Sage Journals: Discover world-class research

Abstract

Background:

Hybrid closed-loop control of glucose levels in people with type 1 diabetes mellitus (T1D) is limited by the requirements on users to manually announce physical activity (PA) and meals to the artificial pancreas system. Multivariable automated insulin delivery (mvAID) systems that can handle unannounced PAs and meals without any manual announcements by the user can improve glycemic control by modulating insulin dosing in response to the occurrence and intensity of spontaneous physical activities.

Methods:

An mvAID system is developed to supplement the glucose measurements with additional physiological signals from a wristband device, with the signals analyzed using artificial intelligence algorithms to automatically detect the occurrence of PA and estimate its intensity. This additional information gained from the physiological signals enables more proactive insulin dosing adjustments in response to both planned exercise and spontaneous unanticipated physical activities.

Results:

In silico studies of the mvAID illustrate the safety and efficacy of the system. The mvAID is translated to pilot clinical studies to assess its performance, and the clinical experiments demonstrate an increased time in range and reduced risk of hypoglycemia following unannounced PA and meals.

Conclusions:

The mvAID systems can increase the safety and efficacy of insulin delivery in the presence of unannounced physical activities and meals, leading to improved lives and less burden on people with T1D.

Keywords

multivariable artificial pancreas automated insulin delivery physical activity exercise additional physiological signals closed-loop control

Introduction

Hybrid closed-loop insulin delivery improves glycemic control in people with type 1 diabetes mellitus (T1D) compared with sensor augmented pump therapy, leading to clinically significant improvements in time in range, reduction in hemoglobin A1C, and lower risk of hypoglycemia.^1
-5 Hybrid closed-loop control continuously adjusts basal insulin delivery in response to measurements of the blood glucose concentration (BGC) levels, with premeal boluses and exercise-related insulin dosing adjustments addressed through manual interventions by the user.^6
-10 The burden on users to manually enter premeal boluses and modulate insulin dosing to accommodate physical activity (PA) limits the safety and efficacy of hybrid closed-loop insulin delivery.^11
-13 Multivariable automated insulin delivery (mvAID) systems, also known as multivariable artificial pancreas, achieve better glycemic control than hybrid closed-loop control systems by supplementing the continuous glucose monitoring (CGM) measurements with real-time inputs of physiological signals received from wearable activity trackers to automatically detect the onset of PA and estimate its type and intensity, and mitigate its effects on glucose variations.¹⁴

Recent advancements in automated insulin delivery (AID) systems have incorporated automation of mealtime boluses, with algorithms developed to analyze historical and current CGM measurements and consider the insulin-on-board to automatically detect the occurrence of meals and estimate the carbohydrate (CHO) consumption amounts.^15
-18 The amount of CHOs consumed can also be estimated in real-time as the meal progresses and BGC rises in response to CHO intake.^17,19
-23 Glucose-only AID systems can handle automation of mealtime boluses by analyzing the effects of meals on the CGM measurements, but they are ineffective in modulating insulin dosing in response to PA since the AID system only detects exercise with a significant delay when the BGCs decrease.²⁴ This limitation in handling PA can be addressed by incorporating additional physiological signals, like heart rate and energy expenditure, which can indicate the occurrence of PA and enable estimation of its intensity.

Physical activity presents a significant challenge to glycemic control for people with T1D.²⁵ Moderate-intensity PA can cause a precipitous decline in BGCs, which increases the risk of hypoglycemia.²⁶ Moreover, PA induces heightened insulin sensitivity and increased glucose disposal, which may persist for several hours following PA. Planned exercise may be managed by hybrid closed-loop control through early interventions initiated by the user in anticipation of the exercise, and timely reductions in insulin delivery well in advance of the exercise. In contrast to planned exercise, unplanned or spontaneous PA presents a major challenge to glycemic control because unanticipated PA cannot be preemptively accommodated by users of AID systems. A new generation of mvAID systems is proposed that incorporates additional physiological signals from a wristband device with artificial intelligence in real-time to automatically detect PA, identify its type and estimate its intensity, and modulate insulin dosing in response to concomitant evolving variations in PA-related glycemic dynamics.

This article presents a novel mvAID system that does not require any manual entries from users to handle the glycemic effects of planned and spontaneous physical activities and meals. The structure, modules, and algorithms encompassing the mvAID system are described in this article, and results evaluating the mvAID system in both simulation studies and pilot clinical studies are presented. The mvAID readily accommodates daily patterns of everyday life to improve glycemic control and is robust to unforeseen glycemic disturbances. Artificial intelligence powered predictive modules are developed and implemented on a mobile application, enabling the mvAID to proactively counteract the glycemic disturbances, including planned and spontaneous physical activities and meals. The mvAID system has multiple components, including modules to process the real-time streaming of the additional physiological signals to impute missing data from sensors, reconcile outliers, and eliminate artifacts in the physiological variables; estimate heart rate from wristband signals; estimate energy expenditure, a valuable indicator for the intensity of PA; estimate the plasma insulin concentration (PIC), a robust measure of the insulin circulating in the bloodstream; detect unannounced meals and estimate CHO consumption; optimize the insulin dosing with an adaptive model predictive control (MPC) formulation that predicts the future BGC trajectory profiles in response to the assessed PA levels, detected meals, and PIC; estimate the likely future risk of potential hypoglycemic and hyperglycemic episodes.^27
-31 Deep neural networks are developed and trained on historical data to capture the daily activity behaviors and patterns of people with T1D. These patterns are used in predictive modeling and anticipating the future glycemic disturbances used by the MPC algorithm for more accurate insulin dosing. The safety and efficacy of the mvAID are demonstrated through simulation studies and the mvAID is translated to pilot clinical studies that validate the improvement in glycemic control over glucose-only AID systems.

