rate. Net IOB is calculated by subtracting from total IOB the amount of insulin that would have been delivered as part of the user’s basal rate. This calculation takes into account insulin infused to neutralize ingested carbohydrates, as well as insulin administered to correct high blood glucose levels. Since net IOB considers insulin that effectively exceeds the basal insulin requirement, it provides a more accurate reflection of the potential effect of insulin on the reduction of blood glucose levels at any given moment. In the context of this disclosure, “basal insulin rate” refers to the insulin administered to a user to maintain normal blood glucose levels in the absence of food. This rate is typically set based on the user’s basal metabolic insulin requirement, which is the fasting or intermeal insulin requirement. The basal insulin rate is designed to mimic the steady-state release of insulin by a healthy pancreas and is often delivered by an insulin pump or through long-acting insulin injections. The basal insulin rate can be adjusted based on a variety of factors, such as the user’s daily life, physical activity, stress levels, or changes in other physiological conditions that may affect insulin sensitivity. It is a fundamental component of insulin therapy for individuals with diabetes, particularly those receiving, for example, intensive insulin therapy. The basal insulin rate is distinct from bolus insulin administered to manage the rise in blood glucose levels due to carbohydrate intake from food.

[0134] In some embodiments, the outlier unit 34 of the control device 30 is configured to analyze the relationships between various features representing a feature space, such as measured blood glucose derivatives, net IOB, carbohydrate onboard (COB), KF residuals, and additional sensor data such as accelerometer or heart rate, using, for example, a support vector machine (SVM). The SVM allows for the creation of a hyperplane that effectively separates the feature space into two distinct categories: out-of-range periods and physiological periods. This separation helps distinguish between normal physiological variability and potential outliers that may indicate defects in measured blood glucose levels or non-physiological events affecting glucose measurements. In the context of this disclosure, “hyperplane” is a geometric concept that generalizes the concept of a plane in higher dimensions. Specifically, in an n-dimensional feature space, the hyperplane is a flat affine subspace of dimension n-1, meaning it is one dimension less than the feature space itself. For example, in a three-dimensional space, the hyperplane is a two-dimensional plane. In the context of machine learning, particularly in the use of SVMs, a hyperplane is used to separate different classes of data points by finding a plane that maximizes the margin between classes. A hyperplane is defined by a set of weights and biases, which are optimized during SVM training to create a decision boundary that can classify new data points based on their position relative to the hyperplane.

[0135] In another embodiment, the outlier unit 34 may utilize a trained random forest method to define a threshold surface in the feature space. The random forest method, consisting of an ensemble of decision trees, can classify the measured blood glucose levels as either outliers or non-outliers based on the aforementioned features. This method is particularly advantageous due to its ability to handle high-dimensional data and its robustness against overfitting, which is beneficial for the accurate detection of outliers in measured blood glucose levels.

[0136] Furthermore, in some cases, the outlier unit 34 may implement a neural network, including a convolutional neural network (CNN) or a recurrent neural network (RNN), for outlier classification. These types of neural networks are adept at capturing complex patterns and time dependencies in the data, which can be particularly useful for identifying subtle or transient anomalies in measured blood glucose levels that may not be readily apparent through other means.

[0137] Furthermore, the outlier unit 34 incorporates Gaussian process regression (GPR) to model the relationship between input features and output classifications. GPR provides not only predictions for classification but also uncertainty estimates, which are of great value when defining threshold surfaces with confidence intervals. These uncertainty estimates can enhance the decision-making process by providing a probabilistic assessment of the data, thereby enabling more informed and careful interventions in blood glucose management.

[0138] The outlier unit 34 is configured to determine that the measured blood glucose value is affected by compression artifacts, based at least on deviations from the predicted blood glucose and the measured blood glucose value. This configuration allows the control device 30 to distinguish between true fluctuations in the measured blood glucose value and fluctuations affected by compression artifacts. By comparing the deviation between the predicted blood glucose and the actually measured blood glucose value, the outlier unit 34 can effectively identify when the reading is likely to be impaired. This ensures that RC recommendations are based on accurate and reliable data, especially in situations where the CGM sensor may be under pressure, such as during sleep or physical activity. The ability to filter and remove these specific types of outliers prevents unnecessary or inaccurate recommendations for carbohydrate intake, thereby optimizing the user's blood glucose management and reducing the risk of hypoglycemia. This feature is particularly advantageous for maintaining the integrity of the RC recommendation process and ensuring that the user receives warnings and interventions that accurately reflect the user's physiological state.

