Linear Model Identification for Personalized Prediction and Control in Diabetes

Autoencoder-Based Detection of Insulin Pump Faults in Type 1 Diabetes Treatment

Individuals with type 1 diabetes (T1D) require lifelong insulin replacement to compensate for deficient endogenous insulin secretion, which would otherwise result in abnormal blood glucose levels. In recent years, significant investments have been made to improve T1D management, leading to the widespread adoption of accurate technology such as continuous glucose monitoring (CGM) sensors and automated insulin delivery systems. However, malfunctions in these devices, particularly pump systems, can cause undesirable interruptions of insulin delivery posing significant safety risks if not promptly addressed. Due to the low frequency of these episodes, developing accurate algorithms to identify insulin pump faults remains a challenge. To address these issues, this paper proposes a novel approach for detecting insulin pump faults (IPFs) by combining the ability of a long short-term memory (LSTM) autoencoder to extract features, with the strength of random forest to distinguish between anomalous and normal patterns. This method was developed and evaluated using data from 100 subjects, simulated over 90 days with the UVa/Padova T1D Simulator, an FDA-approved nonlinear computer simulator of T1D physiology. In the test set, the proposed algorithm identified the 93% of the total faults, while raising 2 false alarms in 3 months on average. These findings suggest that deep learning algorithms can enhance the safety and reliability of insulin pump systems, contributing to more effective therapeutic technologies.

Shortcomings in the Evaluation of Blood Glucose Forecasting

OBJECTIVE

Recent years have seen an increase in machine learning (ML)-based blood glucose (BG) forecasting models, with a growing emphasis on potential application to hybrid or closed-loop predictive glucose controllers. However, current approaches focus on evaluating the accuracy of these models using benchmark data generated under the behavior policy, which may differ significantly from the data the model may encounter in a control setting. This study challenges the efficacy of such evaluation approaches, demonstrating that they can fail to accurately capture an ML-based model's true performance in closed-loop control settings.

METHODS

Forecast error measured using current evaluation approaches was compared to the control performance of two forecasters - a ML-based model (LSTM) and a rule-based model (Loop) - in silico when the forecasters were utilized with a model-based controller in a hybrid closed-loop setting.

RESULTS

Under current evaluation standards, LSTM achieves a significantly lower (better) forecast error than Loop with a root mean squared error (RMSE) of vs. at the 30-minute prediction horizon. Yet in a control setting, LSTM led to significantly worse control performance with only 77.14% (IQR 66.57-84.03) time-in-range compared to 86.20% (IQR 78.28-91.21) for Loop.

CONCLUSION

Prevailing evaluation methods can fail to accurately capture the forecaster's performance when utilized in closed-loop settings.

SIGNIFICANCE

Our findings underscore the limitations of current evaluation standards and the need for alternative evaluation metrics and training strategies when developing BG forecasters for closed-loop control systems.

Towards Personalized AI-Based Diabetes Therapy: A Review

Insulin pumps and other smart devices have recently made significant advancements in the treatment of diabetes, a disorder that affects people all over the world. The development of medical AI has been influenced by AI methods designed to help physicians make diagnoses, choose a course of therapy, and predict outcomes. In this article, we thoroughly analyse how AI is being used to enhance and personalize diabetes treatment. The search turned up 77 original research papers, from which we've selected the most crucial information regarding the learning models employed, the data typology, the deployment stage, and the application domains. We identified two key trends, enabled mostly by AI: patient-based therapy personalization and therapeutic algorithm optimization. In the meanwhile, we point out various shortcomings in the existing literature, like a lack of multimodal database analysis or a lack of interpretability. The rapid improvements in AI and the expansion of the amount of data already available offer the possibility to overcome these difficulties shortly and enable a wider deployment of this technology in clinical settings.

AGATA: A Toolbox for Automated Glucose Data Analysis

BACKGROUND

Analyzing continuous glucose monitoring (CGM) data is a mandatory step for multiple purposes spanning from reporting clinical trial outcomes to developing new algorithms for diabetes management. This task is repetitive, and scientists struggle in computing literature glucose control metrics and waste time in reproducing possibly complex plots and reports. For this reason, to provide the diabetes technology community a unified tool, here we present Automated Glucose dATa Analysis (AGATA), an automated glucose data analysis toolbox developed in MATLAB/Octave.

