Keywords: Physiological modeling
Abstract: Prometeus (Preterm Brain-Oxygenation and Metabolic EU-Sensing: Feed the Brain) develops innovative technology for personalized nutrition in premature neonates. Central to this is the Nutritional Clinical Advisor, an adaptive algorithm providing optimal parenteral nutrition plans to enhance brain oxygenation. This study tests a closed-loop PID controller for regulating glucose levels in virtual patients, offering insights into its effectiveness and guiding future NCA development to support neonatal care.

Keywords: New control system algorithms for Automatic Insulin Delivery
Abstract: We introduce a fully automated insulin delivery (AID) system for type 2 diabetes (T2D) patients. This system initially titrates a basal insulin rate for the patient using an I-controller and subsequently introduces a PD-controller for postprandial management. The algorithm is adaptive, ideally allowing it to be applicable to any patient without prior calibration. Additionally, we present a simulation model for T2D patients, combining the existing Hovorka type 1 diabetes (T1D) model and the UVA/Padova T2D model, enabling simulation of T2D patients undergoing exogenous insulin therapy. In a trial with 84 virtual patients simulated using this T2D model, the proposed AID system demonstrated safe and effective titration and treatment, indicating its potential for providing safe and convenient T2D management.

Keywords: New control system algorithms for Automatic Insulin Delivery, Adjuvant therapies to insulin, Other technological contributions to diabetes management
Abstract: This study investigates the use of Reinforcement Learning (RL) to optimize insulin infusion in Automated Insulin Delivery (AID) systems. The Deep Deterministic Policy Gradient (DDPG) algorithm is evaluated as a potential control strategy and compared to Model Predictive Control (MPC) and PID controllers using a highly realistic in silico simulation model. Simulations for adult Type 1 Diabetes (T1D) patients demonstrate that the DDPG controller effectively reduces hypo- and hyperglycemic episodes and improves glucose stability.

Keywords: New control system algorithms for Automatic Insulin Delivery, Adjuvant therapies to insulin, Physiological modeling
Abstract: Automated Insulin delivery has become standard of care for managing glycemic control in Type 1 Diabetes (T1D), and has been shown to dramatically improve overall glycemic control. However, postprandial hyperglycemia remains a challenge, even with AID systems that require meal announcements. Rapid- and Ultra-rapid-acting insulin analogs that could improve or even prevent postprandial hyperglycemia are now available for both research and clinical care. However, it has been shown that adjusting AID system’s parameters is necessary for maximizing benefits when using accelerated insulins within these systems. In this work, the safety and performance effect of accelerating insulin within AIDANET, a state-of-the-art AID system is evaluated in-silico to address postprandial hyperglycemia generated by unannounced meals. Pharmacokinetic parameters are leveraged to include the use of a novel ultra-rapid insulin formulation within the UVA/Padova Type 1 Diabetes simulator. Additionally, two in-silico scenarios are used to assess the performance of AIDANET with (i) faster insulin analogs and (ii) more concentrated formulations. Results show that AIDANET's built-in adaptive capabilities allow it to adjust its aggressiveness to accommodate faster PK and more concentrated formulations, achieving clinically significant improvements in glycemic control outcomes and maximizing the benefits of alternative formulations

Keywords: New control system algorithms for Automatic Insulin Delivery, Other technological contributions to diabetes management
Abstract: Glucose control in Type 1 Diabetes (T1D) is a challenging task due to several factors among which intra- and inter-patient variability play a key role. This work introduces a Periodic Multiple Model Predictor (PeMMP) integrated in a Periodic Model Predictive Control (PeMPC) to enhance glucose regulation. The PeMMP is able to account for time-varying dynamics by switching smoothly between multiple models. In a case study including a single patient of the UVA/Padova simulator, the PeMMP improves prediction accuracy compared to the traditional Daily Model Predictor (DMP) (fitting index 81.9% vs 39.7%). These more accurate predictions lead to an improvement in control performances when exploited in a PeMPC especially in the postprandial periods increasing the time in range (93.45% to 96.31%) and decreasing hypoglycemia (1.42% to 0%) and hypo treatments (10 to 0). These results highlight PeMPC as a promising approach for better managing T1D, its use on the entire in silico population is currently under study.

