WO2022234032A2

WO2022234032A2 - Methods and systems for estimating fasting glucose values

Info

Publication number: WO2022234032A2
Application number: PCT/EP2022/062183
Authority: WO
Inventors: Henrik Bengtsson; Tina Björk ARADÓTTIR; Sarah Ellinor ENGELL; Zeinab MAHMOUDI; Thomas Emil RYDE; Bo Liang
Original assignee: Novo Nordisk A/S
Priority date: 2021-05-05
Filing date: 2022-05-05
Publication date: 2022-11-10
Also published as: WO2022234032A3

Abstract

A method for estimating a fasting blood glucose (FBG) value for a subject, comprising the steps of obtaining a first data set representing a blood glucose history (BGH), obtaining a second data set representing a basal insulin injection history (IH) of the subject, and based on BGH and IH and using a mathematical model: calculating a predicted FBG for the time t+1, where t+1 is the current time. The method comprises the further steps of obtaining at time t+1 a current measured FBG from the subject, and based on the predicted FBG and the current measured FBG: calculating an updated estimated FBG.

Description

METHODS AND SYSTEMS FOR ESTIMATING FASTING GLUCOSE VALUES

The present disclosure generally relates to systems and methods for assisting patients and health care practitioners in managing insulin treatment to diabetics. In a specific aspect the present invention relates to systems and methods suitable for use in a diabetes management system in which a fasting blood glucose value is to be determined for a patient in treatment with an insulin.

BACKGROUND

Diabetes mellitus (DM) is impaired insulin secretion and variable degrees of peripheral insulin resistance leading to hyperglycaemia. Type 2 diabetes mellitus is characterized by progressive disruption of normal physiologic insulin secretion. In healthy individuals, basal insulin secretion by pancreatic b cells occurs continuously to maintain steady glucose levels for extended peri ods between meals. Also in healthy individuals, there is prandial secretion in which insulin is rapidly released in an initial first-phase spike in response to a meal, followed by prolonged insulin secretion that returns to basal levels after 2-3 hours. Years of poorly controlled hyper glycaemia can lead to multiple health complications. Diabetes mellitus is one of the major causes of premature morbidity and mortality throughout the world.

People suffering from diabetes have been treated with insulin for more than a century. Diabet ics are depending on insulin products to prolong their lives and increase their life quality. In jections of insulin are by far the most used method of diabetics to take their insulin. Different types of diabetes, different stages of each type and different life-styles and individual physio logical differences, all contribute to complicate optimisation of treatment regimens and as a result modern treatment is based on highly personalised treatment regimens based on exten sive research and development as well as carefully supervised introduction of treatment by titration.

Development of treatment is not only focussed on increasing the efficiency and prolonging the life of diabetics, but also on postponing and ultimately avoiding the serious adverse long term effects of the condition, as well as making the treatment itself easier and less of a daily burden for diabetics.

It is known that the better the insulin management fit with the actual need, the better the effi ciency and the longer adverse effects of the condition can be postponed. Furthermore, severe over- or under-dosing can be immediately life threatening and must be avoided. Thus, large fluctuations of insulin levels in the body should be limited to occur only when needed, which is most often related to ingestion of meals and exercising. The better the natural regulation of insulin in the body of a non-diabetic can be mimicked, the better, obviously. An obvious tech nical solution would be to measure the blood glucose level, determine the need for insulin, and inject the required amount at intervals of a few minutes. However, this would be extremely expensive and not improve the life quality of the diabetics, as it would become a full-time oc cupation to regulate blood glucose/insulin levels.

To decrease the number of required injections and thereby the interference with the diabetic’s life and routines, different types of insulin have been developed to allow treatment of different types of diabetes and to allow combination of products to optimise treatment regimens on an individual basis. In particular, products of long-lasting effect of slowly releasing insulin in the user’s body has been developed to provide a basic level of insulin, called basal treatment. This ensures insulin levels do not get too low during sleep and outside meal-situations.

As a new user is introduced to treatment with insulin, the user has to go through a period of titration. For insulins for use at fixed intervals, e.g. once daily or once weekly, the preferred titration method is the 3-FPG method, according to which the FPG (Fasting Plasma Glucose, being a measurement of the plasma glucose concentration after 8 hours of fasting) is meas ured on titration day and the prior two days. The measurements are done as SMPG (Self Measured Plasma Glucose) values and should preferably be in the range between 4.4 mmol/L - 7.2mmol/L. A mean value of these three measurements is used to avoid inefficient titration due to bad measurements and random variations in plasma Glucose variations caused by unusual activities of the user.

Thus, if Sunday is titration day for a new user of a once weekly insulin product, the user is required to obtain SMPG values Friday, Saturday and Sunday and calculate a mean value. If the value is above the target area, the dose should be increased and if it is below, the dose should be decreased. This process continues until the weekly measurements becomes stabile and within target range.

However, such measurements are uncomfortable for the users as they are not only required to prick a finger to obtain blood on which to measure, but also required to be fasting 8 hours each Friday, Saturday and Sunday.

