Real-time adaptive models for the personalized prediction of glycemic profile in type 1 diabetes patients

Blood Glucose Prediction from Nutrition Analytics in Type 1 Diabetes: A Review

INTRODUCTION

Type 1 Diabetes (T1D) affects over 9 million worldwide and necessitates meticulous self-management for blood glucose (BG) control. Utilizing BG prediction technology allows for increased BG control and a reduction in the diabetes burden caused by self-management requirements. This paper reviews BG prediction models in T1D, which include nutritional components.

METHOD

A systematic search, utilizing the PRISMA guidelines, identified articles focusing on BG prediction algorithms for T1D that incorporate nutritional variables. Eligible studies were screened and analyzed for model type, inclusion of additional aspects in the model, prediction horizon, patient population, inputs, and accuracy.

RESULTS

The study categorizes 138 blood glucose prediction models into data-driven (54%), physiological (14%), and hybrid (33%) types. Prediction horizons of ≤30 min are used in 36% of models, 31-60 min in 34%, 61-90 min in 11%, 91-120 min in 10%, and >120 min in 9%. Neural networks are the most used data-driven technique (47%), and simple carbohydrate intake is commonly included in models (data-driven: 72%, physiological: 52%, hybrid: 67%). Real or free-living data are predominantly used (83%).

CONCLUSION

The primary goal of blood glucose prediction in T1D is to enable informed decisions and maintain safe BG levels, considering the impact of all nutrients for meal planning and clinical relevance.

Multi-step ahead predictive model for blood glucose concentrations of type-1 diabetic patients

Continuous monitoring of blood glucose (BG) levels is a key aspect of diabetes management. Patients with Type-1 diabetes (T1D) require an effective tool to monitor these levels in order to make appropriate decisions regarding insulin administration and food intake to keep BG levels in target range. Effectively and accurately predicting future BG levels at multi-time steps ahead benefits a patient with diabetes by helping them decrease the risks of extremes in BG including hypo- and hyperglycemia. In this study, we present a novel multi-component deep learning model BG-Predict that predicts the BG levels in a multi-step look ahead fashion. The model is evaluated both quantitatively and qualitatively on actual blood glucose data for 97 patients. For the prediction horizon (PH) of 30 mins, the average values for root mean squared error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and normalized mean squared error (NRMSE) are [Formula: see text] mg/dL, 16.77 ± 4.87 mg/dL, [Formula: see text] and [Formula: see text] respectively. When Clarke and Parkes error grid analyses were performed comparing predicted BG with actual BG, the results showed average percentage of points in Zone A of [Formula: see text] and [Formula: see text] respectively. We offer this tool as a mechanism to enhance the predictive capabilities of algorithms for patients with T1D.

GLYFE: review and benchmark of personalized glucose predictive models in type 1 diabetes

Due to the sensitive nature of diabetes-related data, preventing them from being easily shared between studies, and the wide discrepancies in their data processing pipeline, progress in the field of glucose prediction is hard to assess. To address this issue, we introduce GLYFE (GLYcemia Forecasting Evaluation), a benchmark of machine learning-based glucose predictive models. We present the accuracy and clinical acceptability of nine different models coming from the literature, from standard autoregressive to more complex neural network-based models. These results are obtained on two different datasets, namely UVA/Padova Type 1 Diabetes Metabolic Simulator (T1DMS) and Ohio Type-1 Diabetes Mellitus (OhioT1DM), featuring artificial and real type 1 diabetic patients respectively. By providing extensive details about the data flow as well as by providing the whole source code of the benchmarking process, we ensure the reproducibility of the results and the usability of the benchmark by the community. Those results serve as a basis of comparison for future studies. In a field where data are hard to obtain, and where the comparison of results from different studies is often irrelevant, GLYFE gives the opportunity of gathering researchers around a standardized common environment.

