Method and means for predicting postprandial blood glucose level
Patent Information
- Application Number
- JP2023563093
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2021-04-14
- Filing Date
- 2022-04-13
- Publication Date
- 2025-04-22
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing methods for predicting postprandial blood glucose levels using patient data-based models are not robust, accurate, or reliable, especially outside controlled clinical environments, due to the complexity of daily life factors affecting blood glucose dynamics.
A computer-implemented method involving data preprocessing steps to clean and filter patient data, followed by inputting the cleaned data into a blood glucose level prediction model, such as a neural network, to identify the best model parameters and improve prediction accuracy.
The method provides unprecedentedly robust, reliable, and accurate predictions of future blood glucose levels, enabling more precise insulin dose recommendations and improved glycemic control.
Smart Images
Figure 00000000_0000_ABST
Abstract
Description
[Technical field]
[0001] technical level The present invention relates to methods and means for monitoring, controlling and predicting blood glucose levels, in particular for post-prandial blood glucose prediction. [Background technology]
[0002] Patients with type 1 diabetes mellitus (T1DM) require insulin injections to mitigate the long-term effects of chronically elevated blood glucose (BG) levels. However, too much insulin leads to hypoglycemia, a potentially life-threatening situation. Managing BG levels by injecting appropriate amounts of insulin is a challenging task and a considerable burden for diabetic patients, especially due to the large day-to-day variations in BG dynamics and the myriad of factors that affect BG levels. Therefore, designing smart control algorithms to help patients manage BG levels has kept control engineers busy for years already.
[0003] These research efforts have resulted in the design of numerous algorithms to assist patients in controlling their blood glucose levels, for example, by alerting the patient about potential hypoglycemia or hyperglycemia, by providing insulin dosing advice via decision support systems, or by providing closed-loop glycemic control via an artificial pancreas.
[0004] Most of these algorithms are based on a mathematical model for the prediction of future BG values. A wide variety of models have been proposed for this purpose. Typically, these can be classified as either physiological models or patient data-based models. Patient data-based models are the focus of current research, since they are much easier to parameterize / individualize than physiological models.
[0005] However, unfortunately, it has been found that models for prediction of future BG values, such as physiological models or models based on patient data, can provide reliable and accurate results for prediction of future BG values only under very specific restricted and well-controlled circumstances, e.g. clinical situations with a strict regimen of well-defined meal intake at well-defined times and with well-defined insulin doses at well-defined insulin injection times.
[0006] In more practical and realistic situations, i.e., during the daily routine of a patient's real life outside of any controlled clinical environment or outside of any medical procedure environment, it has proven very difficult to obtain robust and reliable predictions or estimates of a patient's future BG values using model approaches based on patient data. Summary of the Invention [Problem to be solved by the invention]
[0007] assignment It is therefore an object of the present invention to provide methods and means for obtaining more robust, more accurate and more reliable predictions of a patient's future blood glucose levels. It is further an object of the present invention to identify and obtain more reliable models for describing the overall behavior of blood glucose levels when using patient data based models or physiological model approaches.
[0008] It is an object to provide methods and means for obtaining more robust, more accurate and more reliable predictions of a patient's future blood glucose levels when using a model approach based on patient data, particularly when using medical patient data collected during the patient's daily routine outside of a controlled clinical setting.
[0009] It is a further object of the present invention to provide patients with improved means for monitoring, managing and controlling blood glucose levels based on the obtained output of a patient data based model for predicting future blood glucose levels. [Means for solving the problem]
[0010] solution According to the present invention, the object is achieved by a computer implemented method, a computer system, a computer storage medium and a medical device according to the independent claims.
[0011] For example, a method for predicting future blood glucose levels, in particular for post-prandial blood glucose level prediction, may be a computer-implemented method, comprising one, some or all of the following exemplary possible steps: - receiving a first medical data set of a patient covering a time range, the first medical data set including glucose data and further other medical data of the patient; extracting a second medical dataset from the first medical dataset, the second medical dataset being a subset of the first medical dataset, the extracting comprising: - identifying duplicates in the first medical dataset and removing the identified duplicates; - identifying data values that exceed a predetermined maximum threshold data value and removing data associated with said identified data values; - identifying data values below a predetermined minimum threshold data value and removing data associated with said identified data values; identifying data values that differ from a predetermined expected data value by more than a predetermined amount and removing data associated with said identified data values; ● identifying incomplete data having missing data values and removing the identified incomplete data; - identifying at least one predefined time-varying data pattern and removing data associated with the identified time-varying data pattern; and extracting a second medical data set comprising at least one of: providing the extracted second medical data set as an input to a blood glucose prediction model or a blood glucose prediction model algorithm; predicting future blood glucose levels of the patient using an output of the blood glucose prediction model or blood glucose prediction model algorithm based on the second medical data set; may include:
[0012] In this specification, a blood glucose prediction model may in particular be understood as being a patient data based blood glucose prediction model or as a physiological prediction model.
[0013] The exemplary glucose data of the first medical data set may come from a glucose monitoring system that measures / records a patient's blood or interstitial fluid glucose levels at regular intervals or scheduled times.
[0014] In this specification, the term glucose data may be understood to include glucose data measured in the interstitial fluid of a patient and / or glucose data measured in the patient's blood, i.e. blood glucose data. The term glucose level may also be understood to include the glucose level of glucose in the interstitial fluid and / or the glucose level in the blood, i.e. blood glucose level.
[0015] An exemplary step of providing the extracted second medical dataset as an input to a blood glucose prediction model may include, inter alia, providing the extracted second medical dataset as a training dataset for a blood glucose prediction model algorithm, such as a neural network algorithm, and training the blood glucose prediction model algorithm using the extracted second medical dataset.
[0016] For example, based on the extracted second medical data as a training dataset, a blood glucose prediction model algorithm can be trained to identify the best blood glucose prediction model and / or to identify the best parameters and / or best parameter values of the blood glucose prediction model.