Furthermore, the current titration method and target range is a best compromise for a popula tion based on a “one rule fits all” approach. As the PG values vary during the day, depending on the activities of the individual and the individual itself, it would be expected that users would experience increase stability of BG levels and better treatment outcomes if it was possible and feasible to provide a more personalised titration procedure and treatment.

Correspondingly, it is an object of the present invention to provide methods and systems providing that a titration process can be performed requiring less fasting and less finger pricks to obtain SMPG values. It is a further object to provide better estimates of dose adjustments during a titration process to reduce variations and time of obtaining a stabile dose setting. It is a yet further objective to suggest solutions providing more personalized titration to improve outcome of treatment.

DISCLOSURE OF THE INVENTION

In the disclosure of the present invention, embodiments and aspects will be described which will address one or more of the above objects or which will address objects apparent from the below disclosure as well as from the description of exemplary embodiments.

In a first aspect of the invention a system for estimating a fasting blood glucose (FBG) value for a subject is provided, wherein the system comprises one or more processors and a memory, the memory comprising instructions that, when executed by the one or more proces sors, perform a method responsive to receiving a request for an updated predicted FPG value. The method comprises the steps of obtaining a first data set, comprising a plurality of glucose measurements of the subject taken over a time course and thereby establish a blood glucose history (BGH), each respective glucose measurement in the plurality of glucose measurements comprising (i) a plasma glucose (BG) level, and (ii) a corresponding BG timestamp represent ing when in the time course the respective glucose measurement was made. The method comprises the further steps of, based on BGH and IH and using a mathematical model: calcu lating a predicted FBG for the time t+1, where t+1 is the current time, obtaining at time t+1 a current measured FBG from the subject, and based on the predicted FBG and the current measured FBG: calculating an updated estimated FBG.

By this method it is possible to a high degree to estimate a current FBG value based on only a single current SMBG value. Indeed, for an initial period the estimated FBG value will be less precise or additional SMBG determinations would have to be performed.

The method may comprise the further step of obtaining a second data set, comprising a basal insulin injection history (IH) of the subject, wherein the injection history comprises a plurality of injections during all or a portion of the time course and, for each respective injection in the plurality of injections comprising (i) a corresponding injection amount, and (ii) an injection timestamp representing when in the time course the respective injection occurred, wherein the predicted FBG is calculated based on BGH and IH.

When in the present context the term “BG” as in FBG (fasting plasma glucose) is used this also covers the term “PG” as in FPG (fasting blood glucose). Although the values are not iden tical differing by approximately 11% these terms are often used interchangeably.

The updated estimated FBG may be calculated as a weighted average of the predicted FBG and the current measured FBG. The weight may be based on the distributions of v_t and w_t, where v_t denotes measurement noise when determining the current measured FBG, and w_t is a stochastic term representing the physiological variations of the FBG. The weight may be dynamic and influenced by one or more variables. For example, the weight can be designed to mitigate the effects of outliers, e.g. caused by the incorrect use of SMBG, and is chosen to reflect the domain knowledge about the noise distributions of v_t and w_t.

In an exemplary embodiment the system is adapted to calculate a predicted FBG for the time t+1 using a plurality of mathematical models, each mathematical model representing a sub group of a population of diabetic subjects. The method comprises the steps of for the subject, selecting at least one mathematical model being representative for the subject, based on BGH and IH and for each selected mathematical model: calculating a predicted FBG for the time t+1, and based on the at least one predicted FBG and the current measured FBG: calculating an updated estimated FBG.

The updated estimated FBG may be calculated as a weighted average of the at least one predicted FBG and the current measured FBG. In an exemplary embodiment the updated es timated FBG is calculated as a weighted average of the predicted FBG and the current meas ured FBG: updated estimated FBG = Wi (predicted FBG) + W2(current measured FBG), wherein Wi + W2 = 1 , and Wi and/or W2 are functions of respective FBG values, the W values being negatively correlated with rising FBG values, i.e. low measurements weigh higher than high measurements. In a further aspect of the invention, methods per se for estimating a fasting plasma glucose (FBG) value for a subject as described above for a system are provided, the methods corre spond to the above-described methods performed by a system.

In a yet further aspect of the invention the above-described systems are adapted to provide a long-acting or ultra-long-acting insulin adjustment dose recommendation (ADR) for a subject to treat diabetes mellitus, the memory further comprising instructions that, when executed by the one or more processors, perform a method responsive to receiving a dose guidance re quest (DGR). The method comprises the further steps of: obtaining (A) a first data structure, comprising: (i) a glucose upper target range level (UTR) of the subject, and (ii) a glucose lower target range level (LTR) of the subject, obtaining (B) a second data structure, comprising: (i) a current dose guidance baseline (DGB), wherein the current DGB corresponds to (a) a most recent ADR, or (b) a starting basal dose (SBD), and providing the long-acting or ultra-long- acting insulin ADR, the recommendation being calculated based on data from the first data structure, the second data structure, and the updated estimated FBG.

In an exemplary embodiment the method comprises the further step of calculating a quality measure (QM) for a period of time based on (i.e. fully or partly) one or more of: the variance in FBG, the difference between predicted and measured FBG, and the difference between IH injection amounts and historic ADR. If the QM is at or below a given level: provide the calcu lated insulin ADR, or if the QM is above the given level: provide a reduced calculated insulin ADR.

