Objective: Artificial intelligence and machine learning are transforming many fields including medicine. In diabetes, robust biosensing technologies and automated insulin delivery therapies have created a substantial opportunity to improve health. While the number of manuscripts addressing the topic of applying machine learning to diabetes has grown in recent years, there has been a lack of consistency in the methods, metrics, and data used to train and evaluate these algorithms. This manuscript provides consensus guidelines for machine learning practitioners in the field of diabetes, including best practice recommended approaches and warnings about pitfalls to avoid. Methods: Algorithmic approaches are reviewed and benefits of different algorithms are discussed including importance of clinical accuracy, explainability, interpretability, and personalization. We review the most common features used in machine learning applications in diabetes glucose control and provide an open-source library of functions for calculating features, as well as a framework for specifying data sets using data sheets. A review of current data sets available for training algorithms is provided as well as an online repository of data sources. Significance: These consensus guidelines are designed to improve performance and translatability of new machine learning algorithms developed in the field of diabetes for engineers and data scientists.

Evaluation of metabolic control of diabetic people has been classically performed measuring glucose concentrations in blood samples. Due to the potential improvement it offers in diabetes care, continuous glucose monitoring (CGM) in the subcutaneous tissue is gaining popularity among both patients and physicians. However, devices for CGM measure glucose concentration in compartments other than blood, usually the interstitial space. This means that CGM need calibration against blood glucose values, and the accuracy of the estimation of blood glucose will also depend on the calibration algorithm. The complexity of the relationship between glucose dynamics in blood and the interstitial space, contrasts with the simplistic approach of calibration algorithms currently implemented in commercial CGM devices, translating in suboptimal accuracy. The present review will analyze the issue of calibration algorithms for CGM, focusing exclusively on the commercially available glucose sensors.

The artificial pancreas (AP) system is designed to regulate blood glucose in subjects with type 1 diabetes using a continuous glucose monitor informed controller that adjusts insulin infusion via an insulin pump. However, current AP developments are mainly hybrid closed-loop systems that include feed-forward actions triggered by the announcement of meals or exercise. The first step to fully closing the loop in the AP requires removing meal announcement, which is currently the most effective way to alleviate postprandial hyperglycemia due to the delay in insulin action. Here, a novel approach to meal detection in the AP is presented using a sliding window and computing the normalized cross-covariance between measured glucose and the forward difference of a disturbance term, estimated from an augmented minimal model using an Unscented Kalman Filter. Three different tunings were applied to the same meal detection algorithm: (1) a high sensitivity tuning, (2) a trade-off tuning that has a high amount of meals detected and a low amount of false positives (FP), and (3) a low FP tuning. For the three tunings sensitivities 99 ± 2%, 93 ± 5%, and 47 ± 12% were achieved, respectively. A sensitivity analysis was also performed and found that higher carbohydrate quantities and faster rates of glucose appearance result in favorable meal detection outcomes.

The artificial pancreas aims at the automatic delivery of insulin for glycemic control in patients with type 1 diabetes, i.e., closed-loop glucose control. One of the challenges of the artificial pancreas is to avoid controller overreaction leading to hypoglycemia, especially in the late postprandial period. In this study, an original proposal based on sliding mode reference conditioning ideas is presented as a way to reduce hypoglycemia events induced by a closed-loop glucose controller. The method is inspired in the intuitive advantages of two-step constrained control algorithms. It acts on the glucose reference sent to the main controller shaping it so as to avoid violating given constraints on the insulin-on-board. Some distinctive features of the proposed strategy are that 1) it provides a safety layer which can be adjusted according to medical criteria; 2) it can be added to closed-loop controllers of any nature; 3) it is robust against sensor failures and overestimated prandial insulin doses; and 4) it can handle nonlinear models. The method is evaluated in silico with the ten adult patients available in the FDA-accepted UVA simulator.

Continuous Glucose Monitoring (CGM) has been a springboard of new diabetes management technologies such as integrated sensor-pump systems, the artificial pancreas, and more recently, smart pens. It also allows patients to make better informed decisions compared to a few measurements per day from a glucometer. However, CGM accuracy is reportedly affected during exercise periods, which can impact the effectiveness of CGM-based treatments. In this review, several studies that used CGM during exercise periods are scrutinized. An extensive literature review of clinical trials including exercise and CGM in type 1 diabetes was conducted. The gathered data were critically analysed, especially the Mean Absolute Relative Difference (MARD), as the main metric of glucose accuracy. Most papers did not provide accuracy metrics that differentiated between exercise and rest (non-exercise) periods, which hindered comparative data analysis. Nevertheless, the statistic results confirmed that CGM during exercise periods is less accurate.

The artificial pancreas (AP) system or closed-loop blood glucose (BG) regulation is a technological advancement that aims to relieve diabetic subjects from their current decision-making burden while tightening their BG levels. However, large disturbances such as meals and exercise still pose great challenges to a fully closed-loop system. In this paper, the problem of BG regulation with unannounced physical activity for type 1 diabetic subjects is addressed. We use a coordinated control strategy with insulin infusion and extra carbohydrates (CHO) for hypoglycemia avoidance. The insulin algorithm is based on a proportional-derivative controller with insulin feedback and the so-called safety auxiliary feedback element (SAFE) layer, and the algorithm for CHO is based on a predictive, quantified proportional-derivative controller. The UVa/Padua simulator glucose-insulin model is modified to include the effects of physical activity and is used to test the new AP. We consider scenarios where the subject does not announce physical activity and with challenging meals. Then, we analyze the performance and robustness of the combined insulin and CHO recommender system and compare them to the insulin-only controller. The simulation results show that the new AP system is able to mitigate daily hypoglycemia episodes (0.9 versus 0.2, p < 0.01) and increase the time in range during day (91.5% versus 92.4%, p < 0.01) without increasing the time above 180 mg/dl (6.3% versus 6.4%, p > 0.05).