(A-37) Patient-specific deep offline artificial pancreas for blood glucose regulation in type 1 diabetes

Introduction:: Due to insufficient insulin secretion, patients with type 1 diabetes mellitus (T1DM) are prone to blood glucose fluctuations ranging from hypoglycemia to hyperglycemia. While dangerous hypoglycemia may lead to coma immediately, chronic hyperglycemia increases patients’ risks for cardiorenal and vascular diseases in the long run. In principle, an artificial pancreas -- a closed-loop insulin delivery system requiring patients manually input insulin dosage according to the upcoming meals -- could supply exogenous insulin to control the glucose levels and hence reduce the risks from hyperglycemia. However, insulin overdosing in some type 1 diabetic patients, who are physically active, can lead to unexpected hypoglycemia beyond the control of common artificial pancreas. Therefore, it is important to take into account the glucose decrease due to physical exercise when designing the next-generation artificial pancreas.


Materials and Methods:: Motivated by the features highlighted for next-generation AP design, we propose a novel framework to design a patient-specific artificial pancreas using up-to-date hardware and software technologies with digital twins. By simultaneously considering patient-specificity, meal intakes, insulin infusion, and most importantly physical exercises, we built and optimized patient-specific glucose levels using a system of ordinary differential equations (ODE) developed by Roy and Parker, time-dependent systems biology informed neural networks (SBINN), wearable sensor data from the OhioT1DM dataset and two offline reinforcement learning (RL) algorithms. Specifically, when training the time-dependent SBINN, we implemented self-adaptive weight to automatically adjust the weight of each loss terms, and inversely solve for the dynamics of hidden states and hidden parameters. We designed the offline RL agents such that they learn the insulin dosage with only a short sequence of past glucose levels without any meal or exercise announcements, which significantly further the step towards an authentic closed-loop system for artificial pancreas design.

Results, Conclusions, and Discussions:: In summary, by seamlessly integrating real-world data from wearable sensors, time-dependent systems biology informed neural networks and deep offline reinforcement learning algorithms, we developed a universal framework that could shed light on the patient-specific digital twin designs where adequate biological numerical models are established and enough medical data from sensors are available. In the context of type 1 diabetes, we found SBINN could successfully infer the hidden physiological parameters governing glucose-insulin dynamics and accurately reconstruct the corresponding glucose trajectories at patient's level. Despite of a limited number of patients, but with a complete dataset, we found TD3+BC offline reinforcement learning is not only more effective but also easier to tune than BCQ and it does help significantly stabilize the glucose level with only offline training.

Despite of the aforementioned possible improvements, our framework shows inspiring translational potentials in precision medicine by enriching digital models with extensive medical data collected from wearable devices. For example, with a few slight modifications on the ODE model to address the insulin resistance of tissue, we could model and optimize the insulin usage in patients with type 2 diabetes who develop insulin resistance. Efforts in hardware designs of wearable sensors are also inspiring and beneficial to our model development. For example, novel wearable devices are being designed to probe insulin and other metabolites that are impossible to quantify with existing techniques with a higher resolution. Furthermore, the rapid development in non-invasive wearable devices tremendously increases patients' compliance in using wearable devices, which help reduce data sparsity with more frequent sensor readouts, and consequently improve the performance of our framework. These next-generation sensors will provide us with high-resolution and unprecedented data, which will significantly enhance the performance of our framework.


Acknowledgements (Optional): :

References (Optional): :