In a further exemplary embodiment the method comprises the further steps of calculating a Personalised Target Measure (PTM) based on one or more of: the variance in FBG, the vari ance in minimum average BG for a moving average filter of a predetermined number of hours, and the difference between predicted and measured FBG. If the PTM is lower than a given average for the subject, this indicating that the risk for insulin induced hypoglycaemia is lower, then adjusting the target range level to a lower level than defined by the first data structure.

The system adapted to provide an ADR may comprise instructions in the memory performing the further steps of: obtaining from the subject for a period of time a continuous glucose mon itoring (CGM) data set comprising: (i) a plurality of BG levels, and (ii) a corresponding BG timestamp representing when in the time course the respective CGM measurement was made, calculating for the period of time a hypo-risk value based on the number and/or severity of hypoglycaemic events in the CGM data set, and if the hypo-risk value is below a given value, lower the LTR for the subject. Correspondingly, if the hypo-risk value is below a given value, also the UTR may be lowered for the subject.

In a further aspect of the invention, methods per se for providing a long-acting or ultra-long- acting insulin ADR for a subject to treat diabetes mellitus as described above for a system are provided, the methods correspond to the above-described methods performed by a system.

In a further aspect of the invention a system for estimating a fasting blood glucose (FBG) value for a subject is provided, wherein the system comprises one or more processors and a memory. The memory comprises instructions that, when executed by the one or more proces sors, perform a method responsive to receiving a request for an updated predicted FBG value, the method comprising the steps of: (A) obtaining a first data set, comprising a plurality of glucose measurements of the subject taken over a time course and thereby establish a blood glucose history (BGH), each respective glucose measurement in the plurality of glucose meas urements comprising: (i) a plasma glucose (BG) level, and (ii) a corresponding BG timestamp representing when in the time course the respective glucose measurement was made, and (B) obtaining at time t+1 a current measured FBG from the subject, where t+1 is the current time. The method comprises the further steps of, based on BGH and the current measured FBG from the subject and using a mathematical model: calculating a predicted FBG for the time t+1 , and determining an estimated FBG for the time t+1 as (i) the calculated predicted FBG if the calculated predicted FBG is lower than the current measured FBG, or (ii) the current measured FBG if the calculated predicted FBG is higher than the current measured FBG.

In a yet further aspect of the invention a method for aiding in the classification of a subject as either hypo-prone or non-hypo-prone is provided, comprising the steps of: obtaining individual population data for a plurality of subjects, the data comprising values from one or more of the following groups of data: BG values, blood characterizing values, demographic data, and ge nomic information, obtaining outcome data for each of the plurality of subjects, the outcome data classifying each subject as belonging to one of the following at least two groups deter mined through a specified treatment regimen with a given drug during a period of time: hypo- prone, and non-hypo-prone. The method comprises the further steps of: training an ML model based on the population and outcome data, the trained model allowing a subject represented with a set of population data to be classified with a given certainty as belonging to one of the groups, and based on a set of population data for a subject, utilize the trained ML model to classify the subject. The individual population data may comprise values from one or more of the following further groups of data obtained during the period of time of treatment: meal and activity, and adher ence to prescribed treatment.

By the above method the HCP can categorize a given patient as belonging to a given risk group at the very beginning of treatment with the given drug, this allowing the HCP initially to determine a FBG treatment range that effectively will lower the patients HbA1c yet allows pa tients at hypo-risk to be treated accordingly.

Classification of a subject in treatment with a given drug in a given regimen may be optimized by: classifying the subject in a group utilizing the above-described method, the method com prises the steps of: based on the classification determine an FBG target range, during treat ment obtaining subject outcome data for the subject indicative of treatment outcome, and uti lizing subject outcome data to adjust the FBG target range.

BRIEF DESCRIPTION OF THE DRAWINGS

In the following embodiments of the invention will be described with reference to the drawings, wherein fig. 1 shows a graphical representation of FPG values as a function of time for a linear 1-FPG model, figs. 2A-2C illustrate different approaches for determination of a weight value W, fig. 3 shows distributions of baseline data for different diabetes phenotypes, fig. 4 shows a graphical representation of FPG values as a function of time for a non-linear 1-

FPG model, fig. 5 shows a graphical representation of FPG values as a function of time using a moving average model, fig. 6 shows a graphical representation of FPG values as a function of time using a moving average filter model, fig. 7 shows a graphical representation of FPG values as a function of time using a least square regression model, fig. 8A shows CGM data and personalized treatment for a non-hypo-prone user, fig. 8B shows CGM data and personalized treatment for a hypo-prone user, fig. 9 illustrates training an ML model based on population data, fig. 10 illustrates using the pretrained ML model to examine if a new patient and unseen data are predicted either as hypo-prone or non-hypo-prone, and fig. 11 illustrates follow-up support to the HCP for FPG target decision based on continuous monitoring of the treatment effect with the use of SMBG or CGM data considering different parameters.

DESCRIPTION OF EXEMPLARY EMBODIMENTS

The present invention relates to an algorithm adapted to estimate a fasting plasma glucose (FPG) value for a patient to be used in calculating a recommended dose of an insulin.