METHODS

Automated Glucose dATa Analysis is an open-source software program to visualize and preprocess CGM data, compute glucose control metrics, detect adverse events, evaluate the effectiveness of users' prediction algorithms, and compare study arms. Automated Glucose dATa Analysis can be used as a standalone computer application accessible through a dedicated graphical user interface, particularly suitable for clinicians, or by integrating its functionalities in user-defined MATLAB/Octave scripts, which fits the need of researchers and developers. To demonstrate its features, we used AGATA to analyze CGM data of two subjects extracted from a publicly available data set of individuals with type one diabetes. Finally, AGATA's features are compared against those of 12 noncommercial software programs for CGM data analysis.

RESULTS

Using AGATA, we easily preprocessed, analyzed, and visualized CGM data in a handy way, in compliance with the requirements and the standards defined in the literature. Compared to the other considered software programs, AGATA offers more functionalities and capabilities.

CONCLUSION

Automated Glucose dATa Analysis is easy to use and reduces the burden of CGM data analysis. It is freely available in GitHub at https://github.com/gcappon/agata.

Individualized Models for Glucose Prediction in Type 1 Diabetes: Comparing Black-Box Approaches to a Physiological White-Box One

OBJECTIVE

Accurate blood glucose (BG) prediction are key in next-generation tools for type 1 diabetes (T1D) management, such as improved decision support systems and advanced closed-loop control. Glucose prediction algorithms commonly rely on black-box models. Large physiological models, successfully adopted for simulation, were little explored for glucose prediction, mostly because their parameters are hard to individualize. In this work, we develop a BG prediction algorithm based on a personalized physiological model inspired by the UVA/Padova T1D Simulator. Then we compare white-box and advanced black-box personalized prediction techniques.

METHODS

A personalized nonlinear physiological model is identified from patient data through a Bayesian approach based on Markov Chain Monte Carlo technique. The individualized model was integrated within a particle filter (PF) to predict future BG concentrations. The black-box methodologies considered are non-parametric models estimated via gaussian regression (NP), three deep learning methods: long-short-term-memory (LSTM), gated recurrent unit (GRU), temporal convolutional networks (TCN), and a recursive autoregressive with exogenous input model (rARX). BG forecasting performances are assessed for several prediction horizons (PH) on 12 individuals with T1D, monitored in free-living conditions under open-loop therapy for 10 weeks.

RESULTS

NP models provide the most effective BG predictions by achieving a root mean square error (RMSE), RMSE = 18.99 mg/dL, RMSE = 25.72 mg/dL and RMSE = 31.60 mg/dL, significantly outperforming: LSTM, GRU (for PH = 30 minutes), TCN, rARX, and the proposed physiological model for PH=30, 45 and 60 minutes.

CONCLUSIONS

Black-box strategies remain preferable for glucose prediction even when compared to a white-box model with sound physiological structure and individualized parameters.

Combined Use of Glucose-Specific Model Identification and Alarm Strategy Based on Prediction-Funnel to Improve Online Forecasting of Hypoglycemic Events

BACKGROUND

Advanced decision support systems for type 1 diabetes (T1D) management often embed prediction modules, which allow T1D people to take preventive actions to avoid critical episodes like hypoglycemia. Real-time prediction of blood glucose (BG) concentration relies on a subject-specific model of glucose-insulin dynamics. Model parameter identification is usually based on the mean square error (MSE) cost function, and the model is usually used to predict BG at a single prediction horizon (PH). Finally, a hypo-alarm is raised if the predicted BG crosses a threshold. This work aims to show that real-time hypoglycemia forecasting can be improved by leveraging: a glucose-specific mean square error (gMSE) cost function in model's parameters identification, and a "prediction-funnel," that is, confidence intervals (CIs) for multiple PHs, within the hypo-alarm-raising strategy.

