Personalized artificial intelligence-based insulin dose determination system

Abstract

Provided is a personalized artificial intelligence-based insulin dose determination system for a terminal which can communicate with an insulin pump and a continuous blood glucose meter used by a user, the personalized artificial intelligence-based insulin dose determination system comprising: an interface layer which can communicate with the insulin pump and the continuous blood glucose meter, and signal-processes information from the insulin pump and the continuous blood glucose meter; a control layer which receives the signal-processed information from the interface layer and generates output information associated with an insulin injection amount; an outer safety layer which determines whether the output information generated from the control layer satisfies a preset threshold value condition, and if the output information satisfies the preset threshold value condition, transfers the output information to the interface layer; and a personalized safety layer which determines a safety control variable per individual by receiving, from the outside, prescription information including the total daily dose (TDD) of insulin information of a user, and transmits the determined control variable as an input variable of the control layer.

Description

A personalized AI-based insulin dosage decision system

[0001] The present invention relates to a personalized artificial intelligence-based insulin dosage determination system.

[0002] Diabetes mellitus is characterized by high blood sugar levels and relative insulin deficiency. There are two main types of diabetes: type 1 diabetes (insulin-dependent diabetes) and type 2 diabetes (non-insulin-dependent diabetes). In some cases, diabetes is also characterized by insulin resistance.

[0003] To treat diabetes, various noninvasive, transdermal (e.g., transcutaneous) and / or implantable electrochemical sensors are being developed to continuously detect and / or quantify blood glucose levels. Continuous glucose monitors are gaining popularity as a convenient way to monitor glucose levels.

[0004] A bolus calculator device is used along with these continuous glucose monitors, but in this case, variables such as basal insulin, correction factor (CF), and carbohydrate-to-insulin ratio (CIR) must be measured and entered, which is cumbersome, and its accuracy is also problematic.

[0005] Accordingly, the problem to be solved by the present invention is to provide a system that can more effectively determine variables used in a precision insulin control system and a bolus calculator utilizing the system.

[0006] In order to solve the above problem, the present invention provides a personalized artificial intelligence-based insulin dosage decision system by a computing terminal capable of communicating with an insulin pump and a continuous glucose monitor used by a user, the system including: an interface layer capable of communicating with the insulin pump and the continuous glucose monitor and signal processing information from the insulin pump and the continuous glucose monitor; a control layer that receives signal-processed information from the interface layer and generates output information related to the insulin injection amount; an external safety layer that determines whether the output information generated from the control layer satisfies a preset threshold condition and, if the output information satisfies the preset threshold condition, transmits the output information to the interface layer; and a personalized safety layer that receives prescription information including the user's TDD (Total Daily Dose of Insulin) information from the outside, determines an individual safety control variable, and transmits the determined control variable as an input variable of the control layer.

[0007] In one embodiment of the present invention, the control layer processes the continuous blood glucose amount (g), insulin-on-board (iob), and continuous blood glucose rate (dg / dt or v) measured by the continuous blood glucose meter from the signal-processed information. g ), and continuous glucose acceleration (d2g / dt2 or a g ) is used as input information to determine the amount of insulin to be injected to the user as output information.

[0008] In one embodiment of the present invention, iob at time t t is determined by the following formula.

[0009] (1)

[0010] (here i t-k is the insulin dose injected at time step k prior to time t, and F kis a gamma cumulative density function (CDF) using SF as a scaling factor, k is the number of time steps over which insulin dosage data is collected, and n is the maximum number of time steps over which insulin dosage data is collected during the time it is assumed that accumulated IOB remains in the user's body.

[0011] In one embodiment of the present invention, the control layer determines the amount of insulin to be injected to the user as output information using a deep reinforcement learning model.

[0012] In one embodiment of the present invention, the deep reinforcement learning model uses a SAC (Soft Actor Critic) algorithm, the control layer determines a blood sugar control policy of the SAC (Soft Actor Critic) algorithm at a preset time interval, and the control layer trains the SAC (Soft Actor Critic) algorithm model for the purpose of maximizing an objective function according to the following equation.