Methods

Structure of mvAID System

Figure 1 depicts the framework of the mvAID system. The blue-colored blocks indicate modules that estimate glycemic disturbances in real-time,^27,28,30,31 green-colored blocks are signal processing modules or algorithms used for real-time signal reconciliation,^32,33 red-colored blocks represent the modules and algorithms in the feedback loop, namely, the system identification (modeling), MPC formulation, and safety constraint modules.

Figure 1.

The schematic architecture of mvAID system.

Components of mvAID System

The closed-loop AID system consists of two main parts:

Feedback modules: Feedback modules are the main part of the control system, where the optimal insulin dose to administer is computed by the adaptive MPC algorithm based on CGM measurements and the past infused insulin data. The feedback module is made up of multiple sub-components: CGM signal reconciliation, PIC estimation, prediction of potential future hypoglycemia and hyperglycemia events, adaptation model parameter estimation using nonlinear recursive least squares (NRLS), safety constraints module, and adaptive MPC.^34,35 Details on their implementation will be presented later.

Feedforward modules: The purpose of the feedforward modules is to further modify the controller output based on anticipated future occurrences of physical activities and meals. The feedforward modules estimate the disturbances affecting BGC dynamics using wristband signals. Examples of feedforward modules for PA include energy expenditure estimation through calculated metabolic equivalents of task (MET) values reflecting metabolic rate during activity and heart rate estimation.

The meal detection and the CHO intake amount estimation algorithms are predictor modules developed based on the CGM sensor signal and act similarly to feedforward modules. Hence, they are considered in a separate category, separate from feedback/feedforward modules.

Initialization of the mvAID System

The mvAID system consists of numerous modules that adaptively and recursively predict the variables of interest. Hence, initialization of these modules, variables, and predictive models is performed before computing the insulin dose suggestion by the MPC algorithm.

Initialization of the mvAID system is a crucial step that necessitates providing the demographic and insulin infusion routine information. The initialization module consists of a data-driven model with demographic information entered by the user as model input. The model outputs are time-varying parameters of the glucose-insulin dynamics whose values are calculated adaptively during the execution of mvAID. Initialization of physiological parameters in Hovorka’s model is imperative for estimating the PIC and the amount of CHO intake. To initialize the mvAID system, a data-driven approach is used where kernel partial least squares (KPLS) regression is implemented based on the pipeline model presented in Figure 2. Radial basis function kernel with 95% variance of input and output parameters (as listed in Table 1) are explained by latent vectors trained on data collected from 12 T1D subjects in open-loop clinical experiments.

Figure 2.

A pipeline for the initialization of physiological parameters in mvAID.

Table 1.

The List of Inputs and Predictor Outputs of the Partial Least Squares Model.

Variable	Type	Explanation
Gender	Input	A binary variable (“0” for females and “1” for male subjects)
BMI	Input	Body mass index $(\frac{k g}{c m^{2}})$
Age	Input	Expressed in year
Waist size	Input	The waist size expressed in cm
BW	Input	The body weight in kg
Height	Input	The height of subject in cm
A1C	Input	A hemoglobin A1C measurement $(\frac{m g}{d L})$
$S_{I T}^{f}$	Output	Insulin sensitivities of transport $(\frac{m i n^{- 1}}{m U . L^{- 1}})$
$S_{I D}^{f}$	Output	Insulin sensitivities of disposal $(\frac{m i n^{- 1}}{m U . L^{- 1}})$
$S_{I E}^{f}$	Output	Insulin sensitivities of endogenous glucose production $(\frac{m i n^{- 1}}{m U . L^{- 1}})$
$K_{e}$	Output	The elimination rate of insulin from plasma $(m i n^{- 1})$
$T_{m a x, I}$	Output	Time-to-maximum of absorption of injected insulin $(m i n)$
$T_{m a x, I}$	Output	Time-of-maximum glucose appearance rate in the bloodstream (regular meal) $(m i n)$
$T_{m a x, I}$	Output	The time constant of the dynamic effect of EE on the glucose concentration $(m i n)$
$β_{E E}$	Output	A gain value converting the effect of quantitative physical activity in masses of glucose in the no accessible compartment
$E G P_{0}$	Output	EGP extrapolated to zero insulin concentration $(\frac{m m o l}{k g^{- 1} . m i n^{- 1}})$

Source: The parameters of the model were taken from the work of Hovorka et al.³⁶

Abbreviations: BMI, body mass index; BW, body weight.

Estimating Disturbances Affecting BGC Regulation

Mitigating the effects of disturbances before they affect the outputs is the main responsibility of the feedforward blocks of the control loop.³⁷ In the mvAID system, feedforward modules translate raw physiological variable signals into meaningful features variables, usually to characterize disturbances and adjust the manipulated variables directly or indirectly.