[0139] According to this disclosure, the outlier unit 34 of the control device 30 is configured to utilize a predictive model that takes into account both predicted and actually measured blood glucose levels. The predictive model can predict expected blood glucose levels based on a variety of factors, including but not limited to historical blood glucose data, recent trends in the measured blood glucose levels of a particular user, and the physiological parameters of a particular user, using a variety of algorithms and techniques. When the outlier unit 34 detects a deviation between the predicted blood glucose level and the actually measured blood glucose level that exceeds a predetermined threshold, it determines this to be an indicator of a potential pressure artifact. This threshold can be dynamically adjusted based on the user's current state and the variability of the user's blood glucose measurements. By setting this threshold, the control device 30 can distinguish between normal physiological fluctuations in blood glucose levels and readings that are likely to be affected by external pressure on the glucose monitor. This feature is particularly useful in scenarios where the user is engaged in activities that may put pressure on the CGM, such as sleep or exercise. The ability to accurately identify and ignore these affected measurements helps maintain the integrity of the RC recommendation process and ensures that the user is provided with reliable and actionable information to manage their diabetes.

[0140] The control device 30 includes an RC unit 36. The RC unit 36 ​​is configured to estimate RC recommendations using the slope of the measured blood glucose level. The function of the RC unit 36 ​​to estimate RC recommendations using the slope of the measured blood glucose level enables a dynamic assessment of the user's glucose tendency. This allows the control device 30 to predict potential hypoglycemic events and recommend timely and appropriate interventions, for example, potentially reducing the risk of severe hypoglycemia.

[0141] The RC unit 36 ​​is further configured to estimate an RC recommendation if the most recent measured blood glucose level among the measured blood glucose levels shows a slowdown. This characteristic allows the control device 30 to provide RC recommendations that not only respond to the current blood glucose level but also predict future trends. By analyzing the slowdown in blood glucose levels, the RC unit 36 ​​can identify when the rate of decrease is slowing down, which can indicate an impending stabilization or a potential reversal of a downward trend. This proactive function can be particularly advantageous in preventing hypoglycemia because it allows for early intervention by RC when blood glucose levels are still within a safe range, rather than waiting for blood glucose levels to reach a level that already indicates hypoglycemia. This proactive approach can improve the user's quality of life by reducing the frequency and severity of hypoglycemic episodes and can also minimize the user's cognitive burden by reducing the number of decisions the user has to make regarding glucose management.

[0142] The RC unit 36 ​​is configured to adjust RC recommendations based on physical activity data. This configuration allows the control device 30 to consider the impact of physical activity on measured blood glucose levels, which is a dynamic factor in diabetes management. By including physical activity data in the RC recommendation process, the control device 30 is able to provide more personalized, context-aware recommendations. By adjusting RC recommendations based on the user's physical activity, the system can better predict the physiological effects of exercise, which often include increased insulin sensitivity and glucose uptake by muscles, potentially increasing the risk of exercise-induced hypoglycemia. This feature enhances the control device 30's ability to tailor RC recommendations to the unique user's lifestyle, promoting more effective and safer diabetes management. Furthermore, it helps users maintain an active lifestyle by providing them with confidence that their blood glucose levels are being monitored and managed in consideration of their physical activity.