[0013] The above control layer uses the SAC (Soft Actor Critic) algorithm model for the purpose of maximizing the objective function J(π) according to the following equation.

[0014] (4)

[0015] (from here is the compensation sum, H is the entropy, and α is the temperature parameter)

[0016] In one embodiment of the present invention, the control layer determines the insulin injection amount as an output variable within one or more of the following three limit value ranges.

[0017] - Insulin dose during the day (S) D )

[0018] - Insulin dose during the night (S N )

[0019] - Maximum insulin-on-board value (iob) max )

[0020] In one embodiment of the present invention, the amount of insulin injected is determined by the following formula.

[0021]

[0022] (from here is the output information of the above control layer, t i is the time during the day, i on the right-hand side is the initially determined insulin injection dose using the Deep Reinforced Learning model, τ1 and τ2 are the start and end times of the night time, S D은 Coefficient for daytime, S N is the coefficient for night time, π is BR min / 2, and BR min is BR / 60 [U / min], where BR is the basal insulin rate.)

[0023] In one embodiment of the present invention, the amount of insulin infusion (S) during the night N ) is the amount of insulin injected during the day (S D ) is the value obtained by multiplying a coefficient between 0 and 1.

[0024] In one embodiment of the present invention, the coefficient is determined based on the TDD (Total Daily Dose of Insulin).

[0025] In one embodiment of the present invention, the maximum insulin-on-board value (iob max ) is determined as a constant based on the above TDD (Total Daily Dose of Insulin), and the maximum insulin-on-board value (iob) max ) is provided as a threshold of the above outer safety layer.

[0026] It provides an accurate insulin management (AIM) system applicable to high-volume endocrine clinical trials, and shows an excellent balance between the characteristics of being short, easy to remember, and easy to use to reduce medication errors and high accuracy.

[0027] FIG. 1 is a block diagram of a structure of a fully closed-loop insulin dosage determination system according to one embodiment of the present invention.

[0028] FIG. 2 is a schematic diagram of an operation of a layered structure system according to one embodiment of the present invention.

[0029] Figure 3 is a diagram comparing the generalized iob effect between two subjects with different TDD values.

[0030] Figure 4 is a calibration iteration procedure according to the present invention.

[0031] Figure 5 is a graph showing the correlation between TDD and the control variable SD.

[0032] Figure 6 is a graph comparing the daily intake of three types of CHO with an average content of 40 g, 80 g, and 60 g in rats, comparing (a) an open loop (including meal guidance) using a conventional basal-bolus calculator and (b) a completely closed loop (without meal guidance) according to the present invention.

[0033] Hereinafter, specific embodiments of the present invention will be described with reference to the drawings. However, these are merely examples and the present invention is not limited thereto.

[0034] In describing the present invention, detailed descriptions of known technologies related to the present invention will be omitted if they are deemed to unnecessarily obscure the gist of the invention. Furthermore, the terms described below are defined based on their functions within the present invention and may vary depending on the intent or custom of the user or operator. Therefore, their definitions should be based on the overall content of this specification.

[0035] Figure 1 is a block diagram of a fully closed-loop insulin dosage determination system architecture according to one embodiment of the present invention. Hereinafter, "layer" refers to functionally separated software and / or hardware that form a layered architecture.

[0036] The system according to the present invention can be implemented in a terminal (e.g., a mobile phone, a tablet PC, etc.) that can communicate with an insulin pump and a continuous blood glucose meter attached to the body of a user and that can perform computing on its own, and each layer below can be configured to be distinguished in software within the terminal.

[0037] That is, the system according to the present invention is a device that can be installed in a terminal or the like and processed by a processor of the terminal, and can communicate with an insulin pump and a continuous blood glucose meter used by a user, and the system (10) according to one embodiment of the present invention includes an interface layer (100) that can communicate with an insulin pump (20) and a continuous blood glucose meter (30), and can receive information from the single insulin pump (20) and the continuous blood glucose meter (30) and process the information as a signal. In one embodiment of the present invention, the interface layer (100) can transmit information on the amount of insulin injection determined by the system to the insulin pump (20), so that the insulin pump (20) can actuate by the amount of the determined injection.