Analysis and Design Process for Predicting and Controlling Blood Glucose in Type 1 Diabetic Patients

Engineering smart software that can monitor, predict, and control blood glucose is critical to improving patients' quality of treatments with type 1 Diabetic Mellitus (T1DM). However, ensuring a reasonable glycemic level in diabetic patients is quite challenging, as many methods do not adequately capture the complexities involved in glycemic control. This problem introduces a new level of complexity and uncertainty to the patient's psychological state, thereby making this problem nonlinear and unobservable. In this paper, we formulated a mathematical model using carbohydrate counting, insulin requirements, and the Harris-Benedict energy equations to establish the framework for predicting and controlling blood glucose level regulation in T1DM. We implemented the framework and evaluated its performance using root mean square error (RMSE) and mean absolute error (MAE) on a case study. Our framework had less error rate in terms of RMSE and MAE, which indicates a better fit with reasonable accuracy.

Glucose Prediction under Variable-Length Time-Stamped Daily Events: A Seasonal Stochastic Local Modeling Framework

Accurate glucose prediction along a long-enough time horizon is a key component for technology to improve type 1 diabetes treatment. Subjects with diabetes might benefit from supervision and control systems that accurately predict risks and trigger corrective actions early enough with improved mitigation. However, large intra-patient variability poses big challenges to glucose prediction. In previous works by the authors, clustering and local modeling techniques with seasonal stochastic models proved to be efficient, allowing for good glucose prediction accuracy for long prediction horizons. Continuous glucose monitoring (CGM) data were partitioned into fixed-length postprandial time subseries and clustered with Fuzzy C-Means to collect similar behaviors, enforcing seasonality at each cluster after subseries concatenation. Then, seasonal stochastic models were identified for each cluster and local predictions were integrated into a global prediction. However, free-living conditions do not support the fixed-length partition of CGM data since daily events duration is variable. In this work, a new algorithm is provided to overcome this constraint, allowing better coping with patient's variability under variable-length time-stamped daily events in supervision and control applications. Besides predicted glucose, two real-time indices are additionally provided-a crispness index, indicating good representation of current glucose behavior by a single model, and a normality index, allowing for the detection of an abnormal glucose behavior (unusual according to registered historical data). The framework is tested in a proof-of-concept in silico study with ten patients over four month training data and two independent two month validation datasets, with and without abnormal behaviors, from the distribution version of the UVA/Padova simulator extended with diverse sources of intra-patient variability.

Adversarial multi-source transfer learning in healthcare: Application to glucose prediction for diabetic people

BACKGROUND AND OBJECTIVES

Deep learning has yet to revolutionize general practices in healthcare, despite promising results for some specific tasks. This is partly due to data being in insufficient quantities hurting the training of the models. To address this issue, data from multiple health actors or patients could be combined by capitalizing on their heterogeneity through the use of transfer learning.

METHODS

To improve the quality of the transfer between multiple sources of data, we propose a multi-source adversarial transfer learning framework that enables the learning of a feature representation that is similar across the sources, and thus more general and more easily transferable. We apply this idea to glucose forecasting for diabetic people using a fully convolutional neural network. The evaluation is done by exploring various transfer scenarios with three datasets characterized by their high inter and intra variability.

RESULTS

While transferring knowledge is beneficial in general, we show that the statistical and clinical accuracies can be further improved by using of the adversarial training methodology, surpassing the current state-of-the-art results. In particular, it shines when using data from different datasets, or when there is too little data in an intra-dataset situation. To understand the behavior of the models, we analyze the learnt feature representations and propose a new metric in this regard. Contrary to a standard transfer, the adversarial transfer does not discriminate the patients and datasets, helping the learning of a more general feature representation.

CONCLUSION

The adversarial training framework improves the learning of a general feature representation in a multi-source environment, enhancing the knowledge transfer to an unseen target. The proposed method can help improve the efficiency of data shared by different health actors in the training of deep models.

Ability of Current Machine Learning Algorithms to Predict and Detect Hypoglycemia in Patients With Diabetes Mellitus: Meta-analysis

Background

Machine learning (ML) algorithms have been widely introduced to diabetes research including those for the identification of hypoglycemia.

Objective

The objective of this meta-analysis is to assess the current ability of ML algorithms to detect hypoglycemia (ie, alert to hypoglycemia coinciding with its symptoms) or predict hypoglycemia (ie, alert to hypoglycemia before its symptoms have occurred).