[0017] In this specification, further other medical data may be understood as referring to data regarding food intake, such as amount of carbohydrates in / from food intake(s), data regarding medication, such as data regarding insulin injections, or measured or recorded insulin levels, such as number / amount of recorded bolus insulin injections or current insulin levels, heart rate, blood pressure, hormonal status / levels, mood, oxygen saturation, other analyte data (i.e. further physiological data other than glucose data / BG, e.g. other blood analyte data such as hemoglobin levels), or any other data of medical interest.
[0018] In this specification, the term medical dataset covering a time range may be understood, inter alia, as a dataset that can be ordered in time such that each data entity or each data type, e.g. blood glucose value or glucose level, can be associated with a point in time, e.g. marked with a timestamp.
[0019] The medical dataset(s) may be in an exemplary electronic format, for example the first and second medical datasets may be arranged in a row-based table format, where each row or each row index corresponds to a time point, where each row includes at least one entry for a value of a particular medical data entity or type, for example a glucose level value or blood glucose value, and where each column represents a different or separate data entity or data type, for example a glucose level or blood glucose value, a time, a type of meal, an amount of carbohydrate in a meal intake, an amount of a bolus insulin injection, a hormone level, or other data entity or data type of medical interest.
[0020] However, it is contemplated that the medical dataset may also be organized in a column-based format, with each column or each column index corresponding to a time point.
[0021] The term data or medical data may be understood to include, inter alia, data entities and / or data types and / or data values and / or data points.
[0022] An exemplary first medical dataset may be a hybrid dataset of data automatically acquired from medical sensors and data manually entered by the patient / user.
[0023] The term removing data may be understood as removing data from one / said first medical data set, including inter alia removing data including any associated data values.
[0024] The possible data entities or data types or data formats of the data values may include, inter alia, numeric types, e.g. floating point or fixed point or integer types, string or text types, e.g. characters or strings, Boolean types, or other types such as composite types, e.g. arrays, vectors, etc.
[0025] As used herein, the term extracting may be understood to refer to selecting or generating or filtering, and the term identifying may be understood to refer to ascertaining or searching or scanning.
[0026] The exemplary step of identifying duplicates within the first medical dataset and removing any identified duplicates may be understood to include removal of all occurrences of the identified duplicates, i.e., all data related to or associated with the identified duplicates may be removed.
[0027] However, it is also possible that one unique occurrence / entry of the data of the identified duplicate may be retained, for example, if two identical entries, e.g., two identical rows, occur in the first medical dataset, then only one entry / row is removed from the dataset and the one entry / row associated with the identified duplicate(s) is retained.
[0028] While the exemplary first medical dataset may contain multiple valid identical entries / valid duplicates, e.g. entries with the same data and the same / identical data values at different times, i.e. at / with different time indexes, since, for example, a patient may have exactly the same meal at different times, suspicious or anomalous double entries or duplicates may be identified by checking whether such identical entries are recorded as occurring within a close time interval, or even at the same time, e.g. with the same timestamp.
[0029] For example, in the case of multiple entries having equal values for the amount of carbohydrates from dietary intake and / or multiple entries having equal values for the amount of insulin injected, if such entries are recorded / are within a predetermined short time interval, e.g., less than 5 minutes, the entries may be marked as suspicious or anomalous during the exemplary step of identifying duplicates in the first medical dataset and may then be removed from the exemplary first medical dataset.
[0030] An exemplary step of identifying data values above a predetermined maximum threshold data value may be based on a predetermined maximum threshold data value derived from a statistical analysis of the patient's previously recorded medical data set.
[0031] For example, some data / data values of the first medical data set may indicate an unrealistically high carbohydrate intake(s) and / or an unrealistically high insulin input.
[0032] For example, carbohydrate and bolus insulin intakes with values greater / higher than a certain threshold may be marked as suspicious in the data and removed from the first medical data.
[0033] An exemplary maximum threshold data value may be calculated / defined, for example, as 1.5 times the interquartile range above the 75% quartile (3rd quartile) of all carbohydrate and / or insulin values for a particular patient, and values may be obtained from historical patient records.
[0034] This particular maximum threshold data value is exemplary only and other ranges / limits and / or other data, such as heart rate or other analyte data, may be used to derive a maximum threshold data value at which underlying data may be removed from the first medical data set.
[0035] In other words, the predetermined maximum threshold data value may be based on determining the interquartile range above the 75% quartile of all available data values of a particular data type, including any previously / historically recorded data values for the patient.
[0036] The exemplary step of identifying data values below a predetermined minimum threshold data value may also be based on a predetermined minimum threshold data value derived from a statistical analysis of the patient's previously recorded medical data set. For example, an exemplary minimum threshold data value may be calculated / defined as 1.5 times the interquartile range below the 25% quartile (first quartile) of all carbohydrate and / or insulin values for a particular patient, requiring that said exemplary minimum threshold data value be non-negative, i.e., equal to or greater than zero. Again, said minimum threshold data value is merely exemplary, and other ranges / limits and / or other data, such as heart rate or other analyte data, may be used to derive a minimum threshold data value at which the underlying data may be removed from the first medical data set.
[0037] In other words, the predetermined minimum threshold data value may be based on determining the interquartile range below the 25% quartile of all available data values of a particular data type, including any previously / historically recorded data values for the patient.
[0038] Notably, an exemplary predetermined minimum threshold data value may be set to zero to exclude any negative data / negative data values that may indicate an error in the first medical data set.
[0039] An exemplary step of identifying data values that differ from a predetermined expected data value by more than a predetermined amount and removing data associated with said identified data values may include, for example, identifying unrealistic / suspicious carbohydrate-to-insulin ratios.
