This paper is structured as follows: "Methodology" describes the proposed methodology, focusing on the detection and prediction models for hypoglycemia and hyperglycemia models. Section "Results and discussion" introduces the mathematical framework of the glucose-insulin metabolic system. Section "Hardware implementation" discusses the design of an event-triggered controller aimed at minimizing blood glucose variability. Section 5 presents the results and discussion, while "Conclusion and challenges" concludes the paper by outlining the key challenges in this field.
This paper presents SDHAP, an advanced technology designed to help individuals with T1DM manage glucose levels more effectively by automating insulin and glucagon delivery based on real-time glucose monitoring. The innovation of SDHAP lies in its event-triggered algorithm, which dynamically calculates and administers only the necessary hormone doses, enabling precise, responsive glucose regulation through automated dual-hormone infusion. Figure 1 shows the proposed methodology for event- triggered SDHAP. The main components of SDHAP are:
The T1DiabetesGranada dataset is openly accessible through the Zenodo repository and offers a highly detailed, four-year longitudinal dataset for 736 patients with Type 1 diabetes, captured in Granada, Spain. The glucose monitoring data was collected every 15 min using devices like the Freestyle Libre. This frequency provides over 2,57,000 patient days of data, allowing for fine-grained analyses of glucose trends and fluctuations across different timescales. It also includes a comprehensive demographic as well as relevant clinical and biochemical data. Table 1 provides detailed information related to the dataset. Figure 2. shows the changes in CGM measurement over time for a particular period.
Continuous glucose data were collected as a time series input, consisting of 124 values per day, recorded at five-minute intervals. A range of time series features was extracted across temporal, statistical, and spectral domains utilizing the feature engineering machine learning framework outlined by. In addition, diabetes-related metrics were included, such as Time in Range (TIR), Time Below Range (TBR), and Time Above Range (TAR), which represent the percentage of time that glucose levels remain within, below, or above the target range of 70 to 180 mg/dL. High and low glucose indexes were also incorporated. To tackle the issue of imbalanced data, which can negatively impact classifier performance, feature rescaling was applied during the training process.
Additionally, the Synthetic Minority Over-Sampling TechniquE (SMOTE) was used to generate synthetic samples for the minority class. Before applying SMOTE, the dataset was imbalanced, with hyperglycemic events significantly outnumbering hypoglycemic events. The distribution showed a higher frequency of hyperglycemic cases. After applying SMOTE, synthetic samples were generated for the hypoglycemic class by interpolating closely related positive instances, balancing the dataset. This improved model performance by increasing sensitivity to hypoglycemia, reducing classification errors, and enhancing overall accuracy, particularly for the minority class.
In this section, BG levels detected from the time series and diabetes-related metrics were used for further analysis. Various machine learning models were discussed in the literature relating the importance of AI and machine learning in classification and multi-step ahead prediction of glucose by The following two machine learning algorithms SVM and KNN were used.
Support Vector Machine (SVM) is a powerful supervised learning algorithm that constructs an optimal decision boundary to separate different classes in the feature space. In this study, SVM was employed to classify blood glucose levels into hypoglycemic, normoglycemic, and hyperglycemic states using features derived from the T1DiabetesGranada dataset. Its effectiveness in handling high-dimensional and non-linear physiological data makes it well-suited for glucose classification tasks. SVM provides high accuracy in classifying glucose levels and identifying early signs of glycemic events, which supports its integration into the SDHAP system for timely hormone administration.
Multi-class SVM, the One-vs-One or One-vs-All which divides the multi-class problem into multiple binary class problems. In multi-class SVM, the One-vs-All multi-class methods, box limitations, and the kernel function are all important. The hyperplane's where is weight vector, are input vectors and represent the class label and is a slack variable. For separable data, the optimal margin length is and 'b' is the bias term. The objective of SVM is to minimize the subject to , and 'E' represents the box constraint. Figure 3 shows the classification output using a multi-class SVM classifier.
K-Nearest Neighbor (KNN) is a simple yet effective non-parametric algorithm that classifies new data points based on the majority label of their nearest neighbors. In our framework, KNN was used to classify blood glucose readings by comparing them to past observations with similar feature patterns. Its ability to capture local data structures makes it particularly useful in detecting transitions into hypo- or hyperglycemia.