Reconciliation and Processing of Wristband Physiological Signals

The interpretation of raw physiological signals and converting the measured variables data to informative estimates of disturbance variables requires denoising the raw physiological signals and performing feature engineering. Physiological signals are processed differently depending on disturbance variables. Heart rate plays a crucial role in predicting disturbances, such as PA and is a useful variable to estimate the effects of PA on the glycemic dynamics. Estimation of heart rate from blood volume pulse measurements requires denoising the raw blood volume pulse measurements, motion artifact removal, and designing an algorithm for tracking the frequency of heart rate peaks in real-time. The detailed process of estimating heart rate in real-time is provided by Askari et al.^32,33

Reconciliation of Glucose Measurements

A closed-loop control feedback system relies heavily on BGC measurements captured by CGM sensors. Interruption in wireless connectivity between the CGM sensor and the mobile edge computing device where the AID system resides (cellphone or insulin infusion pumps) is considered a critical situation, where no monitoring of BGC is performed. Hence, temporarily predicting BGC values is vital for maintaining the functionality of the hypoglycemia/hyperglycemia detector, CHO intake and PIC estimators, and making control decisions. In this regard, up to six missing CGM measurements are predicted using a vector autoregressive (VAR) model with a recursive adaptation of model parameters. Predicted CGM values are replaced by cubic spline interpolated values after CGM connection is re-established.

Detecting PA and Quantifying Its Intensity With Energy Expenditure

The detection of PA occurs follows the preprocessing of the blood volume pulse signal for noise cancelation and artifact removal, and estimating heart rate values. A major obstacle to integrating multiple modules is implementing these modules on platforms with limited computational power. Prediction of energy expenditure through MET values is imperative to avoid the risk of hypoglycemia caused by PA and to have a more accurate model for use in the MPC algorithm. Combining the approaches presented in the work of Askari et al³³ for preprocessing the raw signals and using the simplified model developed in previous works,^38,39 we develop and integrate an accurate estimation of MET values for automatically calculating an optimal amount of insulin in real-time. The developed MET estimation quantified the intensity of physical activities by measuring the last ten seconds of wristband signals, including three-axis accelerometer signals, skin temperature, blood volume pulse signal, and galvanic skin response. More details on the MET estimation module are provided in the work of Sevil et al.³⁹

The quantification of PA with MET values results in a continuous variable that can be applied directly to activate or inactivate the insulin pump or be utilized as an additional exogenous input to the linear state-space model employed by the MPC formulation. When the estimated MET values exceed a specific threshold value, a simple approach to ensure the safe regulation of BGC levels is to suspend the insulin delivery by the infusion pump. The estimation of MET values has been used directly in the formulation of the mvAID system. High values of estimated MET result in lowering BGC when performing low-intensity or moderate-intensity PA and this behavior has been considered in the extended exercise-glucose-insulin dynamic model.^40,41

Real-Time Adaptive Estimation of PIC

Plasma insulin concentration is a variable whose value is critical for designing adaptive identification algorithms, predictive models, and safety constraint modules. Currently, there is no commercially available sensor for measuring PIC in real-time. Therefore, estimation of the PIC should be performed using adaptive approaches known as state observers. A nonlinear glucose-insulin dynamic model is integrated with an unscented Kalman filter (UKF) to predict the values of intermediate state variables including PIC.⁴² To avoid degradation in the performance of the PIC estimator, the parameters of the glucose-insulin model must be updated in real-time. As a solution, an extended set of state variables is utilized for state estimation. In addition, a subset of parameters is adjusted in real-time by integrating an NRLS parameter identification algorithm. A parameter saturation function is considered to ensure that the model parameters take values inside a meaningful range and no stability issues occur. Details on the PIC estimation module are reported in the work of Hajizadeh et al.³⁴

Estimation of CHO Amount

A bolus insulin infusion is required if CHO intake is detected by the meal detection and CHO estimation module. The amount of CHO intake is necessary to estimate for handling unannounced meals, and the estimated CHO intake is used in the estimation of PIC with UKF and predicting the future glucose profile trajectory for predictive control.

The primary variable that provides real-time information on CHO intake is the CGM measurements. As part of the fuzzy model, several fuzzy logic rules are used to incorporate human knowledge into the estimation of the model output. The first fuzzy model is developed to detect meal events and triggering the second fuzzy model results in the estimation of meal amount. The structure of the fuzzy rules is comprehensively studied by Samadi et al.^19,20 Given the list of fuzzy rules in previous studies, the meal flag is calculated by a fuzzy Sugeno model. Triangular fuzzy membership functions are used for both fuzzy models, and particle swarm algorithm is used to optimize the location of membership functions. Table 2 displays the comparison between total estimated CHO amount and their ground truth values. Figure 3 visualizes the CHOs consumed by a randomly selected subject to better illustrate the performance of meal estimator.

Table 2.

Total Predicted and Actual CHO Amounts for Each Subject.