[0143] The control device 30 comprises a prediction unit 40, which is configured to calculate at least a predicted blood glucose level based at least on user data. Such features enable the control device 30 to enhance diabetes management by predicting the user's future blood glucose levels. The ability of the prediction unit 40 to calculate predicted blood glucose based on user data allows the control device 30 to predict potential hypoglycemic events before they occur, enabling proactive measures to be taken. This predictive capability provides users and healthcare providers with valuable information for making decisions based on carbohydrate intake and insulin medication, potentially leading to improved glycemic control. Furthermore, the use of user data by the prediction unit 40 for predictive purposes ensures that predictions are personalized and tailored to individual metabolic responses, further improving the accuracy and relevance of RC recommendations. This can result in a reduction in hypoglycemic episodes, improved patient safety, and an overall improvement in the quality of life for individuals with diabetes. In the context of this disclosure, “prediction unit” 40 refers to a component within the control device 30 responsible for calculating future physiological states, such as predicted blood glucose levels, based on current and historical user data. The prediction unit can utilize various mathematical models, algorithms, and data processing techniques to analyze trends and patterns in a unique user's physiological data, thereby enabling the control device 30 to predict future changes in blood glucose levels. For example, the prediction unit 40 can predict blood glucose levels using a linear regression model based on recent measurements, as well as known rates of glucose absorption and insulin action. In another embodiment, the prediction unit 40 can use machine learning algorithms such as neural networks or support vector machines to analyze complex datasets and identify subtle correlations that can improve the accuracy of its predictions. Furthermore, the prediction unit 40 can incorporate time series analysis to detect periodic patterns in blood glucose fluctuations that may be influenced by daily activities, meal times, and sleep cycles.These examples demonstrate the versatility of the prediction unit 40 in adapting to the individualized characteristics of a unique user's metabolic responses and lifestyle factors, thereby enhancing the personalized management of diabetes. In some embodiments, the prediction unit may also employ a model predictive control (MPC) strategy to enhance the prediction of future physiological states. An MPC is a form of control algorithm that involves creating a model of the system for predicting future outcomes and adjusting the control inputs accordingly. In the context of blood glucose management, MPCs may be particularly useful for taking into account the delayed effects of insulin delivery and carbohydrate absorption on blood glucose levels. By considering the range of possible future scenarios and the possible effects of different insulin administration strategies, the prediction unit can optimize blood glucose control over a specified prediction range, thereby contributing to more stable and accurate diabetes management.

[0144] The RC unit 36 ​​is configured to estimate an RC recommendation when the predicted blood glucose is lower than the predicted blood glucose threshold (θ1), the current blood glucose is lower than the current blood glucose threshold (θ2), and the insulin onboard (IOB) is higher than the IOB threshold (θ3). These features allow the control device 30 to provide a more nuanced and responsive approach to managing blood glucose levels, particularly in the context of insulin therapy. By incorporating multiple thresholds that take into account predicted blood glucose, current blood glucose, and IOB, the RC unit 36 ​​can make more informed decisions about when to recommend carbohydrate intake. This multifaceted analysis ensures that RC recommendations are made at the appropriate time, reducing the risk of over- or under-treatment of hypoglycemia. The integration of these thresholds enables more individualized and accurate management of diabetes, which can lead to better overall glycemic control and a reduced incidence of hypoglycemic events. This proactive strategy can improve patient outcomes and reduce long-term complications associated with diabetes. In the context of this disclosure, “insulin onboard” (IOB) refers to the amount of insulin administered and still available in the body of the specific user. IOB is a dynamic value that reflects insulin that has not yet exerted its full glucose-lowering effect. Estimating IOB involves calculating the remaining available insulin from a previous dose, taking into account the time since administration and the known pharmacokinetics of the insulin preparation used. IOB can be influenced by factors such as the type of insulin, the dosage, and the specific insulin sensitivity of the user. For example, rapid-acting insulin has a different IOB profile compared to long-acting insulin. IOB is an integral part of the decision-making process in insulin management because it helps prevent insulin buildup and the risk of hypoglycemia. In some embodiments, the control device 30 is configured to estimate IOB by utilizing a decay function that models the progressive absorption effect of insulin over time. This estimation allows the RC unit 36 ​​to adjust RC recommendations by taking into account the residual effect of previously administered insulin on current and predicted blood glucose levels.

[0145] According to one embodiment, the RC unit 36 ​​is configured to dynamically adjust the predicted blood glucose threshold (θ1), the current blood glucose threshold (θ2), and the IOB threshold (θ3) based on real-time data and the user's current physiological state. For example, the thresholds may be adjusted to take into account factors such as physical activity, stress, illness, or changes in medication. This dynamic adjustment allows the control device 30 to provide RC recommendations tailored to the user's current situation, thereby improving the safety and effectiveness of insulin therapy.