The algorithm of the present invention may be used as a stand-alone solution providing a user with information about his/her FPG, however, the algorithm may also be used as part of an overall diabetes dose guidance system that helps people with diabetes by generating recom mended insulin doses based on estimated FPG values.

In such a system a given algorithm is used to generate recommended insulin doses and treat ment advice for diabetes patients based on BG data, insulin dosing history and, in more ad vanced applications, other factors like meals, physical activity, stress, illness etc. may be taken into consideration.

Essentially such a system comprises a back-end engine (“the engine”) used in combination with an interacting systems in the form of a client and an operating system. The client from the engine’s perspective is the software component that requests dose guidance. The client gath ers the necessary data (e.g. FPG values) and requests dose guidance from the engine. The client then receives the response from the engine.

On a small local scale the engine may run directly as an app on a given user’s smartphone or on a blood glucose meter (BGM) and thus be a self-contained application comprising both the client and the engine. Alternatively, the system setup may be designed to be implemented as a back-end engine adapted to be used as part of a cloud-based large-scale diabetes manage ment system. Such a cloud-based system would allow the engine to always be up-to-date (in contrast to app-based systems running entirely on e.g. the patient’s smartphone), would allow advanced methods such as machine learning and artificial intelligence to be implemented, and would allow data to be used in combination with other services in a greater “digital health” set up. Such a cloud-based system ideally would handle a large amount of patient requests for dose recommendations. Although a “complete” engine may be designed to be responsible for all computing aspects, it may be desirable to divide the engine into a local and a cloud version to allow the patient-near day-to-day part of the dose guidance system to run independently of any reliance upon cloud computing. For example, when the user via the client app makes a request for dose guidance the request is transmitted to the engine which will return a dose recommendation. In case cloud access is not available the client app would run a dose-recommendation calculation using local data. Dependent upon the user’s app-settings the user may or may not be informed.

In a first aspect of the invention, a predicted FPG is calculated using a mathematical model using FPG history data and insulin injection history data. Based on the predicted FPG and a current measured FPG an updated estimated FPG can be calculated which can then be used to calculate a dose recommendation.

This will allow the system to predict the result of next week’s measurement from a given change in dose, whereby the system can be used to optimise suggestion of a new dose setting. As the FPG is then measured the following week and thus added to the historical data, the algorithm can improve the modelling accuracy and gradually adapt the parameters of the model to fit the specific individual’s response to the drug better and better.

As the model quickly becomes able to predict the following weeks FPG value within reasonable tolerances, the system will be able to detect bad SMPG measurements and outliers and re quest a measurement to be redone, accompanied with suggestions of how to avoid errors, such as washing hands just prior to measuring to avoid contamination of the sample and such. By becoming able to detect bad measurements “earlier”, in the sense that they may be de tected even when being just outside limited tolerances and not having to be obviously way off to an extend where the user would become suspicious, fewer measurements are needed, as one of the purposes of requiring three measurements where to reduce the impact of a bad measurement.

By using the predicted FPG value calculated from prior data combined with a single measure ment, a mean value can be established from only a single measurement, without the risk of a random variation caused by unusual activity or a contaminated sample creates a bad input for the model. This will reduce the discomfort experienced by the user to only once a week instead of three days a week. Furthermore, the solution will allow for a more efficient and shorter titra tion process, as a more accurate prediction of the individual user’s response to a new dose setting can be made. This reduces the risk of overshooting dose adjustments. This will be referred to as the FPG-1 method in the following. For the initial calculation of a predicted FPG value additional data may be supplied.

In a first embodiment a simple linear model is used to improve the accuracy of the estimated FPG.

G_t+1 — a(G_t ⁺ — G) — bb + G + w_t Vt = G_t + v_t

The linear model constitutes a prediction model for the current week t + 1 given (1) the FPG at week t (i.e. the previous week), (2) the insulin dose I_t given at week t, and (3) the FPG at zero insulin G. The term w_t is a stochastic term representing the physiological variations of the FPG. The superscript - denotes predicted FPG and the superscript + denotes the estimated FPG. The variable G_t is the actual FPG, which is unknown.

The measured FPG y_t is also subject to a measurement noise denoted v_t. This measurement noise depicts the uncertainties of the glucose sensor.

The second step consists of taking a weighted average between the predicted FPG G_t+1 at week t + 1 and the measured FPG y_t+1 taken at the same time, i.e.

Gt+i = WGf₊1 + (1 - W)y_t+1

The ideal weight 0 < W < 1 between the predicted and measured FPGs provides the best estimate of the actual FPG. The weight is designed to mitigate the effects of outliers, e.g. caused by the incorrect use of SMBG, and is chosen to reflect the domain knowledge about the noise distributions of v_t and w_t.

Table 1 below shows an example of numerical values for linear one-step prediction model.

Fig. 1 shows the corresponding values in graph form. The estimated FPG (squares) is a weighted average between the predicted FPG (circles) and the measured FPG (crosses). The dark line shows the trajectory of the predicted FPG.

Alternatively, each component may have their own weight, normalized to sum to 1, i.e.