[0040] For example, the basis for such confirmation of an unrealistic / suspect carbohydrate to insulin ratio is the following formula: The expected injected bolus insulin amount, BI, which may be calculated according to TIFF2024513607000002.tif12104 expected where CHO is the carbohydrate content of the meal or the amount of carbohydrate in the diet, ΔG is the glucose or blood glucose level, BG, measured / recorded by a glucose monitoring system, e.g., by a continuous glucose monitoring system (CGM). CGM The nominal target glucose or blood glucose level, BG target , for example, BG set to 110 mg / dl target deviation from BG, i.e. △G=BG CGM -BG target It is.
[0041] For example, identifying a data value that differs from a predetermined expected data value by more than a predetermined amount may include determining whether the recorded bolus insulin amount differs from the expected bolus insulin amount by more than a predetermined amount, such as by more than 10, 20, or 40%.
[0042] As used herein, the term continuous glucose monitoring system (CGM) may be understood to refer to a glucose monitoring system that measures / records a patient's glucose values, e.g., blood glucose and / or interstitial glucose values, at regular intervals or scheduled times, and the frequency at which a patient's glucose level may be measured / recorded may be up to 12 measurements per hour, i.e., with a sampling time T of 5 minutes. S or the frequency can be higher, for example every minute or even with a sampling time of less than one minute.
[0043] These measurements may be made, for example, by a transcutaneous sensor implanted in the patient as part of a glucose monitoring system.
[0044] Furthermore, the CIR is a patient-specific carbohydrate-to-insulin ratio value or factor that is set, for example, by the patient himself or herself or with the help of a physician, and the ISF is a patient-specific insulin sensitivity factor that can also be set, for example, by the patient himself or herself or with the help of a physician.
[0045] Patients can rely on the CIR and ISF values to calculate their particular bolus insulin injection needs.
[0046] The expected bolus insulin amount BI defined above is expected is dependent on two ratios, CHO / CIR and ΔG / ISF. Abnormal deviations in any one of said ratios can be reflected in the measured / recorded injected bolus insulin amount BI value.
[0047] Using the above identified exemplary parameters, if the measured / recorded BI value differs from a predetermined expected injected bolus insulin amount or value by more than a predetermined amount or value, the data or data point or data subset or data segment of the first medical dataset may be identified / marked as suspicious / abnormal.
[0048] For example, if the measured / recorded BI value is greater than the expected BI expectedIf it is different by more than 10, 20, or 40%, the data or data points or data subsets or data segments of the first medical dataset can be marked for removal.
[0049] For example, to check whether the measured / recorded BI value of the first medical dataset is outside the following boundaries: TIFF2024513607000003.tif10108. l is a factor for setting the lower limit, u is a factor for setting the upper limit, and l < u. For example, the expected value BI expected If it is required to remove data from the first medical dataset that has a measured / recorded BI value different by more than 40% from the expected value, l may be set to 0.6 and u may be set to 1.4.
[0050] The step of identifying incomplete data with missing data values and removing the identified incomplete data from the first medical dataset may include, among other things, checking for missing time / missing timestamps of the data, and / or checking for incorrect data formats, e.g., the appearance of a string instead of an expected floating-point or integer value, and / or checking for missing data values of an expected data entity or data type.
[0051] The step of identifying at least one predetermined time-dependent data pattern and removing the data associated with the identified time-dependent data pattern may include, for example, checking for any unexplained or invalid increase or decrease in specific data values over time.
[0052] In other words, an unexpected or unexplained or invalid increase or decrease in specific data values over time can be regarded as a predetermined time-dependent data pattern or as an abnormal temporal signature in a subset of the first medical dataset.
[0053] For example, identifying / detecting invalid elevations in glucose levels or blood glucose values in data from a glucose monitoring system / device, e.g., data recorded / measured by a continuous glucose monitoring system / device, CGM, that cannot be associated with a meal intake (which may itself also be an indication of an incomplete data set / incomplete data entry) may be detected / identified using the following exemplary computer-executable methods and criteria: ● Filtering the first medical data set, in particular data relating to glucose levels or blood glucose values, e.g. time-dependent data from a glucose monitoring system / device, using a filter, e.g. a Savitzky-Golay Filter (SGF) having filter parameters d and w, where the filter parameter d refers to the polynomial order and the filter parameter w refers to the window size of the Savitzky-Golay Filter. This makes it possible to obtain a filtered time signal y(t) of the glucose level or blood glucose value. Finding all the minima and maxima of the filtered signal y(t). For each rising segment of the filtered signal y(t), for example for each minimum up to the next maximum of the filtered signal y(t), the following steps can be carried out: a) Identifying when: t min : The time when the minimum value appears in y(t) t max :t min The time when the next maximum occurs in y(t) after t1: the first derivative of y(t) TIFF2024513607000004.tif910 is in the interval [t min , t max ] the time when the maximum value t2: the second derivative of y(t) TIFF2024513607000005.tif910 is in the interval [t min , tmax ] the time when the maximum value t start =t min -ΔT, where ΔT is the time shift t end =t2+△T b) The difference in glucose level between the time of the maximum and the time of the minimum, △y = y(t max )-y(t min ) to determine c) The following conditions: TIFF2024513607000006.tif16122 (where c(t) denotes the corresponding data value for the amount of carbohydrate from a meal intake at a given time, i.e., the time-dependent carbohydrate input signal, and Δy min indicates the threshold for a minimum rise in glucose level / blood sugar level, TIFF2024513607000007.tif913 shows the percentage change threshold for changes in glucose levels / blood glucose levels) Determine whether the requirements are met.
[0054] If the above condition is met, then the time interval [t min , t max ] may be considered an invalid increase in glucose level or blood glucose (BG) value, and the data or data segment associated with the invalid increase may be marked as suspect and removed from the first medical data set.