Methods

Electronic literature searches (from January 1, 1950, to September 14, 2020) were conducted using the Dialog platform that covers 96 databases of peer-reviewed literature. Included studies had to train the ML algorithm in order to build a model to detect or predict hypoglycemia and test its performance. The set of 2 × 2 data (ie, number of true positives, false positives, true negatives, and false negatives) was pooled with a hierarchical summary receiver operating characteristic model.

Results

A total of 33 studies (14 studies for detecting hypoglycemia and 19 studies for predicting hypoglycemia) were eligible. For detection of hypoglycemia, pooled estimates (95% CI) of sensitivity, specificity, positive likelihood ratio (PLR), and negative likelihood ratio (NLR) were 0.79 (0.75-0.83), 0.80 (0.64-0.91), 8.05 (4.79-13.51), and 0.18 (0.12-0.27), respectively. For prediction of hypoglycemia, pooled estimates (95% CI) were 0.80 (0.72-0.86) for sensitivity, 0.92 (0.87-0.96) for specificity, 10.42 (5.82-18.65) for PLR, and 0.22 (0.15-0.31) for NLR.

Conclusions

Current ML algorithms have insufficient ability to detect ongoing hypoglycemia and considerate ability to predict impeding hypoglycemia in patients with diabetes mellitus using hypoglycemic drugs with regard to diagnostic tests in accordance with the Users’ Guide to Medical Literature (PLR should be ≥5 and NLR should be ≤0.2 for moderate reliability). However, it should be emphasized that the clinical applicability of these ML algorithms should be evaluated according to patients’ risk profiles such as for hypoglycemia and its associated complications (eg, arrhythmia, neuroglycopenia) as well as the average ability of the ML algorithms. Continued research is required to develop more accurate ML algorithms than those that currently exist and to enhance the feasibility of applying ML in clinical settings.

Trial Registration

PROSPERO International Prospective Register of Systematic Reviews CRD42020163682; http://www.crd.york.ac.uk/PROSPERO/display_record.php?ID=CRD42020163682

Seasonal Local Models for Glucose Prediction in Type 1 Diabetes

Linear empirical dynamic models have been widely used for blood glucose prediction and risks prevention in people with type 1 diabetes. More accurate blood glucose prediction models with longer prediction horizon (PH) are desirable to enable warnings to patients about imminent blood glucose changes with enough time to take corrective actions. In this study, a blood glucose prediction method is developed by integrating the predictions of a set of seasonal local models (each of them corresponding to different glucose profiles observed along historical data). In the modeling step, the number of sets and their corresponding glucose profiles characteristics are obtained by clustering techniques (Fuzzy C-Means). Then, Box-Jenkins methodology is used to identify a seasonal model for each set. Finally, blood glucose predictions of local models are integrated using different techniques. The proposed method is tested by using 18 60-h closed-loop experiments (including different exercise types and artificial pancreas strategies) and achieving mean absolute percentage error (MAPE) of 2.94%, 3.89%, 5.41%, 6.29% and 8.66% for 15-, 30-, 45-, 60-, and 90-min PHs, respectively.

Short-term prediction of glucose in type 1 diabetes using kernel adaptive filters

This study aims at presenting a nonlinear, recursive, multivariate prediction model of the subcutaneous glucose concentration in type 1 diabetes. Nonlinear regression is performed in a reproducing kernel Hilbert space, by either the fixed budget quantized kernel least mean square (QKLMS-FB) or the approximate linear dependency kernel recursive least-squares (KRLS-ALD) algorithm, such that a sparse model structure is accomplished. A multivariate feature set (i.e., subcutaneous glucose, food carbohydrates, insulin regime and physical activity) is used and its influence on short-term glucose prediction is investigated. The method is evaluated using data from 15 patients with type 1 diabetes in free-living conditions. In the case when all the input variables are considered: (i) the average root mean squared error (RMSE) of QKLMS-FB increases from 13.1 mg dL^-1 (mean absolute percentage error (MAPE) 6.6%) for a 15-min prediction horizon (PH) to 37.7 mg dL^-1 (MAPE 20.8%) for a 60-min PH and (ii) the RMSE of KRLS-ALD, being predictably lower, increases from 10.5 mg dL^-1 (MAPE 5.2%) for a 15-min PH to 31.8 mg dL^-1 (MAPE 18.0%) for a 60-min PH. Multivariate data improve systematically both the regularity and the time lag of the predictions, reducing the errors in critical glucose value regions for a PH ≥ 30 min. Graphical abstract ᅟ.