G_t ⁺ _+i ⁼ W_±G_t+1 + VK₂y_t+1 with each weight set according to the weight functions described below, where G would be either G_t+1 ory_t+1.

Different approaches may be utilized to determine the weight W.

When computing the next estimated value, new measurements may be weighted differently, depending on their importance or how well they are “believed” compared to the model. The weighting function may be “global”, linear or non-linear, depending on glucose value. Fig. 2A illustrates an example of a general “global” non-linear cost function used to weigh measure- ments, whereas fig. 2B illustrates an example of a specific “global” non-linear cost function used to weigh measurements. For both cost functions low measurements weigh higher than high measurements.

Alternatively, a “local” weighting function may be used, linear or nonlinear, that weighs glucose measurements within a range of insulin doses, wherein low glucose measurements higher than high measurements. Fig. 2C illustrates an example of a “local” linear cost function used to weigh measurements. Only measurements following the same dose size are considered, i.e. a new weighting function is created for each dose size. In the figure G is the current glucose measurement following insulin dose I, G_max and G_mm are the maximum and minimum glucose measurements following the same insulin dose size as I (historically), and a=1 and b=2 (could be set otherwise).

Table 2 shows an example of weights using a “local” linear cost, with weights ranging from 0.5- 1. Only measurements following the same dose size are considered, i.e. a new weighting func tion is created for each dose size.

As a further alternative, when computing the next estimated fasting SMBG G_t ⁺ ₊₁, new meas urements may be weighted differently, depending on e.g. risk for hypoglycaemia or hypergly- caemia , and/or how well they are “believed” compared to the model:

where å denotes the inverse of the variance of the model or the measurements. These vari- ances can be known beforehand, e.g. through knowing the uncertainty of the SMBG device or knowing the variance of the model on a prior dataset, and are updated over time using new measurements and predictions from the model. It should also be noted that:

The weights sum up to 1 , only the ratio between the measurement and model variances matters, i.e. multiplying both variances by the same number would yield the same weights, and the way of computing weights can easily be generalized to N models while keeping the properties above, e.g. for model number i:

Table 3 shows an example of updates of fasting SMBG estimates

As a further example a multiple-model embodiment of the invention uses N models de noted f, for i = 1, 2, ..., N

These models are in the form

Gt+i, i = fi(Go, 1. t, lo, 1. t) For example, these models may encompass the different phenotypes of type 2 diabetes:

• SAID (Severe auto-immune diabetes)

• SIDD (Severe insulin-deficient diabetes)

• SIRD (severe insulin-resistant diabetes)

• MOD (Mild obesity-related diabetes)

• MARD (mild age-related diabetes)

Each phenotype is associated with a specific model f,. The likelihood of a given phenotype is determined a priori by the user’s baseline data (HbA1c, BMI, age, estimate of b-cell function, estimate of insulin resistance etc.), and is up dated a posteriori using data during titration.

In fig. 3 distributions of baseline data for different diabetes phenotypes are shown. Given these distributions, a new user may be assessed probabilities to belong to a phenotype given the user’s base-line data. For instance, a 50-year old user with a HbA1c around 50 mmol/mol and a BMI above 35 kg/m² is more likely to belong to the MOD phenotype.

The weighted average providing the next FPG update is

The weights

Wi are all non-negative and must satisfy

As for the first embodiment, the tuning of the weights W, depends on the accuracy of model I and the SMBG noise.

An example of advanced non-linear model would be:

See: Kanderian SS, Weinzimer S, Voskanyan G, Steil GM. Identification of Intraday Metabolic Profiles during Closed-Loop Glucose Control in Individuals with Type 1 Diabetes, Journal of Diabetes Science and Technology, 2009;3(5): 1047-1057. doi: 10.1177/193229680900300508. Table 4 below shows an example of numerical values for a non-linear model.

Fig. 4 shows the corresponding values in graph form. The estimated FPG (squares) is a weighted average between the predicted FPG (circles) and the measured FPG (crosses). The dark line shows the trajectory of the predicted FPG.

In the following an alternative approach to the above-disclosed concept of estimating a FPG value based on a weighted combination of an estimated FPG value based on historical data and a current measured FPG value will be described. More specifically, for calculating an es- timated fasting SMBG based on the historical fasting SMBG and the current measured fasting SMBG the following 3 methods are proposed:

1. Moving Average

When using a Moving Average to calculate a predicted fasting SMBG

each mean is calcu- lated over a sliding window of length k across neighbouring elements of

which denotes the historical fasting SMBG containing n entries. is then computed as (assuming that k is an even number):

In the example shown in table 5 and illustrated in fig. 5 the predicted Fasting SMBG Gjj (circles) is only regarded as the estimated fasting SM BG and used for titration if the value is lower than the actual measured fasting SMBG (crosses). The squares illustrate the selected measure ments for titration and the dark line shows the trajectory of the values used for titration.

Table 5 show examples of numerical values for Moving average one step estimation.