Keywords: New control system algorithms for Automatic Insulin Delivery, Other technological contributions to diabetes management
Abstract: This study introduces a personalized Model Predictive Control (MPC) strategy for managing type 1 diabetes, specifically addressing intra-day variability in glucose regulation. The proposed method uses a soft-switching mechanism combining multiple MPCs, each optimized for distinct day periods (breakfast, lunch, and dinner). Patient-specific models were identified using an enhanced impulse response technique, featuring a novel clinically-based cost function to better capture circadian dynamics. A switching mechanism ensures smooth transitions between controllers, minimizing disruptions. Performance was evaluated using the FDA-approved UVA/Padova simulator, demonstrating a 20% improvement in model fit and enhanced control performance compared to the state-of-art approach. The strategy consistently prevented hypoglycemia while improving time-in-target metrics.

Keywords: New control system algorithms for Automatic Insulin Delivery
Abstract: Personalization is key in developing technologies for Type 1 Diabetes (T1D) management, given the large inter- and intra-variability in patients response. This holds in the design of Artificial Pancreas (AP) systems as well. Model Predictive Control (MPC) is one of the most adopted control strategies for this purpose. Leveraging a model of glucose-insulin dynamics, MPC proactively adjusts insulin infusion based on the predicted impact of this action on glucose concentration level. Personalization strategies in MPC could be based on the use of individual-specific models and/or rely on the customization of the cost function. A comparison of these approaches seems to be missing in literature. Therefore, this work investigates three levels of individualization (of cost only, model only, and both cost and model) for a MPC-based AP system. Their comparison is performed using the UVa/Padova T1D Simulator (v.S2014). The latter two strategies are found to significantly outperform the former, provided that the adopted individualized model is compliant with basic physiological requirements.

Keywords: New control system algorithms for Automatic Insulin Delivery, Reducing the computational footprint for embedded Automatic Insulin Delivery systems, Other technological contributions to diabetes management
Abstract: Over the past few years, automated Insulin Delivery (AID) systems have rapidly revolutionized the management of type 1 diabetes (T1D). As the vast majority of first and second generation AID systems are based on mathematical models, it can be difficult to account for the large variability in insulin needs between patients. Therefore, many AID systems are optimized for an average patient, with limited adaptation to individual insulin requirements. We propose a simple wrapper function, POGO, to adjust an AID system to the insulin needs of any user, by rescaling of the insulin inputs and the controller output linearly based on a ratio of insulin needs. We then demonstrate improvement, in silico, in model prediction and glycemic control performances across a wide range of insulin needs while avoiding the need for complex personalization.

Keywords: New control system algorithms for Automatic Insulin Delivery, Other technological contributions to diabetes management
Abstract: Unannounced meals introduce substantial disturbances, causing large deviations in blood glucose concentrations from the desired range. Accurate estimation of meal timing and size is crucial for precise state estimation in a Kalman filter. Achieving accurate meal estimation remains a challenging task for fully-automated insulin delivery systems. This paper proposes incorporating a correction mechanism for the estimated states, where missed meals are detected by a neural network. Additionally, a Bayesian network is utilized to forecast timing probabilities of the next meal. Our proposed stochastic model predictive controller (SMPC) incorporates predicted meal scenarios. We evaluate the controller performance with respect to the stochasticity of the dietary patterns; the results illustrate that integrating the most likely meal scenarios into SMPC decision-making enhances both robustness and performance.