[0038] The above system (10) includes a control layer (200) that receives signal-processed information from the interface layer (100) and generates output information related to the amount of insulin injected; an external safety layer (300) that determines whether the output information (i.e., the calculated amount of insulin injected) generated from the control layer (200) satisfies a preset threshold condition, and, if the output information satisfies the preset threshold condition, transmits the output information to the interface layer; and a personalized safety layer (400) that receives prescription information including the user's TDD (Total Daily Dose of Insulin) information from the outside, determines an individual safety control variable, and transmits the determined control variable as an input variable of the control layer.

[0039] In particular, a system according to one embodiment of the present invention includes an interface unit (100) for receiving signals from a continuous glucose monitor (CGM) 10 and an insulin pump (imnsulin pump, 20) attached to a user and generating a signal for driving an insulin pump, and a control unit (300) for determining an injection amount of the insulin pump. An outer safety unit (200) is placed between the interface unit (100) and the control unit (300) for determining an injection amount of the insulin pump, so that the insulin injection amount determined by reinforcement learning is checked again based on a threshold value, which is a safety threshold value, before it is sent to the final interface.

[0040] That is, the insulin injection amount determination system according to the present invention has a four-layer structure based on reinforcement learning, thereby maximizing flexibility and adaptability.

[0041] In one embodiment of the present invention, an algorithm according to the present invention is implemented in a smartphone, and therefore, the interface layer in FIG. 1 performs a communication function between an external continuous glucose monitor (CGM, 10), an insulin pump (imnsulin pump, 20), and a control unit, thereby receiving data or providing a signal to drive an external device.

[0042] In a system according to one embodiment of the present invention, several thresholds obtained from silico subjects were set, and for this purpose, an outer safety layer (200) was placed between the interface layer (100) and the control unit (300). This will be described in more detail below.

[0043] A control layer (300) based on deep reinforcement learning is provided on the outer safety layer (200).

[0044] In one embodiment of the present invention, the observed GM(g) value, the insulin dose (insulin-on-board, iob), which means the insulin still effective from the previous bolus, the CGM rate (dg / dt or v g ) and CGM acceleration (d2g / dt2 or a g ) and calculates the amount of insulin based on at least one of them. That is, by combining variables that can be obtained from blood sugar data by at least a plurality of continuous blood sugar meters, the rate of increase and decrease in blood sugar and the acceleration can be predicted, and if at least one or more of these variables is used, it all falls within the scope of the present invention.

[0045] At the top of the system architecture according to the present invention is a personalized safety layer (400) for individual safety. In one embodiment of the present invention, the personalized safety layer (400) has two sub-modules, called the OffSI module (410) and the OnSL module (420), which will be described in more detail below.

[0046] In the present invention, three new adjustable, time-dependent control variables were used to determine the safety of automatic insulin injection prescriptions between users, which are the coefficients of the day unit (S D ), coefficient of night unit (S N ) and maximum insulin dose (maximum IOB, IOB max), which limits the output value of the insulin dose based on the personal information associated with the TDD.

[0047] The OffSI module, which is a lower module of the personalized safety layer (400) according to one embodiment of the present invention, determines the correlation between the total daily dose (TDD) of insulin and the three variables above.

[0048] Another sub-module, the OnSL module, updates and adjusts the control variables to ensure the adaptability of the PersonAI algorithm according to the present invention. The OnSL module adaptively fine-tunes the control variables to ensure the performance and stability of the control algorithm for long-term automated insulin infusion (AID). For example, one of the functions of the OnSL module is to prioritize variables to prevent hypoglycemia and increase the time-in-range (TIR) ​​of blood glucose values ​​each week.

[0049] FIG. 2 is a schematic diagram of an operation of a layered structure system according to one embodiment of the present invention.

