A neural network approach for insulin regime and dose adjustment in type 1 diabetes

LSTMs and Deep Residual Networks for Carbohydrate and Bolus Recommendations in Type 1 Diabetes Management

To avoid serious diabetic complications, people with type 1 diabetes must keep their blood glucose levels (BGLs) as close to normal as possible. Insulin dosages and carbohydrate consumption are important considerations in managing BGLs. Since the 1960s, models have been developed to forecast blood glucose levels based on the history of BGLs, insulin dosages, carbohydrate intake, and other physiological and lifestyle factors. Such predictions can be used to alert people of impending unsafe BGLs or to control insulin flow in an artificial pancreas. In past work, we have introduced an LSTM-based approach to blood glucose level prediction aimed at “what-if” scenarios, in which people could enter foods they might eat or insulin amounts they might take and then see the effect on future BGLs. In this work, we invert the “what-if” scenario and introduce a similar architecture based on chaining two LSTMs that can be trained to make either insulin or carbohydrate recommendations aimed at reaching a desired BG level in the future. Leveraging a recent state-of-the-art model for time series forecasting, we then derive a novel architecture for the same recommendation task, in which the two LSTM chain is used as a repeating block inside a deep residual architecture. Experimental evaluations using real patient data from the OhioT1DM dataset show that the new integrated architecture compares favorably with the previous LSTM-based approach, substantially outperforming the baselines. The promising results suggest that this novel approach could potentially be of practical use to people with type 1 diabetes for self-management of BGLs.

Clinically Accurate Prediction of Glucose Levels in Patients with Type 1 Diabetes

Background and Aims: Accurate prediction of glucose levels in patients with type 1 diabetes mellitus (T1DM) is critical both for their glycemic control and for the development of closed-loop systems. Methods: In this study, we utilized real-life, retrospective, continuous glucose monitoring data from 141 T1DM patients (9,083 connection days, 1,592,506 glucose measurements) and in silico data generated by the UVA/Padova T1DM simulator to evaluate different computational methods for glucose prediction. We evaluated the performance of the models using both measures of numerical accuracy, measured by the root mean square error, and clinical accuracy, measured by the percentage of time in each of the Clarke error grid (CEG) zones, and compared the predictions done by autoregressive (AR) models, tree-based methods, artificial neural networks, and a novel model that we devised and optimized for this task. Results: Our novel model, constructed on real-life data, achieved clinical accuracy of 99.3% and 95.8% in predicting the glucose level 30 and 60 min ahead, respectively, and reduced the percentage of glucose predictions in zones C-E of the CEG by 60.6% and 38.4% in these prediction horizons, compared with a standard AR model. The model was superior to all other models across all age groups and achieved higher clinical accuracy in subgroups of patients with high glucose variability and greater time spent in hypoglycemia. Compared with real-life data, when evaluated on in silico data, the model had a higher clinical and numerical accuracy. Conclusions: A model that optimizes for CEG zones may significantly improve clinical accuracy and clinical outcomes of treatment decisions in T1DM patients. Results obtained from simulated data may overestimate the performance of models on real-life data.

Adaptive System Identification for Estimating Future Glucose Concentrations and Hypoglycemia Alarms

Many patients with diabetes experience high variability in glucose concentrations that includes prolonged hyperglycemia or hypoglycemia. Models predicting a subject's future glucose concentrations can be used for preventing such conditions by providing early alarms. This paper presents a time-series model that captures dynamical changes in the glucose metabolism. Adaptive system identification is proposed to estimate model parameters which enable the adaptation of the model to inter-/intra-subject variation and glycemic disturbances. It consists of online parameter identification using the weighted recursive least squares method and a change detection strategy that monitors variation in model parameters. Univariate models developed from a subject's continuous glucose measurements are compared to multivariate models that are enhanced with continuous metabolic, physical activity and lifestyle information from a multi-sensor body monitor. A real life application for the proposed algorithm is demonstrated on early (30 min in advance) hypoglycemia detection.

Insulin dosage optimization based on prediction of postprandial glucose excursions under uncertain parameters and food intake

Considering the difficulty in selecting correct insulin doses and the problem of hyper- and hypoglycemia episodes in type 1 diabetes, dosage-aid systems are very useful for these patients. A model-based approach to this problem must unavoidably consider uncertainty sources such as large intra-patient variability and food intake. In the present study, postprandial glucose is predicted considering this uncertain information using modal interval analysis. This approach calculates a safer prediction of possible hyper- and hypoglycemia episodes induced by insulin therapy for an individual patient's parameters and integrates this information into a dosage-aid system. Predictions of a patient's postprandial glucose at 5-h intervals are used to predict the risk for a given therapy. Then the insulin dose and injection-to-meal time with the lowest risk are calculated. The method has been validated for three different scenarios corresponding to preprandial glucose values of 100, 180 and 250mg/dl.

Mobile phone technologies and advanced data analysis towards the enhancement of diabetes self-management

Advances in the area of mobile and wireless communication for healthcare (m-Health) along with the improvements in information science allow the design and development of new patient-centric models for the provision of personalised healthcare services, increase of patient independence and improvement of patient's self-control and self-management capabilities. This paper comprises a brief overview of the m-Health applications towards the self-management of individuals with diabetes mellitus and the enhancement of their quality of life. Furthermore, the design and development of a mobile phone application for Type 1 Diabetes Mellitus (T1DM) self-management is presented. The technical evaluation of the application, which permits the management of blood glucose measurements, blood pressure measurements, insulin dosage, food/drink intake and physical activity, has shown that the use of the mobile phone technologies along with data analysis methods might improve the self-management of T1DM.