Artificial Intelligence for Diabetes Management and Decision Support: Literature Review

Background

Artificial intelligence methods in combination with the latest technologies, including medical devices, mobile computing, and sensor technologies, have the potential to enable the creation and delivery of better management services to deal with chronic diseases. One of the most lethal and prevalent chronic diseases is diabetes mellitus, which is characterized by dysfunction of glucose homeostasis.

Objective

The objective of this paper is to review recent efforts to use artificial intelligence techniques to assist in the management of diabetes, along with the associated challenges.

Methods

A review of the literature was conducted using PubMed and related bibliographic resources. Analyses of the literature from 2010 to 2018 yielded 1849 pertinent articles, of which we selected 141 for detailed review.

Results

We propose a functional taxonomy for diabetes management and artificial intelligence. Additionally, a detailed analysis of each subject category was performed using related key outcomes. This approach revealed that the experiments and studies reviewed yielded encouraging results.

Conclusions

We obtained evidence of an acceleration of research activity aimed at developing artificial intelligence-powered tools for prediction and prevention of complications associated with diabetes. Our results indicate that artificial intelligence methods are being progressively established as suitable for use in clinical daily practice, as well as for the self-management of diabetes. Consequently, these methods provide powerful tools for improving patients’ quality of life.

Prediction of Adverse Glycemic Events From Continuous Glucose Monitoring Signal

The most important objective of any diabetes therapy is to maintain the blood glucose concentration within the euglycemic range, avoiding or at least mitigating critical hypo/hyperglycemic episodes. Modern continuous glucose monitoring (CGM) devices bear the promise of providing the patients with an increased and timely awareness of glycemic conditions as these get dangerously near to hypo/hyperglycemia. The challenge is to detect, with reasonable advance, the patterns leading to risky situations, allowing the patient to make therapeutic decisions on the basis of future (predicted) glucose concentration levels. We underline that a technically sound performance comparison of the approaches proposed in recent years has yet to be done, thus it is unclear which one is preferred. The aim of this study is to fill this gap by carrying out a comparative analysis among the most common methods for glucose event prediction. Both regression and classification algorithms have been implemented and analyzed, including static and dynamic training approaches. The dataset consists of 89 CGM time series measured in diabetic subjects for 7 subsequent days. Performance metrics, specifically defined to assess and compare the event-prediction capabilities of the methods, have been introduced and analyzed. Our numerical results show that a static training approach exhibits better performance, in particular when regression methods are considered. However, classifiers show some improvement when trained for a specific event category, such as hyperglycemia, achieving performance comparable to the regressors, with the advantage of predicting the events sooner.

Kernel-based adaptive learning improves accuracy of glucose predictive modelling in type 1 diabetes: A proof-of-concept study

This study aims at demonstrating the need for nonlinear recursive models to the identification and prediction of the dynamic glucose system in type 1 diabetes. Nonlinear regression is performed in a reproducing kernel Hilbert space, by the Approximate Linear Dependency Kernel Recursive Least Squares (KRLS-ALD) algorithm, such that a sparse model structure is accomplished. The method is evaluated on seven people with type 1 diabetes in free-living conditions, where a change in glycaemic dynamics is forced by increasing the level of physical activity in the middle of the observational period. The univariate input allows for short-term (≤30 min) predictions with KRLS-ALD reaching an average root mean square error of 15.22±5.95 mgdL^-1 and an average time lag of 17.14±2.67 min for an horizon of 30 min. Its performance is considerably better than that of time-invariant (regularized) linear regression models.