Subject no.	Actual CHO amount (g)	Predicted CHO amount (g)	The duration of experiment	No. of meal events
1	450.36	403.52	592 samples	8
2	665.04	421.66	660 samples	9
3	603.91	540.58	670 samples	9
4	735.37	627.01	662 samples	11
5	535.10	427.65	606 samples	9
6	627.97	620.14	616 samples	9
7	389.52	450.73	627 samples	7
8	392.98	427.39	664 samples	8
9	656.10	731.21	670 samples	11
10	710.38	694.11	667 samples	13
11	740.08	717.20	669 samples	14

Abbreviation: CHO, carbohydrate.

Figure 3.

Prediction of carbohydrate consumption by a randomly selected subject: The actual amounts of carbohydrate intake recorded during the clinical experiment (solid blue line), the amounts of predicted carbohydrates (red dash-dotted line).

Risk Prediction of Potential Hypoglycemic and Hyperglycemic Events

An important feature of the safety module is the ability to predict in advance hypoglycemia and hyperglycemia for their early detection of potential occurrence, suspension of insulin delivery, or avoiding insulin overdosing.^43,44 In addition, predicting potential hypoglycemia and hyperglycemia can be used to adjust the conservativeness or aggressiveness of the controller. Mamdani fuzzy logic systems were used to predict hypoglycemia and hyperglycemia as they are easy to incorporate human knowledge and leverage the glucose information extracted from the T1Ds.

The variation of the actual and estimated subcutaneous glucose concentration (mg/dL) is illustrated in Figure 4. The glucose concentration data are used in hypoglycemia and hyperglycemia prediction and in the MPC of the MPC.

Figure 4.

The variation of the actual and estimated subcutaneous glucose concentration (mg/dL). (a) The subcutaneously measured glucose concentration (blue), the subcutaneous glucose concentration estimated by UKF (red), the glucose amount in the peripheral tissue (mmol) (black). (b) The variation of adjustable parameters tuned with NRLS and UKF: The glucose transfer rate between accessible and non-accessible compartments (blue), the glucose elimination rate from plasma (red), time-of-maximum glucose appearance rate in the bloodstream (black), the time constant of the effect of energy expenditure on BGC reduction (magenta), and time-to-maximum insulin absorption (venetian red).

The potential hypoglycemia and hyperglycemia is predicted 30 minutes before the expected occurrence time. Although calculating the probability of hypoglycemia and hyperglycemia events is quite useful for automatically adjusting the aggressiveness of the MPC algorithm, converting the probabilities to the label of events is necessary for use in the safety module. The simulation results for the prediction of both hypoglycemia/hyperglycemia events are depicted in Figure 5.

Figure 5.

The prediction of hypoglycemic/hyperglycemic event predictor for a randomly selected subject with T1D (subject no. 11). The variation of the BGC (black line), the probability of hypoglycemic events calculated half an hour prior to the actual events (blue), the probability of hyperglycemic events calculated half an hour prior to the actual occurred events (red).

Adaptive Model Predictive Control System

The closed-loop control algorithm is the main module for determining the optimum amount of insulin to be infused in real-time. The manipulated variable is calculated using adaptive MPC. A constrained quadratic optimization problem is used to summarize the MPC formulation, which calculates the optimal path of manipulated variables in the control horizon and the resulting glucose concentration from trajectories in the prediction horizon. To avoid the effect of model drift and uncertainties in the model formulation, only the first value of the controller action is usually implemented by the mvAID and the sequence of computations is repeated every 5 minutes as new CGM data are reported. A major strength of MPC is accommodating constraints in the optimization problem. Constraints can be applied to the magnitude and rate of the change of input, output, and state variables and they appear in the form of equality and inequalities. More details on the MPC formulation used in this study are given in previous publications.⁴⁵

Safety Considerations

The safety constraint model is the last module to be executed before administering the amount of insulin. The main purpose of this module is to avoid any risky scenarios, such as suggesting insulin during PA or hypoglycemic episodes and avoiding any fast CHO suggestion during hyperglycemic periods. The safety constraint module evaluates several criteria on the probability of hypoglycemic/hyperglycemic episodes, the insulin dosage, and the detection of PA periods. Algorithm 1 summarizes the safety constraint module.

Algorithm 1.

The safety constraint module for final adjustment of the control action and suggesting fast CHO amount to avoid hypoglycemic episodes.

Procedure safety constraint (

H y p o_{p w},

c a r b_{p w},

M E T_{p w},

b a s a l_{k},

b o l u s_{k})

H y p o_{p w} > 0.05

then

b a s a l_{k} = 0

b o l u s_{k} = 0

\sum_{i \in p a s t 3 h o u r s} c a r b_{p w} (i) < 20

then

c a r b_{p w} (k) = 5.0

Else

c a r b_{p w} (k) = 0.0

End if
Else

c a r b_{p w} (k) = 0.0

End if
If

M E T_{p w} > 2.0

then

b a s a l_{k} = 0

b o l u s_{k} = 0

\sum_{i \in p a s t 3 h o u r s} c a r b_{p w} (i) < 20

Then

c a r b_{p w} (k) = 5.0

Else

c a r b_{p w} (k) = 0.0

End if
End if
End procedure

Results

Simulation Results

The multivariable glucose-insulin-physiological variable simulator (mGIPsim) is used for the in silico evaluation of the mvAID system.^40,46 The multivariable glucose-insulin-physiological variable simulator is based on an extended version of a glucose-insulin dynamic model,³⁶ and it has a total of 20 virtual patients with T1D with unique metabolic and physiologic characteristics and behaviors. The multivariable glucose-insulin-physiological variable simulator generates several other variables, such as heart rate and energy expenditure, as output variables for use with in silico testing of multivariable control algorithms designed to automatically handle glycemic disturbances, such as physical activities (Table 3). Each virtual subject follows a daily behavior with stochastic uncertainty according to predefined distributions for the meal timing and amounts and PA timing and intensity. The simulations are conducted for seven days in duration, with the figures showing a week in duration with glycemic events aligned at the same time to make it easy to discern the response to physical activities and meals.