2. Moving Average filter When using a Moving Average filter to estimate based on the historical fasting SMBG G_j a window of length k is slid along the data, computing averages of the data contained in each window. is then computed as:

In the example shown in table 6 and illustrated in fig. 6 the predicted Fasting SMBG Gjj (circles) is only regarded as the estimated fasting SMBG and thus used for titration if the value is lower than the actual measured fasting SMBG (crosses). The squares illustrate the selected meas urements for titration and the dark line shows the trajectory of the values used for titration. Table 6 shows examples of numerical values for Moving Average filter one step estimation.

3. Least squares regression The least squares regression works by making the total of the square of the errors as small as possible and the estimated fasting SMBG G_n is then computed as:

G_n ®C_h T b where x_n is the units of insulin giving in the last dose and a is the slope:

The historical fasting SMBG measurements is denoted as y and the intercept b is calculated as:

In the example shown in table 7 and illustrated in fig. 7 the predicted Fasting SMBG Gjj (circles) is only regarded as the estimated fasting SMBG and used for titration if the value is lower than the actual measured fasting SMBG (crosses). The squares illustrate the selected measure ments for titration and the dark line shows the trajectory of the values used for titration.

Table 7 shows examples of numerical values for least squares regression one step estimation.

Independently of the actual model used, a simple linear model or a number of advanced non linear models, the estimated FBG value can be used as input to an algorithm for calculating a recommended dose of insulin, e.g. to provide a long-acting or ultra-long-acting insulin adjust- ment dose recommendation (ADR) for a subject to treat diabetes mellitus.

However, to personalize the recommendations a quality measure (QM) may be calculated for the patient for a period of time based on one or more of (i) the variance in FPG, (ii) the differ ence between predicted and measured FPG, and (iii) the difference between IH injection amounts and historic ADR.

Depending on (1) the user adherence to the treatment, (2) the variance in FPG, and (3) the difference between the estimated and measured FPG, a quality measure of the estimate com pared to the actual FPG can be determined, and the following guidelines are prescribed:

If the QM is at or below a given level the user will be advised to follow a normal or larger increment/ decrement in the drug dosage if needed. If the QM is above the given level or if another inconsistency is suspected, the user will be advised to take a smaller increment/ dec rement in the drug dosage, or to use the same drug dosage until the next week.

Alternatively, a Personalised Target Measure (PTM) may be calculated based on one or more of: the variance in FPG, the variance in minimum average BG for a moving average filter of a predetermined number of hours, and the difference between predicted and measured FPG. If the Personalised Target Measure (PTM) is lower than a given average for the subject, indicat- ing that the risk for insulin induced hypoglycaemia is lower, adjusting the target range level to a lower level than defined by the first data structure. On the other hand, a high PTM means a high risk of hypoglycaemia and the personalized target is adjusted to a higher level. Correspondingly, a low PTM means low variability and thus a lower risk of hypoglycaemia. In an exemplary embodiment the PTM is the ratio between the standard deviation of the 7 latest FPGs and the mean of the 7 latest FPGs, and is thus a measure of the FPG variability.

Table 8 below shows the PTM values using the variance in FPG.

The target range corresponds to the range where no dose change in drug is required. Table 9 below shows a calculation example.

Additionally or alternatively, a personalized CGM-enabled FPG target may be deter mined based on historical data and decided by the HCP. Using a CGM for a limited period of time, e.g. 2 weeks, gives access to frequent glucose measurements between two FPG meas- urements, e.g. every 5-15 minutes. If the user is at range and no hypoglycaemic event has been observed, the size of the range can be reduced to ensure a lower HbA1c, see fig. 8A in which the indicated area represents the lowered range. Conversely, if the user is prone to hypoglycaemia despite being in the range, the target will be increased to minimize the risk of subsequent hypoglycaemic events, see fig. 8B in which the indicated area represents the in- creased range.

Alternatively, a personalized FPG target may be based on machine learning (ML). More spe cifically, a personalized FPG target may be based on available patient information and histor ical data using a population-based ML classification model. This classification model can func- tion as a decision-support tool for the HCP to decide a personalized FPG target for each pa tient.

The input data may comprise one or more of the following: · 1-FPG

. HbA1c

• Demographics of the patient (e.g. age, sex, BMI, diabetes duration, etc.)

• Genomic information (DNA sample)

• A limited period, e.g. 2 weeks, in which the following data sources are acquired:

• CGM data (expert dependent consensus metrics)

• Information about meal intake (carbohydrates) • Activity level of the patient through activity tracker (e.g. smart watch)

• Adherence level to prescribed treatment (either self-reported or through con nected injection devices)

The chosen ML model is trained in a supervised learning scheme based on existing population data from patients initiated on a given drug, e.g. ICODEC from Novo Nordisk A/S, stratified into e.g. two or more classes:

1) Patients who are hypo-prone

2) Patients who are not hypo-prone

3) ... Any other desired outcome

Fig. 9 diagrammatically shows how an ML model is trained based on population data.

Thus, the input data consists of the proposed patient characteristics/ features, and the output classes of either hypo-prone or non-hypo-prone. These two classes are predefined based on time spent in hypoglycaemia (in %) and/or number of hypoglycaemic events. These two out comes can be extended by other outcomes of interest, e.g. different variants of hypo-prone and non-hypo-prone including those who suffer from cardiovascular diseases or chronic kidney diseases for who different considerations should be accounted for.