Non-linear dynamic modeling of glucose in type 1 diabetes with kernel adaptive filters

We propose a non-linear recursive solution to the problem of short-term prediction of glucose in type 1 diabetes. The Fixed Budget Quantized Kernel Least Mean Square (QKLMS-FB) algorithm is employed to construct a univariate model of subcutaneous glucose concentration, which: (i) handles nonlinearities by transforming the input space into a high-dimensional Reproducing Kernel Hilbert Space and, (ii) finds a sparse solution by retaining a representative subset of the training input vectors. The dataset comes from the continuous multi-day recordings of 15 type 1 patients in free-living conditions. QKLMS-FB produces an average root mean squared error of 18.66±3.19 mg/dl for a prediction horizon of 30 min with 82.04% of hypoglycemic readings and 93.30% of hyperglycemic ones being classified as clinically accurate or with benign errors. The effect of the prediction horizon is more evident in the hypoglycemic range.

How Much Is Short-Term Glucose Prediction in Type 1 Diabetes Improved by Adding Insulin Delivery and Meal Content Information to CGM Data? A Proof-of-Concept Study

BACKGROUND

In type 1 diabetes (T1D) management, short-term glucose prediction can allow to anticipate therapeutic decisions when hypo/hyperglycemia is imminent. Literature prediction methods mainly use past continuous glucose monitoring (CGM) readings. Sophisticated algorithms can use information on insulin delivered and meal carbohydrate (CHO) content. The quantification of how much insulin and CHO information improves glucose prediction is missing in the literature and is investigated, in an open-loop setting, in this proof-of-concept study.

METHODS

We adopted a versatile literature prediction methodology able to utilize a variety of inputs. We compared predictors that use (1) CGM; (2) CGM and insulin; (3) CGM and CHO; and (4) CGM, insulin, and CHO. Data of 15 T1D subjects in open-loop setup were used. Prediction was evaluated via absolute error and temporal gain focusing on meal/night periods. The relative importance of each individual input of the predictor was evaluated with a sensitivity analysis.

RESULTS

For a prediction horizon (PH) ≥ 30 minutes, insulin and CHO information improves prediction accuracy of 10% and double the temporal gain during the 2 hours following the meal. During the night the 4 methods did not give statistically different results. When PH ≥ 45 minutes, the influence of CHO information on prediction is 5-fold that of insulin.

CONCLUSIONS

In an open-loop setting, with PH ≥ 30 minutes, information on CHO and insulin improves short-term glucose prediction in the 2-hour time window following a meal, but not during the night. CHO information improves prediction significantly more than insulin.

Model-Free Machine Learning in Biomedicine: Feasibility Study in Type 1 Diabetes

Although reinforcement learning (RL) is suitable for highly uncertain systems, the applicability of this class of algorithms to medical treatment may be limited by the patient variability which dictates individualised tuning for their usually multiple algorithmic parameters. This study explores the feasibility of RL in the framework of artificial pancreas development for type 1 diabetes (T1D). In this approach, an Actor-Critic (AC) learning algorithm is designed and developed for the optimisation of insulin infusion for personalised glucose regulation. AC optimises the daily basal insulin rate and insulin:carbohydrate ratio for each patient, on the basis of his/her measured glucose profile. Automatic, personalised tuning of AC is based on the estimation of information transfer (IT) from insulin to glucose signals. Insulin-to-glucose IT is linked to patient-specific characteristics related to total daily insulin needs and insulin sensitivity (SI). The AC algorithm is evaluated using an FDA-accepted T1D simulator on a large patient database under a complex meal protocol, meal uncertainty and diurnal SI variation. The results showed that 95.66% of time was spent in normoglycaemia in the presence of meal uncertainty and 93.02% when meal uncertainty and SI variation were simultaneously considered. The time spent in hypoglycaemia was 0.27% in both cases. The novel tuning method reduced the risk of severe hypoglycaemia, especially in patients with low SI.