Table 3.

Scheduling of Meals and Physical Activities Scenarios for the Closed-Loop Simulation.

Meal scenario for seven-day CL experiment using mGIPsim
Meal	Time		Amount
Breakfast	9:00-9:20		60 g
Lunch	12:00-12:30		75 g
Dinner	19:00-19:30		80 g
Exercise scenario for seven-day CL experiment using mGIPsim
	Time	Duration	Power	Speed	Inclination
Bicycling	16:30-17:00	30 minutes	60	—	—
Treadmill run	10:30-11:00	30 minutes	—	5	1

The primary inputs of the glucose-insulin dynamic model include insulin, CHO intake, and MET values, and the BGC measured by CGM sensor is the output. Figure 6 depicts the variation of insulin amounts calculated by UKF. The variation of PIC and insulin amount in two compartments $S_{1} (k)$ and $S_{2} (k)$ is obtained by integrating the UKF, the NRLS, and the dynamic model. The actual and estimated BGC and the concentration of glucose in peripheral tissues are presented in Figure 4a. The estimated values of BGC show a close match to the actual BGC values. Besides, the variation of glucose in non-accessible compartment in Figure 4a aligned with the actual variation of BGC, where small amount of BGC in peripheral tissues overnight and higher postprandial glucose amount can be observed. Several series of first-order filters predict the variation of glucose in meal compartments, and the estimated postprandial and overnight glucose show a reasonable trend.

Figure 6.

The variation of intermediate state variables estimated by UKF. Predicted amount of insulin in the first compartment $S_{1} (k)$ (blue), predicted amount of insulin in the second compartment $S_{2} (k)$ (red), PIC (black).

The parameter adaptation corresponding to the simulation and estimation of state variables is presented in Figure 4b. The transport rate, $K_{12}$ , and endogenous glucose production $E G P_{0}$ , are saturated during simulation that indicates the challenge of practical identifiability of model parameters. However, the set of model parameters are structurally identifiable, over parameterization, introducing decaying weights for older samples, and the sparsity of the data are the main contributors to this issue.

Figure 7 illustrates the results of a closed-loop simulation study of ten subjects with unannounced inducement of PA and CHO intake based on meal and PA descriptions described in Table 3. There are no hypoglycemia events and average BGC exceeds 200 mg/dL only once in seven days.

Figure 7.

A closed-loop simulation study of ten subjects with unannounced inducement of PA and carbohydrate intake.

Exploratory Clinical Experiments

Exploratory clinical experiments are conducted to evaluate the safety and efficacy of our mvAID system. The exploratory study was conducted at the College of Nursing of the University of Illinois Chicago, and the study was approved by the Institutional Review Board (#2018-0989). The subjects were screened and after giving written informed consent completed questionnaires, performed vitals/physiological measurements and hemoglobin A1c test, and maximal exercise stress test. A typical closed-loop experiment consisted of physical activities (treadmill run or stationary bike) and meals. The exercise and CHO intake are unannounced to the mvAID. The exploratory study involves three clinical experiments, with one experiment conducted in a crossover fashion where the glucose-only single-variable is compared with the mvAID (Figure 8). The experiments included meals and medium-intensity PA (treadmill run and stationary bike) without any announcements to the mvAID system. The closed-loop clinical experiments of the mvAID system involved two subjects with T1D (one male and one female), with ages between 35 and 45 years, bodyweight 70 – 110 kg, total daily insulin of 35-80 U, and insulin-to-CHO ratio of 6.5-7 g/U. The mvAID system is operated similar to actual use in free-living conditions, without announcements of meals and PA to the system (Figures 8 and 9).

Figure 8.

Crossover clinical experiment involving unannounced meals and physical activities with (a) glucose-only AP used in the morning and (b) mvAID used in the afternoon, showing the CGM measurements and administered basal and bolus insulin, with various glycemic events, such as meals, snacks, and physical activities indicated. The time spent in various glycemic ranges are summarized in (c) for the glucose-only AP and (d) for the mvAID system.

Figure 9.

Clinical experiment involving unannounced meals and physical activities with the mvAID; (a) the CGM measurements and administered basal and bolus insulin, with various glycemic events, such as meals, snacks, and physical activities indicated and (b) the time spend in various glycemic ranges.

Crossover experiment: Glucose-only versus multivariable AID systems

A crossover experiment was conducted with the morning period involving closed-loop control with glucose-only (single-variable architecture) AP and the afternoon period involving mvAID that integrates additional physiological signals with the glucose measurements (Figure 8). The results show that the glucose-only AP is not able to detect the occurrence of the exercise until the BGC drops, thereby increasing the risk of hypoglycemia. In contrast, the mvAID detects the occurrence of the PA and evaluates its intensity using estimates of energy expenditure and incorporates this information in predictions of the future glucose trajectory profile to proactively modulate insulin dosing, which reduces the risk of hypoglycemia.