[0055] Exemplary parameters and exemplary parameter values for the above exemplary possible steps of using a Savitzky-Golay-Filter (SGF) to identify predefined time-dependent data patterns, such as abnormal or suspicious or invalid elevations in glucose levels / blood glucose (BG) values, are listed in Table I below. TIFF2024513607000008.tif58164
[0056] In addition to the above exemplary set of rules and criteria for detecting incomplete or erroneous or invalid data or data points or data segments, further checks may be performed to identify data or data points or data segments in the first or second medical dataset that are particularly well suited as starting points or data segments to be input into the blood glucose prediction model to facilitate identification / determination of the best blood glucose prediction model parameters / parameter values.
[0057] For example, a suitable valid starting point in the first medical data set or the second medical data set may be identified as a meal intake of a predetermined minimum amount of carbohydrates, e.g. at least 20 g of carbohydrates, accompanied by a concomitant injection of an appropriate amount of bolus insulin, e.g. an amount of bolus insulin that is not marked as invalid or suspect and that is consistent with the expected injection amount of bolus insulin, e.g. as further described above.
[0058] A suitable data segment based on which best-fit model parameter values for the parameters of the blood glucose prediction model can be derived may be defined by starting from a valid starting point and ending with the last data point before a data point marked as suspicious or a hole or gap in the first medical data set, for example a gap longer than a predetermined period, for example longer than 30 minutes. If a segment ends with a data point marked as suspicious from which a data point may be removed, the last data point before the suspicious data point may be used to mark the end of the data segment.
[0059] Additionally or alternatively, the end of a valid data segment may be defined by identifying the next valid starting point marker, or the end of a valid data segment may be defined by the passage of a predetermined period of time after the start of the data segment, for example 2 hours after the first point of the data segment, whichever may occur first.
[0060] To further facilitate the use of data segments of the first or second medical data set to identify the best blood glucose prediction model parameters / parameter values, a minimum length / minimum duration can be set for the identification / selection of data segments of the first medical data.
[0061] For example, the step of providing the extracted second medical dataset as an input to the blood glucose level prediction model data may include identifying at least one data segment in the second medical dataset, the data segment being a subset of a plurality of data points of the extracted second medical dataset covering at least a minimum time range.
[0062] For example, it may be defined / required that a data segment must be at least 30 minutes long.
[0063] The above exemplary second medical dataset may be derived from the first medical dataset by removing data from the first medical dataset according to any of the above exemplary methods, criteria, or rules.
[0064] Additionally or alternatively, the above exemplary second medical dataset may be derived / generated from the first medical dataset by extracting valid data segments from the first medical dataset according to any of the above exemplary methods, criteria, or rules.
[0065] Furthermore, when the second medical dataset is derived from the first medical dataset, the second medical dataset may be divided according to a predetermined schedule, such as, for example, dividing the second medical dataset into three different portions associated with meal times, such as breakfast, lunch and dinner times.
[0066] Following such an exemplary division of the second medical dataset into different portions, each portion of the second medical dataset can be provided as a separate input to a blood glucose level prediction model to derive model parameters for each different portion, thereby better reflecting the effect of time of meal on the predictive performance of the model to predict future blood glucose levels.
[0067] For example, separate model parameter sets may be identified for breakfast, lunch, and dinner times: A breakfast model may be identified, for example, from data segments having start points between 5:30AM and 10:30AM, a lunch model may be identified from data segments having start times between 10:30AM and 2:30PM, and a dinner model may be identified from data segments having start times between 5:00PM and 9:00PM.
[0068] However, it is optional to split such second medical data into different portions and model the portions separately.
[0069] Internal trials and testing have unexpectedly and surprisingly found that when a first medical dataset is pre-processed according to any of the above exemplary methods, standards or rules to derive a second medical dataset that is a subset of the first medical dataset, and the derived second medical dataset is used as an input to a blood glucose prediction model for predicting the patient's future blood glucose levels, particularly for predicting the patient's post-prandial blood glucose levels, the robustness, performance, reliability and accuracy of the prediction of future blood glucose levels is improved to unprecedented quality levels.
[0070] This further advantageously allows for more accurate and more relevant warnings to be provided to the patient regarding critical blood glucose levels, and for more accurate recommendations to administer appropriate doses of insulin, based on which, for example, an insulin pump can be controlled.
[0071] Furthermore, it has been found that when a first medical dataset is pre-processed according to any of the above exemplary methods, criteria, or rules to derive a second medical dataset that is a subset of the first medical dataset, the second medical dataset significantly facilitates the identification of the best blood glucose prediction model and the identification of the best model parameters. The ability to more easily identify the best blood glucose prediction model(s) not only improves results for predicting a patient's future glucose levels, but also improves results for modeling the overall behavior or overall trajectory of the patient's glucose level changes.
[0072] To further support some technical aspects of the present invention, a possible exemplary blood glucose prediction model that can be used to predict a patient's future blood glucose level based on the above second medical dataset is described below.
[0073] For example, a suitable patient data based blood glucose prediction model describes the glycemic response to carbohydrate intake and a bolus insulin injection and is based on the following equation: This is the Kirchsteiger model (denoted PM1) which can be represented as TIFF2024513607000009.tif12122.
[0074] In this formula, BG, which can also be represented by CGM(s), CGM where (s) describes the glucose level(s) / blood glucose value(s) measured / recorded via a continuous glucose monitoring (CGM) system, D(s) carbohydrates of meal intake, and U(s) bolus insulin injection, all in the Laplace domain parameterized by a complex frequency s, and the transformation from the time domain t to the Laplace domain s is given by the following Laplace transform The image can be described by TIFF2024513607000010.tif1152.
[0075] Other influencing inputs such as basal insulin, stress, sports, mixed meal composition, etc. are not incorporated into this model structure.
[0076] The parameters in equation (4) have easy physiological interpretation: K1 describes the effect of 1 gram of carbohydrate on glucose levels, whereas K2 corresponds to the effect of 1 IU of bolus insulin (both for time t→∞). The time constants T1 and T2 are proportional to the response times to carbohydrate and insulin inputs.