The Euclidean distance between the two data points indicates how comparable the two data points are. The Euclidean distance is calculated using closeness.where, '' indicates the dimension of the dataset and ,, .........., . Figure 4 shows glycemic classification using KNN.
Time-series forecasting techniques are widely used to predict BG levels, which leverage historical glucose data to forecast future BG trends, aiding in the prevention of glycemic events like hypoglycemia and hyperglycemia. Common approaches include autoregressive models, machine-learning methods, and advanced neural networks. In our study, we apply ARIMA and GRU models to forecast BG levels, leveraging both linear and non-linear patterns to support proactive glycemic management and minimize risks of hypoglycemia and hyperglycemia.
The Auto-Regressive Integrated Moving Average (ARIMA) model, introduced by Box and Jenkins, is a time series analysis technique widely used for predicting trends in BG data. The ARIMA model's structure includes three parameters: p - autoregressive order, d - differencing order, and q - moving average order. The following equation describes the model:
where: -- current value of the time series; c -- constant term; -- parameters for the AR part; -- parameters for the MA part; -- errorterm.
The adaptability of this model helps to maintain predictive accuracy, and stability, accommodating daily variations in glucose patterns due to meals, exercise, or insulin use. This refined approach enhances the reliability of CGM-based BG forecasting, helping to predict and mitigate hypoglycemic and hyperglycemic events.
The ARIMA model was chosen for its effectiveness in modeling stationary and autocorrelated glucose trends. Prior studies, such as Yang et al. (2019), have demonstrated ARIMA's utility in forecasting short-term glucose levels and hypoglycemia events in Type 1 Diabetes patients [Yang et al., Journal of Healthcare Engineering, 2019]. ARIMA, as a traditional time series model, can be employed for forecasting short-term glucose dynamics in a structured and interpretable manner. Its strength lies in handling linear and stationary blood glucose trends, which can support basic predictive control mechanisms in artificial pancreas systems. For instance, in a dual-hormone closed-loop system, ARIMA can be used to anticipate slight deviations in glucose levels and guide minor insulin or glucagon dosing decisions. However, its univariate structure and reliance on manual preprocessing steps such as differencing limit its application in modeling complex patient-specific glycemic responses. Furthermore, ARIMA lacks adaptability for real-time, multivariate, and nonlinear interactions among physiological variables -- such as hormone kinetics, meal intake, and circadian variation -- which are crucial in an event-triggered drug delivery system. Thus, ARIMA may be better suited as a baseline comparator rather than a core component of adaptive pancreas control.
In contrast, the GRU model, a type of Recurrent Neural Network, was selected for its proven ability to handle non-linear dependencies and learn long-term patterns in physiological time-series data. Mohebbi et al. (2020) showed that GRU outperformed ARIMA in predicting BG levels over longer horizons, especially under noisy and variable conditions [Mohebbi et al.,2002]. GRU networks offer a powerful alternative by learning nonlinear, long-term dependencies in blood glucose dynamics from multivariate data, making them well-suited for real-time, personalized dual-hormone delivery. Unlike ARIMA, GRUs automatically learn temporal patterns from historical glucose, insulin, glucagon, physical activity, and meal-related inputs, enabling robust forecasting even in non-stationary and highly individualized glycemic conditions. In the context of event-triggered control, GRUs can dynamically adapt to sharp glucose excursions and guide timely dosing interventions by predicting upcoming hypo- or hyperglycemic events. Their ability to process large datasets and handle multivariate physiological signals in real-time makes them ideal for building intelligent, adaptive pancreas systems tailored to individual patient responses.
The Gated Recurrent Unit (GRU), introduced by Kyunghyun Cho and colleagues in 2014, is widely used for time-series due to its ability to capture temporal dependencies effectively. For blood glucose prediction, a GRU model captures complex patterns over time, including hypoglycemic and hyperglycemic events. GRUs use two gates: the reset gate (r = σ(W.[h,x] + b)) to decide how much past information of glucose to keep, and the update gate (z = σ(W.[h,x] + b)) to balance past and current glucose information. These gates combine to calculate the candidate hidden state (h҇ = tanh(W.[r*h,x]+b)) and final hidden stage (h = z * h + (1-z * h҇ ), which represents the predicted glucose levels. This design supports accurate blood glucose forecasting, enhancing diabetes management and reducing adverse glycemic events. Figure 5 shows the GRU architecture that efficiently captures temporal dependencies using update and reset gate. It is particularly well-suited for modeling time-series data like blood glucose levels due to its effectiveness in handling sequential information. Table 2 shows the GRU parameters used for blood glucose prediction model.