[0050] Referring to FIG. 2, the present invention configures two safety layers (a personalized safety layer (400) and an external safety laser (100)) on both sides of the control layer (200), so that the insulin injection amount based on actual daily information can be safely and flexibly determined.

[0051] The blood sugar level of a type 1 diabetic patient is controlled using an automated insulin injection system based on artificial intelligence customized for each individual according to the present invention. In this case, the system of the present invention is capable of predicting blood sugar level without the user's direct input of meal information in a fully closed loop manner.

[0052] The present invention illustrated in FIGS. 1 and 2 will be described in more detail through the following examples.

[0053]

[0054] Outpost Safety Layer-Safety Clearance Module

[0055] A system according to one embodiment of the present invention includes a separate outer safety layer (300) between the interface layer (100) and the control layer (200) as described above. The outer safety layer measures real-time metabolic status based on continuous glucose monitoring (CGM) and insulin infusion data, wherein g, v g , a g and iob values ​​are used as input values, where g is the blood glucose level measured by CGM, v g is the rate of blood sugar change, a g is the acceleration of blood glucose change, and iob is insulin-on-board, which indicates the amount of insulin already injected to lower the current blood glucose level. That is, as mentioned above g is the rate of blood sugar change, a g The acceleration of blood sugar change is a variable that can be obtained from blood sugar levels, and at least the above-mentioned g, v g , a g All of these fall within the scope of the present invention as long as at least one of the iob values ​​is used as a variable.

[0056] As shown in Figures 1 and 2, information processed by a signal processing module identical to the input values ​​of the personalized safety layer (400) and the control layer (200) is used in the outer safety layer. In the present invention, the clean g value is measured from the filtered CGM data, and v g , a g simply used the value derived from g.

[0057] In order to obtain another input value, iob, the present invention calculated the total amount of injected insulin according to the following equation (1).

[0058] (1)

[0059]

[0060] Here i t-k is the insulin dose injected at time step k prior to time t, and F k is a gamma cumulative density function (CDF) using SF as a scaling factor, and the present invention assumes that the cumulative IOB remaining in the user's body (e.g., 5 hours) operates using the CDF. k is the number of time steps for which insulin dosage data is collected, and n is the maximum number of time steps for which insulin dosage data is collected during the time period for which it is assumed that the cumulative IOB remains in the user's body. For example, if it is assumed that insulin action occurs in the body for 5 hours and IOB is collected at intervals of 5 minutes, n becomes 60.

[0061] The present invention uses a scale factor (SF) as described above to solve the problem of the prior art that personalized iob is not suitable for fully automatic insulin injection.

[0062] Figure 3 is a diagram comparing the generalized iob effect between two subjects with different TDD values.

[0063] Since the system according to the present invention is based on a model-free approach, an output safety layer is required to improve reliability from problems in severe external environments such as CGM errors.

[0064] The present invention suppresses the CSII pump insulin dose (i) through a so-called hard thresholding method in which coefficients whose absolute values ​​are lower than the threshold are set to 0, instead of proposing a basal rate (BR) as in the prior art.

[0065] In one embodiment of the present invention, three types of limits are provided based on time-dependent upper bounds of i and iob (imax and iobmax, respectively) and lower bounds of g (gmin).

[0066] First, the two conditions for insulin restriction are as follows: Equation (2).

[0067] (2)

[0068] Here, i(t) at the bottom right is the insulin injection amount, which is the final output value of the control layer.

[0069] Acceptable insulin injections are given according to the following single-acting instructions: max =BR[U / h] can be defined, and the equation is as follows (3).

[0070] (3)

[0071] Thirdly, the possibility of problems such as hypoglycemia must also be anticipated, where the CSII pump insulin dose (i) is i(t)=0.

[0072]

[0073] Deep Reinforcement Learning (DRL)-based Control Layer

[0074] As described above, the control layer of the system according to the present invention is based on the deep reinforcement learning (DRL) framework, and therefore, in one embodiment of the present invention, the fully closed rope-automated insulin injection problem is diagrammed as a Markov Decision Process consisting of six states (S, O, A, P, R, γ). Here, S represents the user's hypoglycemic state (probabilistic observation value = 0).