The data are fed to the model and iteratively trained and optimized with respect to model- specific (hyper)parameters. The optimal (hyper)parameter(s) are found by performing cross- validation, thus estimating the test error rate/predictive performance by holding out a subset of the training set from the fitting process until the desired predictive performance is obtained. Once the performance exceeds a predefined (satisfactory) threshold, the model will be (pre)trained for new and unseen data samples, i.e. new patients with same characteristics available.

The classification model may be in the form of a simple K-nearest Neighbours (KNN) where the prediction to one of the two classes is based on the highest frequency from the K-closest instances based on a distance metric (e.g. Euclidean), this as illustrated in fig. 10 showing how the pretrained ML model is used to examine if a new patient and unseen data are predicted either as hypo-prone or non-hypo-prone, supporting HCP to decide personalized FPG target. Alternatives could include probalistic classification models including e.g. logistic regression, gaussian processes or more complex models using deep learning including e.g. feed-forward, convolutional, recurrent neural networks or a combination of these as hybrid ensembled mod els.

When the ML model is fed with new and unseen input data with the presented features, it can predict/classify each patient into one of the two classes of hypo-prone or non-hypo-prone.

• If predicted hypo-prone, the user will be advised to take a smaller increment/decrement in the drug dosage, or to use the same drug dosage until the next week.

• If predicted as non-hypo-prone, the user will be advised to take a smaller incre ment/decrement in the drug dosage, or to use the same drug dosage until the next week.

• If the user is at range and no hypoglycaemic event has been observed, the size of the range is reduced to ensure a lower HbA1c.

The above-described concept works at t = 0 which means that it is based on baseline data without any information about the treatment effect of the given drug, e.g. ICODEC. This will support the HCP in deciding the FPG target at treatment onset. This can be extended by a follow-up/ error check mechanism/ support which continuously monitors the treatment effect development over time (time-variant) from one or several sources including but not limited to:

• SMBG measurements (BG meter)

. CGM

• ... Any other desired data source

From these two or more additional data sources it will be possible to calculate some indicators/ predictors about how the treatment is evolving. These consist of but are not limited to:

• BG trend (up- or downwards)

• Time in ranges (CGM based)

• Variance in BG (CV)

• Any other desired parameter Fig. 11 diagrammatically shows how Follow-up support to the HCP for FPG target decision can be based on continuous monitoring of the treatment effect with the use of SMBG or CGM data considering different parameters.

These parameters are weighted differently based on the severity/ importance with respect to current patient status/treatment. The weighted sum is then calculated as an Importance Score (S). Based on the S value and a predefined threshold (or ranges) by the HCP, different advices are presented/prompted to the HCP. An arbitrary example of this can be as follows where the importance weights and the three presented/scaled parameters sum up to 1, respectively:

51 = 0.5 * 0.2

52 = 0.2 * 0.38

53 = 0.3 * 0.78 Importance score S = 0.41

The threshold is set hypothetically at 0.5. In this case, S < 1 which provides advice A to the HCP to be considered.

This process is iterated and constantly applied as an additional and continuous support/ con firmation to the HCP of the suggested class (hypo-prone, non-hypo-prone or other) by the ML model. Adjustments can thus be made after t = 0 if needed which might deviate from the first choice of FPG target based on the temporal treatment effect of each individual patient. This concept can be considered as an add-on to the above-described 1-FPG titration.

In the above description of exemplary embodiments, the different structures and means provid ing the described functionality for the different components have been described to a degree to which the concept of the present invention will be apparent to the skilled reader. The detailed construction and specification for the different components are considered the object of a nor mal design procedure performed by the skilled person along the lines set out in the present specification.

_*****

Claims

1. A system for estimating a fasting blood glucose (FBG) value for a subject, wherein the system comprises one or more processors and a memory, the memory comprising: instructions that, when executed by the one or more processors, perform a method responsive to receiving a request for an updated predicted FBG value, the method comprising the steps of: obtaining a first data set, comprising a plurality of glucose measurements of the sub ject taken over a time course and thereby establish a blood glucose history (BGH), each re spective glucose measurement in the plurality of glucose measurements comprising:

(i) a plasma glucose (BG) level, and

(ii) a corresponding BG timestamp representing when in the time course the respec tive glucose measurement was made, based on BGH and using a mathematical model: calculating a predicted FBG for the time t+1 , where t+1 is the current time, obtaining at time t+1 a current measured FBG from the subject, and based on the predicted FBG and the current measured FBG: calculating an updated estimated FBG.

2. A system as in claim 1, the method comprising the further step of: obtaining a second data set, comprising a basal insulin injection history (IH) of the subject, wherein the injection history comprises a plurality of injections during all or a portion of the time course and, for each respective injection in the plurality of injections comprising:

(i) a corresponding injection amount, and

(ii) an injection timestamp representing when in the time course the respective injec tion occurred, wherein the predicted FBG is calculated based on BGH and IH.

3. A system as in claim 1 or 2, wherein the updated estimated FPG is calculated as a weighted average of the predicted FBG and the current measured FBG.