The predicted accuracy of blood glucose levels using the dataset is evaluated through the following performance metrics:
where represent the predicted and actual glucose values at time index i, respectively, and n denotes the total number of data points evaluated.
The BMM parameters were derived from the insulin-glucose model, and the parameters related to hypoglycemia and hyperglycemia are discussed in the following section.
The Bergman Minimal Model (BMM) is a mathematical framework that simplifies the analysis of glucose and insulin dynamics within the human body. The core parameters of the BMM include glucose effectiveness, insulin sensitivity, insulin action, and glucose appearance rate. This paper aims to develop a mathematical model derived from the differential equations of BMM, as outlined by. The parameters of the BMM model are described in Table 3.
, ,, h, n, γ parameter values of the BMM model are crucial in modeling hypoglycemic and hyperglycemic condition.
For patients experiencing hypoglycemia, the BMM model has been adjusted to account for the absence of insulin infusion. In this modification, the term representing exogenous insulin infusion U(t), has been removed from the original BMM equations.
For patients experiencing hyperglycemia, the BMM model has been adapted by eliminating the glucose infusion term G. This modification reflects scenarios where glucagon infusion is focused on managing elevated blood glucose levels.
Blood glucose regulation in insulin-dependent diabetic patients can be managed either through open-loop control, which involves multiple daily insulin injections, or through closed-loop control using APs. Insulin and glucagon are the primary hormones regulating blood glucose levels. This physiological process has been mathematically modeled, with several control-oriented models providing valuable insights and visualization of this complex system, as demonstrated by. The SDHAP system leveraged the event-triggered responses to dynamically adjust both insulin and glucagon infusion. By combining the event-triggered response mechanism with the dual-hormone model, the system effectively mimicked the physiological processes of glucose regulation, offering a more refined and adaptive treatment option for T1DM patients.
In the context of glucose regulation, a PI (Proportional - Integral) controller could be used to maintain glucose levels within a desired range. The PI controller would adjust the insulin dosing/glucagon dosing based on the difference between the measured glucose level (the output) and the target glucose levels (the reference point).
where u (t) represents the overall control action (insulin/glucose infusion rate) of the PI controller, K and K represent proportional, integral gain respectively.
A comprehensive overview of the control strategy used in improving glucose regulation was addressed by. In an AP system, MPC treats glucose levels as the dependent variable and adjusts independent factors such as body mass index, carbohydrate intake, and insulin. This real-time modeling, updating every few minutes, allows the MPC algorithm to address external disturbances like meals and physical activity while compensating for the inherent delay in BG monitoring. The versatility of MPC enables an event-triggered model to handle challenges like inter-patient variability, unannounced disturbances, and safety constraints. By linearizing the glucose-insulin relationship into a state-space form, MPC ensures precise glucose regulation, making it a crucial tool for enhancing the safety and efficacy of SDHAP.
For this study, a detailed explanation of event-triggered response protocol fully discussed by was used to manage blood glucose levels and associated vital parameters in the patients. Figure 6 shows the block diagram of the same.
The effectiveness of the proposed controller was evaluated using the following performance metrics:
The block diagram of the FB-FF controller is shown on Fig. 7. The FF controller was designed to reject known disturbances in glucose due to food intake/exercise in hyperglycemia and hypoglycemia conditions. The transfer function of FF controller is given below: -
In this work, a static feedback controller with a gain of - 1 was designed to reject the disturbance by predicting the effect of disturbances variables on blood glucose.
Figure 8 depicts an MPC system tailored for blood glucose regulation. The system represents the mathematical model for human glucose-insulin dynamics in the human body, where blood glucose levels are the output signal. The MPC aims to maintain blood glucose at a target level (set point) by calculating an optimal control signal (e.g., insulin or glucagon administration). It considers measured disturbances such as food intake, exercise, and insulin dosing, directly fed into the controller, and unmeasured disturbances managed indirectly through a feedback mechanism. The MPC operates using two main strategies: feedforward control, which proactively adjusts the control signal based on measured disturbances, and feedback control, which continuously corrects deviations by comparing the actual blood glucose levels (output) to the target.