Multi-model data fusion to improve an early warning system for hypo-/hyperglycemic events

Correct predictions of future blood glucose levels in individuals with Type 1 Diabetes (T1D) can be used to provide early warning of upcoming hypo-/hyperglycemic events and thus to improve the patient's safety. To increase prediction accuracy and efficiency, various approaches have been proposed which combine multiple predictors to produce superior results compared to single predictors. Three methods for model fusion are presented and comparatively assessed. Data from 23 T1D subjects under sensor-augmented pump (SAP) therapy were used in two adaptive data-driven models (an autoregressive model with output correction - cARX, and a recurrent neural network - RNN). Data fusion techniques based on i) Dempster-Shafer Evidential Theory (DST), ii) Genetic Algorithms (GA), and iii) Genetic Programming (GP) were used to merge the complimentary performances of the prediction models. The fused output is used in a warning algorithm to issue alarms of upcoming hypo-/hyperglycemic events. The fusion schemes showed improved performance with lower root mean square errors, lower time lags, and higher correlation. In the warning algorithm, median daily false alarms (DFA) of 0.25%, and 100% correct alarms (CA) were obtained for both event types. The detection times (DT) before occurrence of events were 13.0 and 12.1 min respectively for hypo-/hyperglycemic events. Compared to the cARX and RNN models, and a linear fusion of the two, the proposed fusion schemes represents a significant improvement.

Personalized tuning of a reinforcement learning control algorithm for glucose regulation

Artificial pancreas is in the forefront of research towards the automatic insulin infusion for patients with type 1 diabetes. Due to the high inter- and intra-variability of the diabetic population, the need for personalized approaches has been raised. This study presents an adaptive, patient-specific control strategy for glucose regulation based on reinforcement learning and more specifically on the Actor-Critic (AC) learning approach. The control algorithm provides daily updates of the basal rate and insulin-to-carbohydrate (IC) ratio in order to optimize glucose regulation. A method for the automatic and personalized initialization of the control algorithm is designed based on the estimation of the transfer entropy (TE) between insulin and glucose signals. The algorithm has been evaluated in silico in adults, adolescents and children for 10 days. Three scenarios of initialization to i) zero values, ii) random values and iii) TE-based values have been comparatively assessed. The results have shown that when the TE-based initialization is used, the algorithm achieves faster learning with 98%, 90% and 73% in the A+B zones of the Control Variability Grid Analysis for adults, adolescents and children respectively after five days compared to 95%, 78%, 41% for random initialization and 93%, 88%, 41% for zero initial values. Furthermore, in the case of children, the daily Low Blood Glucose Index reduces much faster when the TE-based tuning is applied. The results imply that automatic and personalized tuning based on TE reduces the learning period and improves the overall performance of the AC algorithm.

GoCARB in the Context of an Artificial Pancreas

BACKGROUND

In an artificial pancreas (AP), the meals are either manually announced or detected and their size estimated from the blood glucose level. Both methods have limitations, which result in suboptimal postprandial glucose control. The GoCARB system is designed to provide the carbohydrate content of meals and is presented within the AP framework.

METHOD

The combined use of GoCARB with a control algorithm is assessed in a series of 12 computer simulations. The simulations are defined according to the type of the control (open or closed loop), the use or not-use of GoCARB and the diabetics' skills in carbohydrate estimation.

RESULTS

For bad estimators without GoCARB, the percentage of the time spent in target range (70-180 mg/dl) during the postprandial period is 22.5% and 66.2% for open and closed loop, respectively. When the GoCARB is used, the corresponding percentages are 99.7% and 99.8%. In case of open loop, the time spent in severe hypoglycemic events (<50 mg/dl) is 33.6% without the GoCARB and is reduced to 0.0% when the GoCARB is used. In case of closed loop, the corresponding percentage is 1.4% without the GoCARB and is reduced to 0.0% with the GoCARB.

CONCLUSION

The use of GoCARB improves the control of postprandial response and glucose profiles especially in the case of open loop. However, the most efficient regulation is achieved by the combined use of the control algorithm and the GoCARB.