[0146] Furthermore, the RC unit 36 ​​may incorporate safety features that prevent RC recommendations from being issued when they are not due to safety or user interference reasons. For example, if the predicted blood glucose level exceeds θ1 and the current blood glucose level exceeds θ2, the RC unit 36 ​​may withhold an RC recommendation even if the IOB is greater than θ3. This safety mechanism ensures that RC recommendations are made cautiously, minimizing the risk of unnecessary carbohydrate intake and potential subsequent hyperglycemia, in accordance with the user's actual blood glucose status.

[0147] The RC unit 36 ​​is configured to calculate RC recommendations based on the distance between a blood glucose target and a linearly predicted blood glucose level for a given predicted range, which is then converted to carbohydrate units via a glycation ratio. Such a configuration allows the control device 30 to provide accurate and personalized RC recommendations by quantifying the precise amount of carbohydrates that is likely to bring a particular user's blood glucose level back to the desired target range. By calculating the distance between the blood glucose target and the linearly predicted blood glucose level, the RC unit 36 ​​can determine the appropriate size of intervention for a given situation. Converting this distance to carbohydrate units via a glycation ratio simplifies the process for the user, transforming complex data into readily usable and understandable guidance. This approach not only helps in the immediate correction of hypoglycemia but also contributes to the long-term stability of blood glucose levels by providing tailored recommendations that are directly aligned with the unique physiological requirements of the user. The ability to predict blood glucose and translate it into tangible RC recommendations allows users to manage their condition with greater confidence and accuracy, potentially reducing the frequency of hypoglycemic episodes and improving overall quality of life.

[0148] In the context of this disclosure, the “glycation ratio” is calculated based on the relationship between the amount of carbohydrates (CHO) expected to raise blood glucose levels to a target range and the expected increase in blood glucose levels over a given predicted range. The glycation ratio is estimated using a formula that takes into account the specific user’s body weight (BW) and a standard absorption coefficient typically derived from empirical data. For example, the glycation ratio can be estimated from the formula sugRatio=20BW / (0.680) in [gCHO / g / L], where BW is in kilograms. This ratio is then used to convert the distance between predicted blood glucose and the blood glucose target into carbohydrate units, which represents the recommended amount of carbohydrates to be consumed to correct or prevent hypoglycemia. Thus, the glycation ratio functions as a conversion factor that translates expected blood glucose fluctuations into equivalent amounts of carbohydrates, facilitating the calculation of personalized RC recommendations.

[0149] The RC unit 36 ​​is further configured to adjust the RC recommendation based at least on IOB, such that the adjusted RC recommendation is an increasing function of IOB. Such a configuration allows the control device 30 to dynamically adjust the RC recommendation according to the amount of insulin still available in the unique user's body, known as IOB. By considering IOB, the RC unit 36 ​​can more accurately determine the amount of carbohydrates a unique user may need to prevent or treat hypoglycemia. This is particularly advantageous because it takes into account the continuous glucose-lowering effect of insulin, which can vary depending on factors such as the type of insulin used, the time since administration, and the unique user's insulin sensitivity. The ability to adjust the RC recommendation as an increasing function of IOB ensures that the user is not advised to consume more carbohydrates than actually needed based on their current physiological state, thereby avoiding potential hyperglycemia. This feature increases the accuracy of the RC recommendation, leading to more effective and safer blood glucose management, and ultimately contributing to the overall health and metabolic stability of the unique user.

[0150] At least IOB-based RC recommendations correspond to net IOB-based recommendations and ensure that RC units 36 consider dynamic insulin levels when determining the appropriate amount of rescue carbohydrates recommended to prevent or treat hypoglycemia. This approach aligns recommendations with the current physiological insulin profile of the specific user, thus enabling more personalized and accurate RC recommendations.

[0151] The control device 30 includes a transmitting unit 38. The transmitting unit is configured to transmit RC recommendations. The transmitting unit is configured to transmit an RC recommendation only if the last measured blood glucose value among the measured values ​​is not determined to be an outlier. The conditional operation of the transmitting unit 38, which transmits an RC recommendation only if the last measured blood glucose value is not considered to be an outlier, ensures that the recommendations provided to the user are based on reliable and accurate data. This selective transmission serves as a safeguard against providing potentially harmful advice based on misreadings.