Keywords: New control system algorithms for Automatic Insulin Delivery, Reducing the computational footprint for embedded Automatic Insulin Delivery systems, Other technological contributions to diabetes management
Abstract: Controlling postprandial glucose excursions in type 1 diabetes remains a challenge for most automated insulin delivery systems due to the slow absorption of insulin analogs. This limitation prevents controllers from responding aggressively enough to the rapid glucose spikes following meals to avoid the risk of insulin over-delivery and late hypoglycemia. In this paper, we propose an algorithmic solution to this problem based on the concept of meal anticipation. We propose a data-driven method to predict meal disturbances and anticipate insulin delivery with Model Predictive Control (MPC) frameworks. In contrast to previous works, this strategy does not add significant computational complexity to the original MPC algorithm and, therefore, it is better suited for implementation in wearable devices. Simulations using the University of Virginia Type 1 Diabetes Simulator and its clinically validated MPC algorithm demonstrate an 8% increase in time-in-range along with reductions in hypoglycemia events.

Keywords: New control system algorithms for Automatic Insulin Delivery, Adjuvant therapies to insulin, Physiological modeling
Abstract: Current closed-loop insulin delivery algorithms need to be informed of carbohydrate intake disturbances. This can be a burden on people using these systems. Pramlintide is a hormone that delays gastric emptying, which enables insulin kinetics to align with the kinetics of carbohydrate absorption. Integrating pramlintide into an automated insulin delivery system can be helpful in reducing the postprandial glucose excursion and may be helpful in enabling fully-closed loop whereby meals do not need to be announced. We present an AI-enabled dual-hormone model predictive control (MPC) algorithm that delivers insulin and pramlintide without requiring meal announcements that uses a neural network to automatically detect and deliver meal insulin. The MPC algorithm includes a new pramlintide pharmacokinetics and pharmacodynamics model that was identified using data collected from people with type 1 diabetes undergoing a meal challenge. Using a simulator, we evaluated the performance of various pramlintide delivery methods and controller models, as well as the baseline insulin-only scenario. Meals were automatically dosed using a neural network meal detection and dosing (MDD) algorithm. The primary outcome was the percent time glucose is in the target range (TIR: 70-180 mg/dL). Results show that delivering pramlintide at a fixed ratio of 6 mcg pramlintide:1 u insulin using an MPC that is aware of the pramlintide achieved the most significant improved TIR compared with an insulin-only MPC (91.6% vs. 64.1%). Delivering pramlintide as a fixed ratio was better than delivering basal pramlintide at a constant rate, indicating the benefit of the MDD algorithm. There was no advantage of independent control of insulin and pramlintide compared with insulin and pramlintide delivered as a fixed ratio. Preliminary real-world results from a human subject further indicate the benefit of pramlintide delivery.

Keywords: Reducing the computational footprint for embedded Automatic Insulin Delivery systems, New control system algorithms for Automatic Insulin Delivery, Physiological modeling
Abstract: Neural Networks (NNs) can transform Model Predictive Controllers (MPCs) by serving as efficient and computation-friendly alternatives for control strategies. However, NNs struggle with extrapolation, where performance is unknown when inputs deviate from the training data distribution. To investigate this, a methodology is developed to analyze the relationship between out-of-distribution input states and MPC-trained NN performance. The strategy employs fast nearest-neighbor search algorithms to generate out-of-distribution NN inputs at increasingly far distances from the training set. By generating out-of-distribution elements at controlled distances from the input domain, we quantify NN prediction errors against the ground truth MPC outputs. Results show degradation in NN performance as perturbation distance increases, highlighting the network's sensitivity to extrapolations. In addition, the extrapolation capacity is influenced by the number of learnable parameters in the NN architecture. The larger network tends to have higher accuracy in the testing dataset but exhibits overfitting and more out-of-distribution errors. These findings underscore the need for extra safety layers, such as real-time anomaly detection, to handle inputs beyond the MPC operating space. The development of such techniques is crucial in safety-critical systems such as type-1 diabetes management.