[0075] In the present invention, four values ​​were used as input variables of the control layer, which are blood sugar g(t) (mg / dL), insulin dose (insulin-on-board, iob), CGM speed (dg / dt or v g) and CGM acceleration (d2g / dt2 or a g )am.

[0076] The action interval A is assumed to be continuous, and the amount of insulin delivered by the CSII pump is is defined as

[0077] For learning, the present invention sets the motion interval A to 0< A <basal max was set to . Here, basal max is the basal rate (BR) size at which the CSII pump is allowed to inject insulin (i.e., 30U).

[0078] The state transition function P defined in the physiological model depends on the metabolic dynamics of each individual (i.e., UVA / Padova model and Hovorka model).

[0079] The reward function R is related to the degree of blood sugar control. This is R:(s t ,a t ) and is a factor reflecting the optimization objective using a discount factor ν ⊆ [0.1] that determines the trade-off between immediate and delayed rewards.

[0080] The present invention utilizes two types of reward functions: a long-term reward function and a short-term reward function. The long-term reward function simulates basal insulin secretion from β-cells, while the short-term reward function simulates basal excess insulin secretion from β-cells. Therefore, the discount factor in the present invention was set to γ ​​= 0.99.

[0081] In one embodiment of the present invention, in relation to the insulin prescription strategy, the Soft Actor Critic (SAC) algorithm was adopted to learn the blood sugar control strategy (see T. Haarnoja, A. Zhou, K. Hartikainen, G. Tucker, S. Ha, J. Tan, et al., “Soft actor-critic algorithms and applications,” arXiv preprint arXiv vol.1812, no. 05905, 2018.; T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine, “Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor,” International conference on machine learning, vol. 35, no. 80, pp. 1861-1870. PMLR, 2018.)

[0082] In one embodiment of the present invention, the policy action, which is the output value of the algorithm, corresponds to the desired insulin injection amount at each preset time step (5 minutes in one embodiment of the present invention). The policy network in the present invention produces μ, log(σ) as an output value, which is a parameterization of the normal distribution N(μ,σ), and the action is a sigmoid normal distribution or It is distributed according to .

[0083] According to the SAC algorithm according to one embodiment of the present invention, maximum entropy learning is corrected in a continuous operating region where a probabilistic policy π is expressed by a parameter Φ.

[0084] According to one embodiment of the present invention, a deep learning policy π Φ The maximum entropy objective function can be formulated as follows, where this equation is learned to maximize the objective function J(π).

[0085] (4)

[0086] Here is the sum of rewards (or cumulative rewards), H is the entropy, and the relative importance of rewards is determined by the temperature parameter α. In one embodiment of the present invention, a conventional automatic temperature adjustment method was used to derive the optimal temperature parameter. Each variable and parameter of the objective function was considered with reference to T. Haarnoja, A. Zhou, K. Hartikainen, G. Tucker, S. Ha, J. Tan, et al., “Soft actor-critic algorithms and applications,” arXiv preprint arXiv, vol. 1812, no. 05905, 2018.

[0087] To ensure user-to-user prescription safety, which corresponds to universal safety rather than individual safety, the present invention sets a time-dependent insulin delivery limit value for the output value of the DRL-based control layer. That is, at each step k, the limit value of the control layer is applied as follows.

[0088] (5)

[0089] Here, the daily time between τ1=00:00 am and τ2=07:00 a.m., π is defined as BRmin / 2, where BRmin corresponds to BR / 80 [U / min)], is the output value of deep reinforcement learning according to the present invention.

[0090] In one embodiment of the present invention, the adjustment coefficient S during the day D > S, the adjustment coefficient during the night N and S D is determined by the correction method described below, and S N is defined by the following equation.

[0091] (6)

[0092] Here, b1 and b2 are the limits of 0.3 and 0.2, respectively, and b1 > b2.

[0093] Finally, the present invention is iob max is defined by the following equation, which is included in the outer safety layer equation (2).