[0152] The transmitting unit 38 is configured to transmit RC recommendations to another device, such as a mobile phone, smartwatch, dedicated receiver, or healthcare provider's monitoring system. This transmission can be made via a variety of communication protocols, including but not limited to Bluetooth®, Wi-Fi, cellular networks, or Near Field Communication (NFC). The ability to transmit RC recommendations to another device allows the user or healthcare provider to receive alerts in a timely manner and take appropriate action, thereby improving user safety and the effectiveness of diabetes management. The transmitting unit 38 is further configured to send an alert to the user or designated recipient when an RC recommendation is transmitted. The alert functions to notify the user or recipient of the RC recommendation and may prompt the user or recipient to take action based on the provided recommendation. Depending on the user's preference and the capabilities of the receiving device, the alert may take the form of a visual notification, an audible alarm, vibration, or any combination thereof. This feature enhances the user's or caregiver's responsiveness to potential hypoglycemic events and contributes to improved safety and preventive diabetes management.

[0153] According to one embodiment, if one of the measured blood glucose levels is considered an outlier, the transmitting unit is configured not to transmit an RC recommendation until a bypass period has elapsed. This bypass period is maintained until at least one of the measured blood glucose levels recovers and is no longer considered an outlier. The bypass period ends based on two conditions: either the time has reached or exceeded the time obtained by adding a predetermined rise time to the lowest point time, or the measured blood glucose level has risen above a certain level. According to one embodiment, the bypass period is designed with safety exceptions to ensure the well-being of a particular user. If the most recently measured blood glucose level falls below a safety threshold, such as 55 mg / dL, or if the bypass period extends beyond its maximum duration, the transmitting unit is configured to transmit an RC recommendation again. With such a configuration, the control device 30 can prevent the risk of a measured blood glucose level being an outlier and failing to detect hypoglycemia even though it actually is.

[0154] In one embodiment, the control device 30 is configured to evaluate the blood glucose level prediction up to a predetermined prediction range before the bypass period. If this blood glucose level prediction falls below a predetermined hypoglycemia safety threshold, the bypass period is shortened to zero. This is because it indicates a true risk of hypoglycemia. With such a configuration, the control device 30 can prevent the risk of actual hypoglycemia. For the purpose of performing these evaluations, the control device 30 may be configured to use any suitable prediction means, such as a linear model, a nonlinear model, or a machine learning algorithm. The selection of the prediction means can be tailored to the specific requirements of the diabetes management system, taking into account factors such as computational efficiency, prediction accuracy, and the individual physiological response patterns of the user.

[0155] The present invention also relates to a method for determining the recommended amount of RC. This method is carried out by the control device 30 as described above and includes the following steps. Step 50 to retrieve user data, A prediction step 52 predicts or estimates the current state of the system, Update step 54 involves updating or improving the estimate by incorporating the latest measured blood glucose levels, Step 56, which determines the RC recommendation using at least the slope of the measured blood glucose levels, Step 58: Submit an RC recommendation.

[0156] Step 58, which sends an RC recommendation, is performed only if the last measured blood glucose level among the measured levels is not determined to be an outlier based on the predictions in step 52 and the updates in step 54 as described above.

[0157] The embodiments, technical effects, and definitions disclosed herein with respect to the control device 30 are also applicable to the method described herein. This method encompasses steps that fully utilize the functions and features of the control device 30 described herein. Therefore, all embodiments, technical effects, and definitions relating to the device, including but not limited to changes in ISF over time and methods for calculating ISF, are equally applicable to the method. This ensures a comprehensive and unified understanding of both the embodiments of the control device 30 and the method of the present invention, facilitating the implementation and use of the disclosed technology across a range of applications.