[0077] In the Kirchsteiger model according to equation (4), carbohydrate and insulin inputs are assumed to have a sustained effect on glucose levels caused by the fact that both transfer functions contain integral terms (pole at s=0). Such behavior is not strictly physiological, but has some similarities to the simple heuristic of Advanced Carbohydrate Counting (ACC) used to calculate the need for meal bolus insulin. The parameter K2 has the same interpretation as the insulin sensitivity factor ISF, while the ratio K2 / K1 indicates how many grams of carbohydrate are compensated by 1 IU, as is the case for the carbohydrate-to-insulin ratio CIR in ACC.
[0078] To identify best-fit model parameters / parameter values of a selected patient data-based blood glucose prediction model, such as the Kirchsteiger model described above, a second medical data set can be input into the patient data-based blood glucose prediction model, and the best-fit model parameters / parameter values can be determined by minimizing a cost function.
[0079] A possible suitable exemplary cost function to optimize / minimize to determine best-fit model parameters for a blood glucose predictive model, such as the Kirchsteiger model described above, is the following exemplary cost function J(θ): TIFF2024513607000011.tif20113, where: The image is JPEG2024513607000012.jpg18115.
[0080] In this cost function, y k corresponds to the measured output of a glucose monitoring system, such as a continuous glucose monitoring system, i.e., y k Glucose Data / Blood Glucose Data / Blood Glucose BG CGM Corresponding to, TIFF2024513607000013.tif99 shows the output of a selected blood glucose prediction model, e.g., the Kirchsteiger model.
[0081] Furthermore, d denotes a data segment, e.g., a data segment from a second medical dataset, identified from the first medical dataset as described above. In other words, a data segment may be understood as a subset of the first medical dataset, the data of the data segment being extracted / selected / filtered from the first medical dataset according to one or more criteria and / or rules described above.
[0082] The total number of data segments is d tot That is, 1...d tot There are data segments. In order to make the cost function expression easier to read, the index of the data segments is intentionally omitted in the above definition of the cost function J(θ). However, d can be expressed as an index of itself, e.g., d m (m=1,...d tot ).
[0083] Each data segment d has a start index k0 and an end index k N i.e., the index k can be used to index individual data points / individual data within a given data segment d.
[0084] The vector θ describes the model parameters, i.e., in the exemplary case of the Kirchsteiger model, said model vector θ may depend on the parameters K1, K2, T1 and T2.
[0085] Model Output TIFF2024513607000014.tif99 is the model parameter θ and the initial state It can be calculated using the estimate of TIFF2024513607000015.tif99.
[0086] This initial state estimate may correspond to the state of the model at the start of the data segment, k0, and may correspond to a best estimate of the glucose level value at the start of the data segment. The initial state estimate may be estimated / derived / determined for each data segment.
[0087] As will be further exemplarily explained below, the initial state estimate may be derived, for example, by using a Kalman filter and / or by using an autoregressive approach.
[0088] The same model parameter vector θ can be used for each data segment of a patient.
[0089] Minimization of this exemplary cost function may be performed using, among other techniques, a brute force search algorithm over a grid of model parameter vector θ values, or by a local gradient search algorithm, such as the simplex algorithm, or by a global optimization algorithm, such as a genetic algorithm, a simulated annealing algorithm, a Markov Chain Monte Carlo algorithm, or other techniques.
[0090] The derived best-fit model parameters θ best may be used to predict a patient's future blood glucose levels, and in particular, to predict a patient's post-prandial blood glucose levels.
[0091] In particular, the derived best-fit model parameters θ bestmay be used together with further information from data from a second medical dataset, such as meal size and / or injection bolus volume, for example, to simulate the patient's future glucose trajectory after meals.
[0092] As described above, it has been found that using a blood glucose prediction model such as the example above on a medical dataset extracted / selected / generated from the first medical dataset, i.e., the second medical dataset, results in unprecedented robust, reliable and accurate prediction of a patient's future blood glucose levels, superior to any currently known predictive techniques.
[0093] Optionally, the application of the above blood glucose prediction model may be further refined by using a Kalman filter to estimate the initial states of the model.
[0094] For example, a Kalman filter can be used to estimate the state of the model prior to the start of the (first) data segment, for example 6 hours prior to the start of the (first) data segment, and calculate estimates of the state at each of those times up to the start of the first data segment d.
[0095] Therefore, The last estimate (at just the beginning of the identification data segment) in TIFF2024513607000016.tif99 is the initial state of the blood glucose prediction model. It may correspond to TIFF2024513607000017.tif99.
[0096] The model in the Kalman filter (derived from the model with parameters θ) may be updated at every iteration step of the optimization using the latest estimates of the model parameters θ, i.e., the model used for prediction and the (same) model used in the Kalman filter may be optimized simultaneously.
[0097] A new Kalman filter may be instantiated and applied at each iteration step, time Input (d, u) and output (△y=yG b ) data based on each hour State estimates in TIFF2024513607000019.tif99 TIFF2024513607000020.tif913 can be used to calculate G b may be an estimate of the patient's basal glucose level, and u is the time domain input.
[0098] time Estimated state in TIFF2024513607000021.tif99 Using TIFF2024513607000022.tif913 as the initial condition, we can calculate a simulation with the process model k steps into the future. TIFF2024513607000023.tif922.
[0099] Therefore, time TIFF2024513607000024.tif919(T S is the sampling time), the prediction can be done, for example, by It can be calculated as TIFF2024513607000025.tif12113.
[0100] The combination of the exemplary Kirchsteiger model (PM1) and the exemplary Kalman filter (KF) may be referred to as KF-PM1.
[0101] If we do not use a Kalman filter, i.e., simply use the basic Kirchsteiger model (PM1), then the above equation (8) also becomes Estimated state in TIFF2024513607000026.tif99 TIFF2024513607000027.tif913 can be used to predict future glucose levels by setting it to zero.