[0094] (7)

[0095]

[0096] Personal safety layer

[0097] The present invention provides a novel PDSA algorithm for a personal safety layer, which is described in more detail below.

[0098] 1) Offline-Safe Start OffSI module

[0099] The OffSI module according to the present invention provides individual initial values ​​for the control variables of the artificial intelligence-based insulin infusion system according to the present invention for all treatment patients.

[0100] As described above, the most practical way to personalize a control algorithm model without retraining is to segment the algorithm processing aggressiveness based on the ICR, BW, TDD, age, and gender of each type 1 diabetic patient. This personal information is linked to the inter-individual variability in insulin sensitivity (IS). Therefore, the module according to the present invention is an OffSI module based on a regression model, which defines the relationship between TDD and the optimized control variable.

[0101] In the present invention, the optimized control variables (i.e., S) of each T1D-VP (type 1 virtual diabetic patient) are determined through a compensation iteration procedure. D ) was found. On the other hand, other parameters, i.e., iob max Wow S N is determined from the hard threshold based on TDD, as shown in Equations (6) and (7) above, respectively.

[0102] In silico testing experiments were conducted according to a protocol specifying meal times and amounts.

[0103] High S D The values ​​represent aggressive control values, which can be seen at higher infused insulin values ​​and may be suitable for some type 1 diabetics with higher TDD.

[0104] Meanwhile, low S D The values ​​represent stable control values ​​around the baseline value, which may be sufficient to maintain glucose in the normal range and suppress blood glucose elevation in some patients.

[0105] Figure 4 is a calibration iteration procedure according to the present invention.

[0106] The calibration iteration procedure according to the present invention is illustrated in Algorithm 1 of Fig. 4. Here, And, . Therefore, the input value of the correction iteration procedure according to the present invention is a SD set with an increment of 0.1 within the range of 0-5 for TDD≥50, and an increment of 0.05 for TDD<50. In each iteration of Algorithm 1, the following performance index was calculated in one embodiment of the present invention.

[0107]

[0108] Here, TIR and time below the range were calculated from the time of glucose concentration in the normal range (between 70 and 180 mg / dL per day) and the time of glucose in the hypoglycemic state (less than 70 mg / dL), respectively. β>1 is an adjustable parameter.

[0109] In each experimenter repetition, a quick rest was performed if some condition was met. For example, the "good" condition was 5 consecutive s kIt is defined as a repeat TIR ≥ 90% for day, and a "bad" condition is defined as 5 consecutive s k It is defined as a repeat TIR ≥ 5% for day.

[0110] Finally, the regression formula used in the present invention is illustrated in Fig. 5, which is a graph showing the correlation between TDD and the control variable SD.

[0111] This can be expressed by the following equation (9).

[0112] S d Init = ae b.TDD

[0113] Here, the constants a and b were 0.013 and 0.086, respectively.

[0114]

[0115] OnSL module

[0116] In one embodiment of the present invention, the purpose of the OnSL module is to update control variables and improve performance indices to prevent hypoglycemic events.

[0117] Since intra-daily insulin sensitivity (intra-daily IS) varies in each type 1 diabetic patient, the purpose of the OnSL module in one embodiment of the present invention is divided into daily and weekly units.

[0118] For example, the diurnal pattern of insulin sensitivity (IS) has been shown to be lower, on average, in the morning (morning) than at noon and at night (lunch and dinner), and other uncontrolled parameters (e.g., continuous glucose monitoring) can be subject to error. Specifically, one embodiment of the present invention focused on preventing daily hypoglycemia and increasing weekly TIR.

[0119] At each date k, the update rule is that blood sugar is less than 60 mg / dL (TBR60), and TBR60 x = {TBR60 D , TBR60 N , TBR60W} (where D is day, N is night, and W corresponds to TBR60 for the entire day) is based on the performance percentage.

[0120] In an update algorithm according to one embodiment of the present invention, TBR 60 is selected in consideration of the error from the CGM value, and the update rule according to the present invention is as follows.