Experiment with multivariable AID system

A subject used the mvAID system during a full day of activities, including two bouts of exercise and meal consumption that were unannounced to the system. The pilot study results again showed the mvAID proactively detect the PA and reduced the basal insulin infusion rate to mitigate the glycemic effects of the PA. The first rescue carb followed by a meal prevented the drop in glucose levels after the morning run, but the combined effect of the two rescue carbs and meal caused a mild hyperglycemia that disappeared with the stationary bike PA in the afternoon. No insulin bolus was given after 1 PM, and the basal insulin was shut off at the beginning of the exercise by the mvAID system.

Discussion of Results

This study showed the effectiveness of using additional physiological signals from a wristband device to handle unannounced glycemic disturbances in real-time. Results of previous in silico studies demonstrate that closed-loop control of BGC can be improved by integrating daily behavior and activity pattern information. Although planned exercise may be handled by considering historical trends and patterns, the capabilities of techniques that rely only on historical information are limited. Rather, the historical information can be synergistically incorporated with real-time estimation of disturbances, such as PA from additional physiological signals reported by the wristband device, which is more reliable approach for proactive control of BGC levels.

Developed modules include the multistep predictor of hypoglycemia/hyperglycemia, the safety constraint modules, CHO estimator modules, the state-space system identification module, and PIC estimator module, which are integrated to complete the feedback-feedforward features of the mvAID. The stability analysis of the system was studied using symbolic and numerical computations. A preliminary one-week in silico study of the proposed mvAID indicates more than 83% time in range calculated over all ten subjects. On average, a few number of hypoglycemic/hyperglycemic episodes were obtained as well by calculating < 4% and > 13% time-below-range and time-above-range, respectively.

The crossover clinical experiment (Figure 8) shows the controller performance effectiveness in the presence of different unannounced disturbance scenarios, including treadmill run, stationary bike, and CHO consumption. The closed-loop clinical experiment indicates the safety and efficacy of the mvAID and the improvement in performance over the single-variable AID system. The single-variable (CGM) AID system is not able to proactively take safe actions to mitigate the effects of unannounced disturbances, and as a result the AID system is more aggressive in infusing insulin and less capable in avoiding the occurrence of hypoglycemic events following unannounced PA. In contrast, the mvAID demonstrates well-controlled BGC because of the ability of automatically detecting and handling unannounced PA through the incorporation of additional physiological signals. One limitation of this work is the small number of exploratory clinical experiments conducted, though the pilot study was useful in demonstrating the safety and efficacy of the mvAID. Future studies will evaluate the mvAID in randomized controlled trials.

More research work in the future will be required to tailor the aggressiveness of the mvAID to individual subjects. The adaptive modeling of glucose-insulin dynamics and the adaptive MPC algorithm mean that performance will improve over time. This is a limitation of the current work, as the exploratory study had short duration that did not enable the testing and evaluation of the mvAID over long-term outpatient experiments, though the in silico simulation studies show the long-term performance of the adaptive learning approach and the mvAID for long time durations. The other improvements relate to the longer duration step-ahead prediction of BGC while the stability analysis of the state-space model is guaranteed.

Conclusion

An mvAID system that supplements the glucose concentration measurements with additional physiological signals from a wristband device to automatically handle unannounced PA is developed and successfully translated to exploratory clinical study. The in silico evaluation of the mvAID system and the exploratory clinical study involving unannounced PA and meals illustrate the safety and efficacy of the mvAID to handle unannounced glycemic disturbances, reduction of hypoglycemia risk during PA, and increase of time in range.

Footnotes

Acknowledgements

Financial support from NIH grants R01DK135116, 1R01DK130049 and 1DP3DK101075, and JDRF grants 2-SRA-2017-506-M-B and 3-APF-2022-1134-A-N made possible through collaboration between JDRF and The Leona M. and Harry B. Helmsley Charitable Trust are gratefully acknowledged.

Abbreviations

AID, automated insulin delivery; BGC, blood glucose concentration; CGM, continuous glucose monitoring; CHO, carbohydrate; KPLS, Kernel partial least squares; MDI, multiple insulin injection; MET, metabolic equivalents of task; MPC. model predictive control; mvAID, multivariable insulin delivery; NRLS, nonlinear recursive least squares; PA, physical activity; PIC, plasma insulin concentration; RNN, recurrent neural network; SAP, sensor augmented pump; T1D, type 1 diabetes; UKF, unscented Kalman filter.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by NIH grants R01DK135116, 1R01DK130049 and 1DP3DK101075, and JDRF grants 2-SRA-2017-506-M-B and 3-APF-2022-1134-A-N made possible through collaboration between JDRF and The Leona M. and Harry B. Helmsley Charitable Trust.

ORCID iDs

Mohammad Reza Askari

Mudassir Rashid

Ali Cinar

References

Phillip

Nimri

Bergenstal

, et al. Consensus recommendations for the use of automated insulin delivery technologies in clinical practice. Endocrine Reviews. 2023;44(2):254-280.