[0158] The present invention also relates to a computer program that includes instructions causing the control device 30 described above to perform the steps of the method described above. [Explanation of symbols]

[0159] 10 Systems 12. Continuous glucose monitor (CGM) 20. Insulin infusion device 30 Control device 32 units acquired 34 Outlier Units 36 RC units 38 Transmitter Unit 40 prediction units 50 Steps to retrieve user data 52 Prediction Steps 54 Update Steps 56 Steps to determine RC recommendations Steps to submit 58 RC recommendations

Claims

1. A control device (30) for determining the recommended amount of carbohydrates (RC), wherein the control device is An acquisition unit (32) is configured to acquire user data, each piece of user data has a timestamp, the user data is associated with a specific user, and the user data is at least The amount of insulin injected into the aforementioned specific user, The amount of carbohydrates consumed by the aforementioned specific user, An acquisition unit comprising: a plurality of physiological values ​​of the specified user, the plurality of physiological values ​​of the specified user comprising at least a measured blood glucose level; An outlier unit (34) is configured to determine whether or not one of the measured blood glucose values ​​is judged to be an outlier, An RC unit (36) is configured to estimate an RC recommendation using a mathematical function of the measured blood glucose level, The system includes a transmitting unit (38) configured to transmit the RC recommendation, The control device (30) is configured such that the transmitting unit (38) transmits the RC recommendation only if the last measured blood glucose value among the measured blood glucose values ​​is not determined to be an outlier.

2. The control device (30) according to claim 1, wherein the outlier unit (34) is configured to determine outliers using a Kalman filter (KF).

3. The KF is configured to predict the system state by integrating a personalized sixth-order insulin-vs-glucose linear model based on user settings, and the model parameters are at least: Diet ratio (MR), and / or Insulin sensitivity factor (ISF), and / or Diffusion time constant (τ ＩＯＢ The insulin diffusion rate expressed by, and / or Digestion time constant (τ Ｄ The control device (30) according to claim 2, comprising a food digestion rate represented by ).

4. The outlier unit (34) is defined as the distance (d) related to the innovation matrix and the KF residuals when the dynamic threshold (d) is reached. ＴＨＲ The control device (30) according to claim 3, further configured to determine that the last measured blood glucose value among the measured blood glucose values ​​is an outlier if it falls below ).

5. The control device (30) according to any one of claims 1 to 4, wherein the outlier unit (34) is further configured to identify outliers, particularly during the postprandial period.

6. The control device (30) according to any one of claims 1 to 5, wherein the RC unit (36) is further configured to estimate an RC recommendation when the most recent measured blood glucose level among the measured blood glucose levels shows a slowdown.

7. The control device (30) according to any one of claims 1 to 6, wherein the control device (30) comprises a prediction unit (40), and the prediction unit (40) is configured to calculate at least a predicted blood glucose level based at least on the user data.

8. The control device (30) according to claim 7, wherein the RC unit (36) is configured to estimate the RC recommendation when the predicted blood glucose is less than the predicted blood glucose threshold (θ1), the current blood glucose is less than the current blood glucose threshold (θ2), and insulin onboard (IOB) is greater than the IOB threshold (θ3).

9. The control device (30) according to any one of claims 1 to 8, wherein if it is determined that one of the measured blood glucose values ​​is affected by compression artifacts due to the outlier unit (34), it is determined to be an outlier.

10. The control device (30) according to claims 7 and 9, wherein the outlier unit (34) is configured to determine, based on deviations from at least predicted blood glucose and measured blood glucose values, that one of the measured blood glucose values ​​is affected by compression artifacts.

11. The control device (30) according to any one of claims 1 to 10, wherein the acquisition unit (32) is further configured to acquire user data including the user's physical activity data, and the RC unit (36) is configured to adjust the RC recommendation based on the physical activity data.

12. The control device (30) according to any one of claims 1 to 11, wherein the RC unit (36) is configured to calculate the RC recommendation based on the distance between a blood glucose target and a linearly predicted blood glucose for a predetermined prediction range, the distance being converted to carbohydrate units via a glycation ratio.

13. The control device (30) according to claim 12, wherein the RC unit (36) is further configured to adjust the RC recommendation based on at least the IOB, for example, the adjusted RC recommendation is an increasing function of the IOB.

14. A method for determining the recommended amount of RC, which is implemented by the control device (30) described in claim 1, The steps include: acquiring the user data (50); and determining the RC recommendation (56) using at least a mathematical function of the measured blood glucose level. The process includes the step (58) of transmitting the aforementioned RC recommendation, The step of sending the RC recommendation (58) is performed only if the last measured blood glucose value among the measured blood glucose values ​​is not determined to be an outlier.

15. A computer program comprising instructions for causing the control device (30) described in claim 1 to perform each step of the method described in claim 14.

Images