[0121] (10)

[0122] Here, δ = {S D , S N} are control variables for day and night updated by the algorithm according to the present invention. In one embodiment of the present invention, iob is used to consider the stability of control. max It does not update on a daily basis.

[0123] Constraints f1 and f2 represent gains of 0.2 and 0.1, respectively.

[0124] A similar update rule was applied to the weekly update. In the d-day embodiment of the present invention, a percentage constant was used to update the control variable. Therefore, on each date j, the average TIR x = {TIR D , TIR N , TIR W} was calculated,

[0125] At date k, the above update rule follows the following formula.

[0126] (11)

[0127] Here, δ can be varied based on the specific performance matrix obtained daily, and the constants θ1 and θ2 represent gains of 0.2 and 0.1, respectively. In one embodiment of the present invention, TIR W When > 0.9, θ 1 > θ2, and the control variable was maintained. However, at the subsequent date (k+1), TBR60W > 0 and TIR W If any of the conditions <80 is satisfied, the update process is resumed. From the above equation (10), it can be seen that if any of the conditions is satisfied, the control parameters are updated. Therefore, in one iteration, one or all of the δ variables may or may not be updated. Furthermore, if the conditions are not satisfied in a particular iteration, the current variable values ​​are maintained (i.e., S DK = S DK+1 )

[0128]

[0129] Experimental Example 1

[0130] Figure 6 is a graph comparing the daily intake of three types of CHO with an average content of 40 g, 80 g, and 60 g in rats, comparing (a) an open loop (including meal guidance) using a conventional basal-bolus calculator and (b) a completely closed loop (without meal guidance) according to the present invention.

[0131] Referring to FIG. 6, the control results of the basal-bolus method according to the prior art, in which a high insulin bolus is required at each meal time, and the micro-bolus-based method of the fully automated system according to the present invention can be seen. It can be seen that when insulin is predicted in advance through the system according to the present invention, insulin control with performance at the level of or better than the conventional control method that requires inputting the conventional meal amount is possible.

[0132] In one embodiment of the present invention, a comparative analysis was performed using the first two weeks of data obtained from an input diet-open loop treatment and the last four weeks of data obtained from a non-input diet-completely closed loop of the present invention with a personal AI.

[0133] In this experiment, a total of 110 VPs (virtual patients) were used for Scenario 1 and Scenario 2, whereas only 10 VPs from UVA / Padova were used for Scenario 3, and therefore each point on the graph for Scenario 3 was calculated based on 24-hour CGM readings.

[0134] Here, in Scenario 1, the Hovorka model of the UVA / Padova model and the built-in variability of daily IS embedded in DMMS.R were used. Each T1D (type 1 diabetes) VP consumed a fixed amount of food at specific intake times. In Scenario 2, the algorithm was tested with a relatively high intake of 90 g of carbohydrates for breakfast, lunch, and dinner.

[0135] This scenario closely mimicked the test protocol of Dassau et al. (2015). As previously mentioned in Dassau et al. (2015), this scenario can present a demanding stress test for the AP controller, especially for FCL-AID. Finally, Scenario 3 was used to determine the importance of the proposed OnSL module, a more challenging and realistic scenario (Gondhalekar et al., 2018). Here, we tested the designed scenario in the UVA / Padova T1D VP, but modified the original version and introduced additional daily variability in insulin absorption and IS.

[0136] The nominal IS pattern proposed by Visentin et al. (2015) was implemented and randomly selected for each in silico subject. Furthermore, the daily variability in sensitivity was characterized by a random ±50% variation in the nominal IS from the beginning to the end of the trial.

[0137] As a result of the comparative experiment, the algorithm of the present invention was successfully customized and adapted to each patient's condition during the first 45 days of the open loop simulation scenario according to the prior art and the last 31 days of the proposed algorithm.

[0138] The present invention is a personalized artificial intelligence-based insulin dosage determination system, and its industrial applicability is recognized.