Keywords: Other technological contributions to diabetes management, New control system algorithms for Automatic Insulin Delivery
Abstract: We utilize dynamical modes as features derived from Continuous Glucose Monitoring (CGM) data to detect meal events. By leveraging the inherent properties of underlying dynamics, these modes capture key aspects of glucose variability, enabling the identification of patterns and anomalies associated with meal consumption. This approach not only improves the accuracy of meal detection but also enhances the interpretability of the underlying glucose dynamics. By focusing on dynamical features, our method provides a robust framework for feature extraction, facilitating generalization across diverse datasets and ensuring reliable performance in real-world applications. The proposed technique offers significant advantages over traditional approaches, improving detection accuracy, detection delay, and system robustness.

Keywords: Other technological contributions to diabetes management
Abstract: Insulin pump faults pose significant safety risks for patients with Type 1 Diabetes. Since these malfunctions can severely compromise glucose regulation and endanger patient safety, their reliable and timely detection is of critical importance. In a previous study, we proposed a detection approach leveraging unsupervised anomaly detection algorithms that showed promising performance in detecting complete occlusions. Nevertheless, real-world scenarios often involve partial occlusions, where insulin delivery is only partially obstructed, and the performance of the unsupervised techniques in this set-up remain unexplored. This work addresses this gap by assessing the effectiveness of two algorithms, namely Isolation Forest (IF) and Histogram-Based Outlier Score (HBOS), in detecting partial occlusions. Using the UVA/Padova T1D Simulator, datasets were generated to simulate varying occlusion severity. While a performance degradation in detecting less severe occlusion should be expected, results show that the effectiveness reduction is gradual. Severe occlusions, where only 25% of insulin reaches the patient, were detected with a limited deterioration of the recall, of about 10% less than for complete occlusion, when no insulin is received by the patient. The algorithms were still able to detect about half of the occlusions where half of insulin reached the patient.

Keywords: New control system algorithms for Automatic Insulin Delivery, Other technological contributions to diabetes management, Adjuvant therapies to insulin
Abstract: Several control strategies aimed at developing an artificial pancreas (AP) for patients with type 1 diabetes mellitus exhibit significant limitations. Specifically, the glucose ingested by the patient cannot be directly measured; instead, only its effect (e.g., an increase in blood glucose levels) can be observed. As a result, aggressively compensating for high blood glucose levels following postprandial episodes with high insulin rates does not always ensure blood glucose levels remain within a healthy range and may instead expose patients to potential hypoglycemic conditions.

Some control approaches, whether based on artificial intelligence techniques, control theories, or a combination of both, may exhibit suboptimal performance if they do not incorporate meal intake announcements or meal size information, as suggested by recent studies. o address this limitation, several meal detection algorithms reported in the literature have demonstrated improvements in blood glucose regulation while reducing the occurrence of both hypoglycemic and hyperglycemic events.

This paper, therefore, explores supervised learning-based techniques for meal detection and carbohydrate counting. For meal detection, decision boundaries are presented for each algorithm to offer a clearer understanding of their performance. For carbohydrate counting, regression algorithms are trained and evaluated by comparing their predictions to real measured glucose curves within a postprandial time horizon. The algorithms are validated using the FDA-approved UVA/PADOVA Type 1 Diabetes Simulator

Keywords: Other technological contributions to diabetes management, Physiological modeling
Abstract: Data-driven technologies are becoming widespread in Type 1 Diabetes (T1D) research to deliver personalized care to patients. Missing or erroneous meal records represent a major factor affecting the quality of T1D data and, therefore, the development of these tools. This work proposes an algorithm based on logistic regression to retrospectively detect meals in T1D data. The algorithm is developed using real data from supervised clinical trials. During testing, absolute detection delay was 15 min on average and recall was > 0.80 – results that remained consistent despite of the proportion of missed and erroneous meals. False positive detection increased from 0.33 events/day (when 33% of meals were missing) to 0.66 (when all meals were missing).