Latres

Finan

Greenstein

Kowalski

Kieffer

. Navigating two roads to glucose normalization in diabetes: automated insulin delivery devices and cell therapy. Cell Metabolism. 2019;29(3):545-563.

Renard

. Automated insulin delivery systems: from early research to routine care of type 1 diabetes. Acta Diabetol. 2023;60(2):151-161.

Sherr

Heinemann

Fleming

, et al. Automated insulin delivery: benefits, challenges, and recommendations—a consensus report of the Joint Diabetes Technology Working Group of the European Association for the Study of Diabetes and the American Diabetes Association. Diabetes Care. 2022;45(12):3058-3074.

Askari

Hajizadeh

Rashid

Hobbs

Zavala

Cinar

. Adaptive-learning model predictive control for complex physiological systems: automated insulin delivery in diabetes. Annu Rev Contr. 2020;50:1-12.

Kesavadev

Saboo

Krishna

Krishnan

. Evolution of insulin delivery devices: from syringes, pens, and pumps to DIY artificial pancreas. Diabetes Ther. 2020;11(6):1251-1269.

Haidar

Tsoukas

Bernier-Twardy

, et al. A novel dual-hormone insulin-and-pramlintide artificial pancreas for type 1 diabetes: a randomized controlled crossover trial. Diabetes Care. 2020;43(3):597-606.

Tsoukas

Majdpour

Yale

, et al. A fully artificial pancreas versus a hybrid artificial pancreas for type 1 diabetes: a single-centre, open-label, randomised controlled, crossover, non-inferiority trial. Lancet Digit Health. 2021;3(11):e723-e732.

Kesavadev

Srinivasan

Saboo

Krishna

B M

Krishnan

. The do-it-yourself artificial pancreas: a comprehensive review. Diabetes Ther. 2020;11(6):1217-1235.

10.

Cobelli

Renard

Kovatchev

. Artificial pancreas: past, present, future. Diabetes. 2011;60(11):2672-2682.

11.

Moon

Jung

Park

. Current advances of artificial pancreas systems: a comprehensive review of the clinical evidence. Diabetes Metab J. 2021;45(6):813-839.

12.

Lee

Kim

Park

Jin

Park

. Toward a fully automated artificial pancreas system using a bioinspired reinforcement learning design: in silico validation. IEEE J Biomed Health Inform. 2021;25(2):536-546.

13.

Boughton

Hovorka

. Is an artificial pancreas (closed-loop system) for Type 1 diabetes effective? Diabet Med. 2019;6(3):279-286.

14.

Tyler

Mosquera-Lopez

Young

El Youssef

Castle

Jacobs

. Quantifying the impact of physical activity on future glucose trends using machine learning. iScience. 2022;25(3):103888.

15.

Sanz

García

Díez

Bondia

. Artificial pancreas system with unannounced meals based on a disturbance observer and feedforward compensation. IEEE T Contr Syst Technol. 2021;29(1):454-460.

16.

Garcia-Tirado

Diaz

Esquivel-Zuniga

, et al. Advanced closed-loop control system improves postprandial glycemic control compared with a hybrid closed-loop system following unannounced meal. Diabetes Care. 2021;44(10):2379-2387.

17.

Zheng

Kleinberg

. Automated meal detection from continuous glucose monitor data through simulation and explanation. J Am Med Inform Assoc. 2019;26(12):1592-1599.

18.

Cameron

Niemeyer

Bequette

. Extended multiple model prediction with application to blood glucose regulation. J Process Contr. 2012;22(8):1422-1432.

19.

Samadi

Rashid

Turksoy

, et al. Automatic detection and estimation of unannounced meals for multivariable artificial pancreas system. Diabet Technol Ther. 2018;20(3):235-246.

20.

Samadi

Turksoy

Hajizadeh

Feng

Sevil

Cinar

. Meal detection and carbohydrate estimation using continuous glucose sensor data. IEEE J Biomed Health Inform. 2017;21(3):619-627.

21.

Cameron

Niemeyer

Wilson

, et al. Inpatient trial of an artificial pancreas based on multiple model probabilistic predictive control with repeated large unannounced meals. Diabet Technol Therap. 2014;16(11):728-734.

22.

Cameron

Buckingham

, et al. Closed-loop control without meal announcement in Type 1 diabetes. Diabet Technol Therap. 2017;19:527-532.

23.

Forlenza

Cameron

, et al. Fully closed-loop multiple model probabilistic predictive controller artificial pancreas performance in adolescents and adults in a supervised hotel setting. Diabetes Technol Ther. 2018;20(5):335-343.

24.

Zaharieva

Messer

Paldus

O’Neal

Maahs

Riddell

. Glucose control during physical activity and exercise using closed loop technology in adults and adolescents with Type 1 diabetes. Can J Diabet. 2020;44(8):740-749.

25.

Tagougui

Taleb

Rabasa-Lhoret

. The benefits and limits of technological advances in glucose management around physical activity in patients type 1 diabetes. Front Endocrinol. Frontiers Media SA;. 2019;9:818.

26.

Riddell

Scott

Fournier

, et al. The competitive athlete with type 1 diabetes. Diabetologia. 2020;63(8):1475-1490.