Claims

1. A personalized artificial intelligence-based insulin dosage determination system using a computing terminal that can communicate with the insulin pump and continuous blood glucose meter used by the user. An interface layer capable of communicating with the insulin pump and the continuous glucose monitor, and signal processing information from the Singgi insulin pump and the continuous glucose monitor; A control layer that receives signal-processed information from the interface layer and generates output information related to the amount of insulin injection; An outer safety layer that determines whether the output information generated from the above control layer satisfies a preset threshold condition, and if the output information satisfies the preset threshold condition, transmits the output information to the interface layer; and A personalized artificial intelligence-based insulin dosage determination system including a personalized safety layer that receives prescription information including the user's TDD (Total Daily Dose of Insulin) information from an external source, determines an individual safety control variable, and transmits the determined control variable as an input variable of the control layer.

2. In paragraph 1, The above control layer processes the continuous blood glucose amount (g), insulin-on-board (iob), and continuous blood glucose rate (dg / dt or v) measured by the continuous blood glucose meter from the signal-processed information. g ), and continuous glucose acceleration (d2g / dt2 or a g ) is a personalized artificial intelligence-based insulin dosage determination system characterized in that it determines the amount of insulin to be injected to the user as output information using at least one of the input information.

3. In paragraph 2, The insulin-on-board (iob) at time t is iob t A personalized artificial intelligence-based insulin dosage determination system characterized in that it is determined by the following formula. (1) (here i t-k is the insulin dose injected at time step k prior to time t, and F k is a gamma cumulative density function (CDF) using SF as a scaling factor, k is the number of time steps over which insulin dosage data is collected, and n is the maximum number of time steps over which insulin dosage data is collected during the time it is assumed that accumulated IOB remains in the user's body.

4. In paragraph 2, The above control layer is a personalized artificial intelligence-based insulin dosage determination system characterized in that it determines the insulin dosage to be injected to the user as output information using a deep reinforcement learning model.

5. In paragraph 4, A personalized artificial intelligence-based insulin dosage decision system, characterized in that the above-mentioned deep reinforcement learning model uses a SAC (Soft Actor Critic) algorithm, and the control layer determines the blood sugar control policy of the SAC (Soft Actor Critic) algorithm at preset time intervals.

6. In paragraph 5, The above control layer is a personalized artificial intelligence-based insulin dosage decision system characterized in that it uses the SAC (Soft Actor Critic) algorithm model for the purpose of maximizing the objective function J(π) according to the following equation. (4) (from here is the compensation sum, H is the entropy, and α is the temperature parameter) 7. In paragraph 4, A personalized artificial intelligence-based insulin dosage determination system, characterized in that the above control layer determines the insulin injection amount as an output variable within one or more of the following three limit value ranges. - Insulin dose during the day (S D ) - Insulin dose during the night (S N ) - Maximum insulin on-board value (iob) max ) 8. In paragraph 7, A personalized artificial intelligence-based insulin dosage determination system, characterized in that the insulin injection amount is determined by the following formula. (Here, the left side is the final output information of the above control layer, the amount of insulin injected, t i is the time during the day, i on the right side is the initially determined insulin injection dose using the Deep Reinforced Learning model, τ1 and τ2 are the start and end times of the night time, S D은 Coefficient for daytime, S N is the coefficient for night time, π is BR min / 2, and BR min is BR / 60 [U / min], where BR is the basal insulin rate.) 9. In paragraph 8, Insulin dose during the night (S N ) is the amount of insulin injected during the day (S D ) is a personalized artificial intelligence-based insulin dosage decision system characterized by a value obtained by multiplying a coefficient between 0 and 1.

10. In paragraph 9, A personalized artificial intelligence-based insulin dosage determination system, characterized in that the above coefficient is determined based on the TDD (Total Daily Dose of Insulin).

11. In paragraph 3, The maximum insulin-on-board value (iob) max ) is determined as a constant based on the above TDD (Total Daily Dose of Insulin), and the maximum insulin-on-board value (iob) max ) is a personalized artificial intelligence-based insulin dosage decision system characterized in that it is provided as a threshold of the above outer safety layer.

Images