Abstract: While a typical way for diabetes therapy is discrete insulin infusion based on long-time interval measurement, in this paper, we design a closed-loop control system for continuous drug infusion to improve the traditional discrete methods and make diabetes therapy automatic in practice. By exploring the accumulative function of drug to insulin, a continuous injection model is proposed. Based on this model, proportional-integral-derivative (PID) and fuzzy logic controllers are designed to tackle a control problem of the resulting highly nonlinear plant. Even with serious disturbance of glucose, such as nutrition absorption at meal time, the proposed scheme can perform well in simulation experiments.
Introduction
Managing Type I diabetes is traditionally a "discrete" process: patients measure their blood glucose a few times a day and self-administer 3–5 insulin injections. This approach is reactive and prone to the "rollercoaster effect"—significant fluctuations where glucose levels swing between dangerous highs (hyperglycemia) and lows (hypoglycemia).
The challenge in automating this via a "closed-loop" system (an artificial pancreas) lies in the complex, nonlinear relationship between insulin and glucose. Most existing models treat insulin doses as isolated events. However, in a continuous infusion scenario, the effects of the drug overlap and accumulate over time. This paper seeks to model this "accumulative function" and design a controller that can handle the resulting nonlinearities to keep a patient within a tight healthy glucose range (60–140 mg/dL).
Closed-Loop Control System Design
The proposed system architecture follows a standard feedback control loop, where the "Plant" represents the patient's physiological response to insulin and glucose disturbances (like meals). The controller continuously adjusts the drug delivery rate of a pump based on the error between the required blood glucose level and the measured level.
The authors break the problem into two distinct modeling components before designing the controllers:
- The Accumulative Drug Model: Unlike previous works that ignore overlapping doses, the authors propose an integral form model. This calculates the current plasma insulin concentration by integrating the dose of the drug (specifically Lispro insulin) over a 360-minute sliding window, accounting for how the drug's effect fades over time.
- The Glucose-Insulin Plant: They utilize the Bergman "Minimal Model," a standard mathematical representation of how insulin promotes glucose uptake. They linearize aspects of this model to make it more manageable for control theory applications.
Controller Implementation and Performance
With the model in place, two control strategies were compared in a simulation environment using pharmacokinetic data for Lispro and standardized Bergman model parameters for a typical Type I diabetic patient:
- PID Controller: A classical engineering approach that adjusts insulin delivery based on the current error (Proportional), the history of the error (Integral), and the predicted future error (Derivative).
- Fuzzy Logic Controller: A rule-based system that mimics human reasoning (e.g., "If glucose is high and rising fast, then deliver much more insulin"). They defined 10 specific rules based on glucose concentration and its rate of change.
Key Results
Both controllers successfully kept blood glucose within the target range of 60–140 mg/dL during a simulated 24-hour period with three meals. As shown in the simulation results, the PID controller significantly dampens the glucose spikes caused by meal disturbances compared to the uncontrolled baseline.
Quantitatively, the PID controller maintained glucose between 66–135 mg/dL, while the Fuzzy Logic controller showed slightly better precision for scheduled events, maintaining levels between 71–128 mg/dL. The system achieved these results with very small, continuous adjustments (generally less than 0.4 units of insulin per minute), which is much gentler on the body than large bolus injections.
Discussion and Limitations
The paper highlights that the mathematical "memory" of previous doses (the integral model) is essential for preventing insulin stacking, which could lead to fatal hypoglycemia. While Fuzzy Logic provided a tighter control range during standard meal cycles, the PID controller proved more robust when faced with unexpected disturbances (like an unscheduled snack) because it is mathematically tied to the rate of change.
However, this study is purely simulation-based. Real-world factors like sensor noise, subcutaneous glucose sensing lag (5–15 minutes), and the physical constraints of a pump were not modeled. Future work should consider more adaptive controllers to handle variables like exercise, stress, and illness, as well as optimizing sampling frequency to balance patient safety with device battery life.
Summary updated