Blood Glucose Prediction Using Artificial Neural Networks Trained with the AIDA Diabetes Simulator: A Proof-of-Concept Pilot Study

Diabetes mellitus is a major, and increasing, global problem. However, it has been shown that, through good management of blood glucose levels (BGLs), the associated and costly complications can be reduced significantly. In this pilot study, Elman recurrent artificial neural networks (ANNs) were used to make BGL predictions based on a history of BGLs, meal intake, and insulin injections. Twenty-eight datasets (from a single case scenario) were compiled from the freeware mathematical diabetes simulator, AIDA. It was found that the most accurate predictions were made during the nocturnal period of the 24 hour daily cycle. The accuracy of the nocturnal predictions was measured as the root mean square error over five test days (RMSE5 day) not used during ANN training. For BGL predictions of up to 1 hour aRMSE5 dayof (±SD)0.15±0.04 mmol/L was observed. For BGL predictions up to 10 hours, aRMSE5  dayof (±SD)0.14±0.16 mmol/L was observed. Future research will investigate a wider range of AIDA case scenarios, real-patient data, and data relating to other factors influencing BGLs. ANN paradigms based on real-time recurrent learning will also be explored to accommodate dynamic physiology in diabetes.

Artificial neural network algorithm for online glucose prediction from continuous glucose monitoring

BACKGROUND AND AIMS

Continuous glucose monitoring (CGM) devices could be useful for real-time management of diabetes therapy. In particular, CGM information could be used in real time to predict future glucose levels in order to prevent hypo-/hyperglycemic events. This article proposes a new online method for predicting future glucose concentration levels from CGM data.

METHODS

The predictor is implemented with an artificial neural network model (NNM). The inputs of the NNM are the values provided by the CGM sensor during the preceding 20 min, while the output is the prediction of glucose concentration at the chosen prediction horizon (PH) time. The method performance is assessed using datasets from two different CGM systems (nine subjects using the Medtronic [Northridge, CA] Guardian and six subjects using the Abbott [Abbott Park, IL] Navigator. Three different PHs are used: 15, 30, and 45 min. The NNM accuracy has been estimated by using the root mean square error (RMSE) and prediction delay.

RESULTS

The RMSE is around 10, 18, and 27 mg/dL for 15, 30, and 45 min of PH, respectively. The prediction delay is around 4, 9, and 14 min for upward trends and 5, 15, and 26 min for downward trends, respectively. A comparison with a previously published technique, based on an autoregressive model (ARM), has been performed. The comparison shows that the proposed NNM is more accurate than the ARM, with no significant deterioration in the prediction delay.

CONCLUSIONS

The proposed NNM is a reliable solution for the online prediction of future glucose concentrations from CGM data.

A real time simulation model of glucose-insulin metabolism for type 1 diabetes patients

In this paper, a simulation model of glucose-insulin metabolism for Type 1 diabetes patients is presented. The proposed system is based on the combination of Compartmental Models (CMs) and artificial Neural Networks (NNs). This model aims at the development of an accurate system, in order to assist Type 1 diabetes patients to handle their blood glucose profile and recognize dangerous metabolic states. Data from a Type 1 diabetes patient, stored in a database, have been used as input to the hybrid system. The data contain information about measured blood glucose levels, insulin intake, and description of food intake, along with the corresponding time. The data are passed to three separate CMs, which produce estimations about (i) the effect of Short Acting (SA) insulin intake on blood insulin concentration, (ii) the effect of Intermediate Acting (IA) insulin intake on blood insulin concentration, and (iii) the effect of carbohydrate intake on blood glucose absorption from the gut. The outputs of the three CMs are passed to a Recurrent NN (RNN) in order to predict subsequent blood glucose levels. The RNN is trained with the Real Time Recurrent Learning (RTRL) algorithm. The resulted blood glucose predictions are promising for the use of the proposed model for blood glucose level estimation for Type 1 diabetes patients.

A preliminary evaluation of neural network analysis for pharmacodynamic modeling of the dosing of the hydroxymethylglutaryl coenzyme A-reductase inhibitors simvastatin and atorvastatin

BACKGROUND

Neural networks have been used in diagnosing and treating many diseases, including the diagnosis of myocardial infarction and insulin dosing in diabetes mellitus.

OBJECTIVE

The goal of this study was to develop a preliminary pharmacodynamic model for dosing of the hydroxymethylglutaryl coenzyme A (HMG-CoA)-reductase inhibitors simvastatin and atorvastatin using neural network analysis.

METHODS

Using NeuralSIM neural network software (NeuralWare, Carnegie, Pa), lipid panels from patients at a hospital lipid clinic were entered as inputs for the model, and the doses of simvastatin or atorvastatin that achieved those lipid results were used as outputs. The network was trained using 2 different data sets and was run using a subset of the data as inputs. The results of the neural network run (predictions) were compared with the actual doses used as directed by the hospital pharmacist in accordance with the lipid clinic protocol.

RESULTS

Complete data sets were available for 17 patients (11 men, 6 women). The mean age of these patients was 56.7 years (range, 40-67 years). The neural network model based on data set 1 predicted a dose that was within 95% of the actual dose 7 of 12 times and predicted use of the drug actually used 13 of 19 (68.4%) times. The neural network based on data set 2 predicted use of the drug actually used 10 of 17 (58.8%) times, but the predicted doses never approximated the actual doses by > or = 90%.

CONCLUSIONS

A neural network model for the dosing of the HMG-CoA-reductase inhibitors simvastatin and atorvastatin demonstrated an ability to predict appropriate dosing, but inclusion of other factors (eg, age, body weight, sex) and a larger sample size may be necessary for development of a more accurate model.