[0102] Alternatively, a hybrid approach may be adopted, where the initial state TIFF2024513607000028.tif99 may be assumed to be zero for the model (i.e., the initial conditions have no effect on the model output), but the effect of the initial conditions may instead be captured by an autoregressive (AR) model.
[0103] The predicted glucose output of this exemplary hybrid model approach may correspond to the sum of the selected blood glucose prediction model prediction(s) and the prediction(s) from the AR model.
[0104] For this purpose, a population-averaged AR model with parameter values identified from nighttime (i.e., unaffected by food) data may be used.
[0105] TIFF2024513607000029.tif935, max, time Information from TIFF2024513607000030.tif99 given time This is the prediction for TIFF2024513607000031.tif919.
[0106] The composite model output of the predicted glucose trajectory is then calculated using the following formula: TIFF2024513607000032.tif10125, where: TIFF2024513607000033.tif920 is the initial condition This is the output of a simulated blood glucose prediction model using the assumptions in TIFF2024513607000034.tif929.
[0107] The combination of the exemplary Kirchsteiger model (PM1) and the exemplary autoregressive model (AR) may be referred to as the AR-PM1 approach.
[0108] Although the blood glucose prediction model based on the Kirchsteiger model has already provided very satisfactory performance results when used with the second medical dataset, as mentioned above, it is also possible that the second medical dataset can be provided as input not only to the exemplary model PM1 or its variants, KF-PM1 or AR-PM1, but also to other blood glucose prediction models.
[0109] Further, in addition to or as an alternative to the above steps for identifying the best model and / or the best model parameters and / or the best model parameter values, it is also possible to use a neural network algorithm or similar algorithm trained on the exemplary extracted second medical data to identify the best blood glucose level prediction model and / or the best model parameters and / or the best model parameter values.
[0110] For completeness, the following further possible examples of other blood glucose prediction models that can be used are exemplarily described below: It is also possible to simultaneously feed the second medical data set to multiple different blood glucose prediction models to compare, measure and calibrate their relative performance.
[0111] ● Zero-Order Hold: The Zero-Order Hold (ZOH) model estimates the BG up to time t by holding the last available value constant. CGM Based on data y, time t+kT S Predictions in Calculate TIFF2024513607000035.tif99, i.e. The file is TIFF2024513607000036.tif10118.
[0112] ● Global AR model: A simple second-order global autoregressive AR model of the form TIFF2024513607000037.tif27118 can be used (wherein, b is again a patient-specific estimate of the basal glucose level). k , b k) is optimized using least squares (LS) optimization for each prediction horizon k.
[0113] Note that the structure of the global AR model used for this purpose may be identical to that in the hybrid AR-PM1 approach described above, although the model parameters may differ.
[0114] While the AR models in the AR-PM1 context are identified only from nighttime data, the global AR models used as baselines may be based on data from the entire 24-h period of each day.
[0115] • Alternative model structures: To put the predictive performance of the PM1 model into context, an alternative model structure without the integral term is further analyzed. TIFF2024513607000038.tif12123
[0116] This alternative model may be called PM2 and may be implemented using a Kalman filter or using an autoregressive (AR) model, in which case the variant may be called KF-PM2 or AR-PM2.
[0117] Whatever blood glucose prediction model is selected to obtain the patient's future blood glucose levels based on the second medical dataset, the obtained prediction of the future blood glucose levels may be displayed to the patient, for example on a display of the glucose monitoring system or on a display of a computing device, for example a smartphone or personal computer.
[0118] Additionally, a recommended dose of insulin to be administered can be displayed to the patient based on the obtained prediction of future blood glucose levels.
[0119] Additionally, based on the obtained predictions of future blood glucose levels, a recommended dose of insulin can be automatically administered via automatic control of the insulin pump.
[0120] In other words, the present invention facilitates the monitoring, prediction and control of a patient's blood glucose levels, and also facilitates the control of administering appropriate doses of insulin in a reliable and accurate manner.
[0121] All method steps described herein for predicting blood glucose levels are computer-executable, i.e., may be performed by, for example, a computing system comprising a computer memory, one or more processors, and optionally a display.
[0122] The computing system may, for example, be of one of the following types: a computer server, a personal computer, or a mobile computing system, such as a smartphone, a tablet, or a laptop.
[0123] Furthermore, the present invention may be implemented as a glucose monitoring system comprising a sensor for acquiring glucose / blood glucose data of a patient, computer memory, one or more processors, and a display, and the computer memory of the glucose monitoring system may include computer-executable instructions that, when executed by the one or more processors of the glucose monitoring system, cause the one or more processors to perform one, some, or all of the steps described herein for predicting blood glucose levels.
[0124] The exemplary glucose monitoring system may further comprise an insulin pump, and the glucose monitoring system may be further configured to control the insulin pump, particularly to control the dosage of insulin that may be administered by the insulin pump.
[0125] An exemplary glucose monitoring system may also be configured to determine a recommended dosage of insulin to be administered based on the obtained predictions of future blood glucose levels.
[0126] Furthermore, the computer-executable method steps described above and below may be stored on one or more computer storage media, such as non-volatile computer storage media, that store computer-executable instructions that, when executed by a computer system, can perform a method according to any one, some, or all of the exemplary method steps described above and below for predicting blood glucose levels. [Brief description of the drawings]
[0127] The following figures are only intended to illustrate the following technical aspects.
[0128] [Figure 1] 1 is an exemplary flow diagram for predicting blood glucose levels. [Diagram 2] Exemplary filtered blood glucose data / exemplary filtered glucose data. [Diagram 3] 13 is an exemplary prediction of post-prandial blood glucose levels. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0129] FIG. 1 shows in a schematic manner exemplary possible steps of a computer-executable method (100) for predicting blood glucose levels, in particular for post-prandial blood glucose level prediction.