27.

Sevil

Rashid

Hajizadeh

, et al. Assessing the effects of stress response on glucose variations. In 2019 IEEE 16th International Conference on Wearable and Implantable Body Sensor Networks (BSN), Chicago, IL, May 19-22, 2019.

28.

Sevil

Rashid

Askari

Maloney

Hajizadeh

Cinar

. Detection and characterization of physical activity and psychological stress from wristband data. Signals. 2020;1:188–208.

29.

Sevil

Rashid

Hajizadeh

, et al. Discrimination of simultaneous psychological and physical stressors using wristband biosignals. Comput Methods Programs Biomed. 2021;199:105898.

30.

Askari

Abdel-Latif

Rashid

Sevil

Cinar

. Detection and classification of unannounced physical activities and acute psychological stress events for interventions in diabetes treatment. Algorithms. 2022;15:352.

31.

Abdel-Latif

Askari

Rashid

, et al. Multi-task classification of physical activity and acute psychological stress for advanced diabetes treatment. Signals. 2023;4:167–192.

32.

Askari

Hajizadeh

Sevil

, et al. Application of neural networks for heart rate monitoring. IFAC: PapersOnline. 2020;53(2):16161-16166.

33.

Askari

Rashid

Sevil

, et al. Artifact removal from data generated by nonlinear systems: heart rate estimation from blood volume pulse signal. Ind Eng Chem Res. 2020;59(6):2318-2327.

34.

Hajizadeh

Rashid

Samadi

, et al. Adaptive and personalized plasma insulin concentration estimation for artificial pancreas systems. J Diabetes Sci Technol. 2018;12(3):639-649.

35.

Hajizadeh

Rashid

Turksoy

, et al. Incorporating unannounced meals and exercise in adaptive learning of personalized models for multivariable artificial pancreas systems. J Diabetes Sci Technol. 2018;12(5):953-966.

36.

Hovorka

Canonico

Chassin

, et al. Nonlinear model predictive control of glucose concentration in subjects with type 1 diabetes. Physiol Meas. 2004;25(4):905-920.

37.

Garcia-Tirado

Brown

Laichuthai

, et al. Anticipation of historical exercise patterns by a novel artificial pancreas system reduces hypoglycemia during and after moderate-intensity physical activity in people with type 1 diabetes. Diabet Technol Therap. 2020;23(4):277-285.

38.

Dempster

Petitjean

Webb

. ROCKET: exceptionally fast and accurate time series classification using random convolutional kernels. Data Min Knowl Discov. 2020;34(5):1454-1495.

39.

Sevil

Rashid

Maloney

, et al. Determining physical activity characteristics from wristband data for use in automated insulin delivery systems. IEEE Sens J. 2020;20(21):12859-12870.

40.

Rashid

Samadi

Sevil

, et al. Simulation software for assessment of nonlinear and adaptive multivariable control algorithms: glucose–insulin dynamics in Type 1 diabetes. Comput Chem Eng. 2019;130:106565.

41.

Hobbs

Samadi

Rashid

, et al. A physical activity-intensity driven glycemic model for type 1 diabetes. Comput Methods Programs Biomed. 2022;226:107153.

42.

Hajizadeh

Rashid

Turksoy

, et al. Plasma insulin estimation in people with type 1 diabetes mellitus. Ind Eng Chem Res. 2017;56(35):9846-9857.

43.

Paiement

Frenette

, et al. Is better understanding of management strategies for adults with type 1 diabetes associated with a lower risk of developing hypoglycemia during and after physical activity. Can J Diabetes. 2022;46(5):526-534.

44.

Paldus

Morrison

Lee

Zaharieva

Riddell

O’Neal

. Strengths and challenges of closed-loop insulin delivery during exercise in people with type 1 diabetes: potential future directions. J Diabet Sci Technol. 2022;17:1077-1084.

45.

Rashid

Sun

Askari

Cinar

. Model identification and model predictive control of biopharmaceutical and biomedical systems. IFAC: PapersOnline. 2022;55(7):51-56.

46.

Rashid

Samadi

Sevil

, et al. Virtual patients: an enabling technology for multivariable control of biomedical systems. IFAC: PapersOnline. 2020;53(2):16233-16238.

Multivariable Automated Insulin Delivery System for Handling Planned and Spontaneous Physical Activities

Abstract

Background:

Methods:

Results:

Conclusions:

Keywords

Introduction

Methods

Structure of mvAID System

Components of mvAID System

Initialization of the mvAID System

Estimating Disturbances Affecting BGC Regulation

Reconciliation and Processing of Wristband Physiological Signals

Reconciliation of Glucose Measurements

Detecting PA and Quantifying Its Intensity With Energy Expenditure

Real-Time Adaptive Estimation of PIC

Estimation of CHO Amount

Risk Prediction of Potential Hypoglycemic and Hyperglycemic Events

Adaptive Model Predictive Control System

Safety Considerations

Results

Simulation Results

Exploratory Clinical Experiments

Crossover experiment: Glucose-only versus multivariable AID systems

Experiment with multivariable AID system

Discussion of Results

Conclusion

Footnotes

Acknowledgements

Abbreviations

Declaration of Conflicting Interests

Funding

ORCID iDs

References