A therapy parameter-based model for predicting blood glucose concentrations in patients with type 1 diabetes

In this paper, the problem of predicting blood glucose concentrations (BG) for the treatment of patients with type 1 diabetes, is addressed. Predicting BG is of very high importance as most treatments, which consist in exogenous insulin injections, rely on the availability of BG predictions. Many models that can be used for predicting BG are available in the literature. However, it is widely admitted that it is almost impossible to perfectly model blood glucose dynamics while still being able to identify model parameters using only blood glucose measurements. The main contribution of this work is to propose a simple and identifiable linear dynamical model, which is based on the static prediction model of standard therapy. It is shown that the model parameters are intrinsically correlated with physician-set therapy parameters and that the reduction of the number of model parameters to identify leads to inferior data fits but to equivalent or slightly improved prediction capabilities compared to state-of-the-art models: a sign of an appropriate model structure and superior reliability. The validation of the proposed dynamic model is performed using data from the UVa simulator and real clinical data, and potential uses of the proposed model for state estimation and BG control are discussed.

Jump neural network for real-time prediction of glucose concentration

Prediction of the future value of a variable is of central importance in a wide variety of fields, including economy and finance, meteorology, informatics, and, last but not least important, medicine. For example, in the therapy of Type 1 Diabetes (T1D), in which, for patient safety, glucose concentration in the blood should be maintained in a defined normoglycemic range, the ability to forecast glucose concentration in the short-term (with a prediction horizon of around 30 min) might be sufficient to reduce the incidence of hypoglycemic and hyperglycemic events. Neural Network (NN) approaches are suitable for prediction purposes because of their ability to model nonlinear dynamics and handle in their inputs signals coming from different domains. In this chapter we illustrate the design of a jump NN glucose prediction algorithm that exploits past glucose concentration data, measured in real-time by a minimally invasive continuous glucose monitoring (CGM) sensor, and information on ingested carbohydrates, supplied by the patient himself or herself. The methodology is assessed by tuning the NN on data of ten T1D individuals and then testing it on a dataset of ten different subjects. Results with a prediction horizon of 30 min show that prediction of glucose concentration in T1D via NN is feasible and sufficiently accurate. The average time anticipation obtained is compatible with the generation of preventive hypoglycemic and hyperglycemic alerts and the improvement of artificial pancreas performance.

An early warning system for hypoglycemic/hyperglycemic events based on fusion of adaptive prediction models

INTRODUCTION

Early warning of future hypoglycemic and hyperglycemic events can improve the safety of type 1 diabetes mellitus (T1DM) patients. The aim of this study is to design and evaluate a hypoglycemia/hyperglycemia early warning system (EWS) for T1DM patients under sensor-augmented pump (SAP) therapy.

METHODS

The EWS is based on the combination of data-driven online adaptive prediction models and a warning algorithm. Three modeling approaches have been investigated: (i) autoregressive (ARX) models, (ii) auto-regressive with an output correction module (cARX) models, and (iii) recurrent neural network (RNN) models. The warning algorithm performs postprocessing of the models' outputs and issues alerts if upcoming hypoglycemic/hyperglycemic events are detected. Fusion of the cARX and RNN models, due to their complementary prediction performances, resulted in the hybrid autoregressive with an output correction module/recurrent neural network (cARN)-based EWS.

RESULTS

The EWS was evaluated on 23 T1DM patients under SAP therapy. The ARX-based system achieved hypoglycemic (hyperglycemic) event prediction with median values of accuracy of 100.0% (100.0%), detection time of 10.0 (8.0) min, and daily false alarms of 0.7 (0.5). The respective values for the cARX-based system were 100.0% (100.0%), 17.5 (14.8) min, and 1.5 (1.3) and, for the RNN-based system, were 100.0% (92.0%), 8.4 (7.0) min, and 0.1 (0.2). The hybrid cARN-based EWS presented outperforming results with 100.0% (100.0%) prediction accuracy, detection 16.7 (14.7) min in advance, and 0.8 (0.8) daily false alarms.

CONCLUSION

Combined use of cARX and RNN models for the development of an EWS outperformed the single use of each model, achieving accurate and prompt event prediction with few false alarms, thus providing increased safety and comfort.