[0130] Here, an exemplary first medical data set of a patient covering a time range may be acquired or received for processing (101).
[0131] The exemplary first medical data set includes blood glucose data and further other medical data of the patient.
[0132] For example, the blood glucose data may originate from / be received from / be obtained from a sensor of the glucose monitoring system, and the other medical data may include at least one of data regarding a meal intake, e.g., an amount of carbohydrates from a meal intake, or data regarding a medication, e.g., data regarding an insulin injection or measured insulin levels, and / or other analyte data, where the exemplary other analyte data or insulin data may also originate from a medical sensor.
[0133] Alternatively or additionally, the patient may enter data themselves, for example data relating to dietary intake, for example amount of carbohydrates from dietary intake, or data relating to medications, blood glucose or insulin levels.
[0134] In other words, the first medical dataset may be a hybrid dataset of data automatically acquired from medical sensors and data manually entered by the patient / user.
[0135] To generate a second medical dataset, the exemplary second medical dataset may be extracted (102) from the exemplary first medical dataset based on at least one of the following rules or criteria: - identifying (103) duplicates in the first medical data set and removing the identified duplicates; and / or - identifying (104) data values that exceed a predetermined maximum threshold data value; or - identifying (105) data values that are below a predetermined minimum threshold data value and removing data associated with said identified data values; and / or identifying (106) data values that differ from predetermined expected data values by more than a predetermined amount and removing data associated with the identified data values; and / or Identifying (107) incomplete data that is missing data values and removing the identified incomplete data; and / or • Identifying (108) at least one predefined time-varying data pattern and removing data associated with said identified time-varying data pattern.
[0136] The exemplary second medical dataset generated / extracted from the first medical dataset may then be automatically fed / provided / input into the blood glucose prediction model (109).
[0137] Any of the above or other blood glucose prediction models may be used on the extracted second medical data set.
[0138] Based on the output of the blood glucose prediction model based on the second medical dataset, the patient's future blood glucose levels, in particular the patient's post-prandial blood glucose levels, may be predicted 110 with unprecedented accuracy and robustness.
[0139] FIG. 2 illustrates the process of filtering (raw, unfiltered) data on blood glucose levels (201), e.g., time-dependent data BG from a glucose monitoring system, using a filter, e.g., the Savitzky-Golay-Filter (SGF) described above, and with the exemplary parameters of Table I. CGM 2A illustrates an exemplary filtered blood glucose data signal y(t) (200) obtained from filtering glycemic index y(t).
[0140] Here, reference numeral 203 indicates an exemplary vertical axis of glucose or blood glucose value BG, and reference numeral 202 indicates an exemplary horizontal axis of time t.
[0141] As mentioned above, exemplary identification / detection of invalid elevations in blood glucose levels in data from a glucose monitoring system, e.g., data recorded / measured by a continuous glucose monitoring system, CGM, that cannot be associated with a meal intake (which may itself also be an indication of an incomplete data set / incomplete data entry) may be detected / identified using the following exemplary computer-executable methods and criteria: - Finding 200 all minima and maxima of the filtered signal y(t), where the exemplary illustrated signal (200) has two maxima and one minimum. For each rising segment of the filtered signal y(t), for example for each minimum up to the next maximum of the filtered signal y(t), the following steps can be carried out: a) Identifying when: t min : The time when a minimum value appears in y(t) (207) t max :t min The time at which the next maximum value appears in y(t) after t1: the first derivative of y(t) TIFF2024513607000039.tif910 is in the interval [t min , t max ] the time when the maximum value is t2: the second derivative of y(t) TIFF2024513607000040.tif910 is in the interval [t min , t max ] the time when the maximum value is t start (204):t start =t min -ΔT, where ΔT(211) is the time shift t end (205), t end =t2+△T b) The difference in glucose level between the time of the maximum and the time of the minimum, Δy (210): Δy = y(t max )-y(t min ) to determine c) The following conditions: where c(t) denotes the corresponding data value for the amount of carbohydrate from a meal intake at a given time, i.e., the time-dependent carbohydrate input signal, and Δy min indicates the threshold for a minimum rise in glucose level / blood sugar level, TIFF2024513607000042.tif913 shows the percentage change threshold for changes in glucose levels / blood glucose levels) Determine whether the requirements are met.
[0142] If the above condition is met, then the time interval [t min , t max ] (212) may be considered an invalid increase in glucose level / blood glucose (BG) value, and the data or data segment associated with the invalid increase may be marked as suspect and removed from the first medical data set.
[0143] Identifying such an exemplary invalid increase in glucose level / blood glucose (BG) value is one example of identifying a predetermined time-dependent data pattern such that data associated with the identified time-dependent data pattern may be removed from the first medical data set.
[0144] Figure 3 shows the glucose data / blood sugar data of a patient. CGM Illustrated diagrammatically by way of example is (t) (304), where the data may be obtained from a continuous glucose monitoring system.
[0145] The presented glucose data / blood glucose data (304) may be glucose data / blood glucose data from an exemplary second medical dataset, i.e., glucose data / blood glucose data extracted from the first medical dataset according to at least one of the above steps, rules, or criteria.
[0146] As in the previous figure, reference numeral 303 indicates an exemplary vertical axis of blood glucose level BG, and reference numeral 302 indicates an exemplary horizontal axis of time t.
[0147] Stars indicate events on the timeline (307) such as food intake and / or drug intake, for example bolus insulin intake.
[0148] For example, event (307) may represent a patient's breakfast meal event. Additionally, two prediction trajectories for predicting postprandial blood glucose levels from two different blood glucose prediction models based on the exemplary second medical dataset are shown.