4. A system as in claim 3, wherein the updated estimated FBG is calculated as a weighted average of the predicted FBG and the current measured FBG: updated estimated FBG = Wi (predicted FBG) + W2(current measured FBG), wherein:

Wi + W2 = 1 , and

Wi and/or W2 are functions of respective FBG values, the W values being negatively correlated with rising FBG values.

5. A system as in claim 3, wherein the weight is based on the distributions of v_t and w_t, where v_t denotes measurement noise when determining the current measured FBG, and w_t is a stochastic term representing the physiological variations of the FBG.

6. A system as in claim 2, wherein the system is adapted to calculate a predicted FBG for the time t+1 using a plurality of mathematical models, each mathematical model represent ing a subgroup of a population of diabetic subjects, the method comprising the steps of: for the subject, selecting at least one mathematical model being representative for the subject, based on BGH and IH and for each selected mathematical model: calculating a pre dicted FBG for the time t+1, based on the at least one predicted FBG and the current measured FBG: calculating an updated estimated FBG.

7. A system as in claim 6, wherein the updated estimated FBG is calculated as a weighted average of the at least one predicted FBG and the current measured FBG.

8. A system as in any of claims 1-7 adapted to provide a long-acting or ultra-long-acting insulin adjustment dose recommendation (ADR) for a subject to treat diabetes mellitus, the memory further comprising: instructions that, when executed by the one or more processors, perform a method responsive to receiving a dose guidance request (DGR), the method comprising the further steps of: obtaining a first data structure, comprising:

(i) a glucose upper target range level (UTR) of the subject, and

(ii) a glucose lower target range level (LTR) of the subject, obtaining a second data structure, comprising:

(i) a current dose guidance baseline (DGB), wherein the current DGB corresponds to (a) a most recent ADR, or (b) a starting basal dose (SBD), providing the long-acting or ultra-long-acting insulin ADR, the recommendation being calculated based on data from the first data structure, the second data structure, and the up dated estimated FPG.

9. A system as in claim 8, the method comprising the further steps of: calculating a Personalised Target Measure (PTM) based on one or more of: the variance in FPG, the variance in minimum average BG for a moving average filter of a prede termined number of hours, the difference between predicted and measured FBG, if the Personalised T arget Measure (PTM) is lower than a given average for the subject, indicating that the risk for insulin induced hypoglycaemia is lower, adjusting the target range level to a lower level than defined by the first data structure.

10. A system as in claim 8 or 9, the method comprising the further steps of: obtaining from the subject for a period of time a continuous glucose monitoring (CGM) data set comprising:

(i) a plurality of BG levels, and

(ii) a corresponding PG timestamp representing when in the time course the respec tive CGM measurement was made, calculating for the period of time a hypo-risk value based on the number and/or se verity of hypoglycaemic events in the CGM data set, and if the hypo-risk value is below a given value, lower the LTR for the subject.

11. A system as in claim 10, the method comprising the further step of: if the hypo-risk value is below a given value, lower the UTR for the subject.

12. A system for estimating a fasting blood glucose (FBG) value for a subject, wherein the system comprises one or more processors and a memory, the memory comprising: instructions that, when executed by the one or more processors, perform a method responsive to receiving a request for an updated predicted FBG value, the method comprising the steps of: obtaining a first data set, comprising a plurality of glucose measurements of the sub ject taken over a time course and thereby establish a blood glucose history (BGH), each re spective glucose measurement in the plurality of glucose measurements comprising:

(i) a plasma glucose (BG) level, and

(ii) a corresponding BG timestamp representing when in the time course the respec tive glucose measurement was made, obtaining at time t+1 a current measured FBG from the subject, where t+1 is the cur rent time, and based on BGH and the current measured FBG from the subject and using a mathe matical model: calculating a predicted FBG for the time t+1 , determining an estimated FBG for the time t+1 as:

(i) the calculated predicted FBG if the calculated predicted FBG is lower than the cur rent measured FBG, or

(ii) the current measured FBG if the calculated predicted FBG is higher than the cur rent measured FBG.

13. A method for aiding in the classification of a subject as either hypo-prone or non-hypo- prone, comprising the steps of: obtaining individual population data for a plurality of subjects, the data comprising values from one or more of the following groups of data,

PG values, blood characterizing values, demographic data, and genomic information, obtaining outcome data for each of the plurality of subjects, the outcome data classi fying each subject as belonging to one of the following at least two groups determined through a specified treatment regimen with a given drug during a period of time, hypo-prone, non-hypo-prone, training an ML model based on the population and outcome data, the trained model allowing a subject represented with a set of population data to be classified with a given cer tainty as belonging to one of the groups, and based on a set of population data for a subject, utilize the trained ML model to classify the subject.

14. A method as in claim 13, wherein the individual population data comprise values from one or more of the following further groups of data obtained during the period of time of treat ment, meal and activity, and adherence to prescribed treatment.

15. A method of optimizing classification of a subject in treatment with a given drug in a given regimen, comprising: classifying the subject in a group utilizing the method of claims 13 or 14, based on the classification, determine an FBG target range , during treatment obtaining subject outcome data for the subject indicative of treatment outcome, and - utilizing subject outcome data to adjust the FBG target range.

_*****