[0149] Here, reference numeral 305 denotes a prediction of postprandial blood glucose level based on providing a second medical data set to a blood glucose level prediction model based on a Kirchsteiger model with a Kalman filter applied, i.e., the KF-PM1 model, and reference numeral 306 denotes a prediction of postprandial blood glucose level based on providing a second medical data set to a blood glucose level prediction model based on a Kirchsteiger model combined with an autoregressive model, i.e., the AR-PM1 model.
[0150] As exemplified above, the second medical data set may be fed into a blood glucose prediction model and the model parameters may be optimized, for example by minimizing a cost function, to derive best-fit model parameters that best fit the data and enable prediction of the patient's future blood glucose levels.
Claims
1. A method (100) for predicting blood glucose levels, said method being computer-implemented, and Receiving (101) a first medical dataset of a patient covering a time range, the first medical dataset comprising glucose data and further other medical data of the patient; Extracting (102) a second medical data set from the first medical data set, the second medical data set being a subset of the first medical data set, the extracting comprising: identifying (103) duplicates in the first medical dataset and removing the identified duplicates; identifying (104) data values that are above a predetermined maximum threshold data value or identifying (105) data values that are below a predetermined minimum threshold data value and removing data associated with the identified data values; identifying (106) data values that differ from a predetermined expected data value by more than a predetermined amount and removing data associated with the identified data values; identifying (107) incomplete data having missing data values and removing the identified incomplete data; Identifying (108) at least one predetermined time-dependent data pattern and removing data associated with the identified time-dependent data pattern. Extracting (102) a second medical data set including at least one of: Providing (109) the extracted second medical data set as an input to a blood glucose prediction model; predicting (110) future blood glucose levels of the patient using an output of the blood glucose prediction model based on the second medical data set; A method (100).
2. The method (100) of claim 1, wherein the blood glucose value is a postprandial blood glucose value.
3. The method (100) of claim 1, wherein the further other medical data of the patient comprises at least one of data regarding dietary intake, or data regarding medications, and / or other analyte data.
4. The method (100) of claim 1, wherein the further other medical data of the patient includes at least one of data regarding the amount of carbohydrates from dietary intake, or data regarding insulin injections, and / or other analyte data.
5. 2. The method of claim 1, wherein identifying overlaps in the first medical data set is based on determining whether the overlaps are within a time period of less than five minutes.
6. 2. The method of claim 1, wherein identifying data values above a predetermined maximum threshold data value is based on the predetermined maximum threshold data value derived from a statistical analysis of a previously recorded medical data set of the patient.
7. 7. The method (100) of claim 6, wherein the predetermined maximum threshold data value is based on determining an interquartile range above the 75% quartile of all available data values for carbohydrate amount and / or insulin amount.
8. 2. The method (100) of claim 1, wherein providing the extracted second medical dataset as an input to a blood glucose level prediction model comprises identifying at least one data segment in the second medical dataset, the data segment being a subset of a plurality of data points of the extracted second medical dataset covering at least 30 minutes.
9. 2. The method (100) of claim 1, wherein identifying a data value that differs from a predetermined expected data value by more than a predetermined amount comprises determining whether the recorded bolus insulin amount differs from an expected bolus insulin amount by more than a predetermined amount.
10. The method (100) of claim 1, wherein identifying data values that differ from a predetermined expected data value by more than a predetermined amount includes determining whether the recorded bolus insulin amount differs from the expected bolus insulin amount by more than 10, 20, or 40%.
11. 2. The method of claim 1, wherein identifying at least one predetermined time-dependent data pattern and removing data associated with the identified time-dependent data pattern comprises detecting invalid elevations in glucose levels.
12. Providing (109) the extracted second medical data set as an input to a blood glucose prediction model comprises: providing the extracted second medical data set as a training data set for a blood glucose prediction model algorithm; and training the blood glucose prediction model algorithm with the extracted second medical data set; And / or the method (100) of claim 1 , wherein the blood glucose prediction model is based on the Kirchsteiger model.
13. The method (100) of claim 12, wherein the blood glucose prediction model algorithm is a neural network algorithm.
14. 2. The method (100) of claim 1, wherein the obtained prediction of future blood glucose levels is displayed to the patient, and / or a recommended dose of insulin to be administered based on the obtained prediction of future blood glucose levels is displayed to the patient, and / or the recommended dose of insulin is automatically administered via automatic control of an insulin pump based on the obtained prediction of future blood glucose levels.
15. 1. A computing system comprising: Computer memory; One or more processors; a display; Equipped with the computing system is configured to receive a first medical data set of a patient; 15. A computing system, the computer memory comprising computer-executable instructions that, when executed by the one or more processors, cause the one or more processors to perform the method (100) of any one of claims 1 to 14 for predicting blood glucose levels.
16. 16. The computing system of claim 15, wherein the computing system is one of the following types: a computer server, a personal computer, or a mobile computing system.
17. The computing system of claim 16, wherein the mobile computing system is a smartphone, a tablet, or a laptop.
18. 1. A glucose monitoring system comprising: a sensor for acquiring glucose data of the patient; A computer memory, one or more processors, and a display. Equipped with 15. A glucose monitoring system, wherein the computer memory includes computer executable instructions that, when executed by the one or more processors, cause the one or more processors to perform a method (100) according to any one of claims 1 to 14 for predicting blood glucose levels.
19. 20. The glucose monitoring system of claim 18, further comprising an insulin pump, the glucose monitoring system further configured to control the insulin pump, and the glucose monitoring system configured to determine a recommended dosage of insulin to be administered based on the obtained prediction of future blood glucose levels.
20. The glucose monitoring system of claim 18, further comprising an insulin pump, wherein the glucose monitoring system is further configured to control the insulin pump to control a dosage of insulin that may be administered by the insulin pump, and wherein the glucose monitoring system is configured to determine a recommended dosage of insulin to be administered based on the obtained prediction of future blood glucose levels.
21. A computer readable storage medium for storing computer executable instructions which, when executed by a computer system, perform the method (100) of any one of claims 1 to 14 for predicting blood glucose levels.