METHODS

Autoregressive integrated moving average with exogenous input (ARIMAX) models are selected to illustrate the proposed solution (use of gMSE and prediction-funnel) and its assessment against the conventional approach (MSE and single PH). The gMSE penalizes the model misfit in unsafe BG ranges (e.g., hypoglycemia), and the prediction-funnel allows raising an alarm by monitoring if the CIs cross a suitable threshold. The algorithms were evaluated by measuring precision (P), recall (R), F1-score (F1), false positive per day (FP/day), and time gain (TG) on a real dataset collected in 11 T1D individuals.

RESULTS

The best performance is achieved exploiting both the gMSE and the prediction-funnel: P = 65%, R = 88%, F1 = 75%, FP/day = 0.29, and mean TG = 15 minutes.

CONCLUSIONS

The combined use of a glucose-specific metric and an alarm-raising strategy based on the prediction-funnel allows achieving a more effective and reliable hypoglycemia prediction algorithm.

In silico design and validation of a time-varying PID controller for an artificial pancreas with intraperitoneal insulin delivery and glucose sensing

Type 1 diabetes (T1D) is a chronic autoimmune disease featured by the loss of beta cell function and the need for lifetime insulin replacement. Over the recent decade, the use of automated insulin delivery systems (AID) has shifted the paradigm of treatment: the availability of continuous subcutaneous (SC) glucose sensors to guide SC insulin delivery through a control algorithm has allowed, for the first time, to reduce the daily burden of the disease as well as to abate the risk for hypoglycemia. AID use is still limited by individual acceptance, local availability, coverage, and expertise. A major drawback of SC insulin delivery is the need for meal announcement and the peripheral hyperinsulinemia that, over time, contributes to macrovascular complications. Inpatient trials using intraperitoneal (IP) insulin pumps have demonstrated that glycemic control can be improved without meal announcement due to the faster insulin delivery through the peritoneal space. This calls for novel control algorithms able to account for the specificities of IP insulin kinetics. Recently, our group described a two-compartment model of IP insulin kinetics demonstrating that the peritoneal space acts as a virtual compartment and IP insulin delivery is virtually intraportal (intrahepatic), thus closely mimicking the physiology of insulin secretion. The FDA-accepted T1D simulator for SC insulin delivery and sensing has been updated for IP insulin delivery and sensing. Herein, we design and validate-in silico-a time-varying proportional integrative derivative controller to guide IP insulin delivery in a fully closed-loop mode without meal announcement.

Assessment of Seasonal Stochastic Local Models for Glucose Prediction without Meal Size Information under Free-Living Conditions

Accurate blood glucose (BG) forecasting is key in diabetes management, as it allows preventive actions to mitigate harmful hypoglycemic/hyperglycemic episodes. Considering the encouraging results obtained by seasonal stochastic models in proof-of-concept studies, this work assesses the methodology in two datasets (open-loop and closed-loop) recorded in free-living conditions. First, similar postprandial glycemic profiles are grouped together with fuzzy C-means clustering. Then, a seasonal stochastic model is identified for each cluster. Finally, real-time BG forecasting is performed by weighting each model’s prediction. The proposed methodology (named C-SARIMA) is compared to other linear and nonlinear black-box methods: autoregressive integrated moving average (ARIMA), its variant with input (ARIMAX), a feed-forward neural network (NN), and its modified version (NN-X) fed by BG, insulin, and carbohydrates (timing and dosing) information for several prediction horizons (PHs). In the open-loop dataset, C-SARIMA grants a median root-mean-squared error (RMSE) of 20.13 mg/dL (PH = 30) and 27.23 mg/dL (PH = 45), not significantly different from ARIMA and NN. Over a longer PH, C-SARIMA achieves an RMSE = 31.96 mg/dL (PH = 60) and RMSE = 33.91 mg/dL (PH = 75), significantly outperforming the ARIMA and NN, without significant differences from the ARIMAX for PH ≥ 45 and the NN-X for PH ≥ 60. Similar results hold on the closed-loop dataset: for PH = 30 and 45 min, the C-SARIMA achieves an RMSE = 21.63 mg/dL and RMSE = 29.67 mg/dL, not significantly different from the ARIMA and NN. On longer PH, the C-SARIMA outperforms the ARIMA for PH > 45 and the NN for PH > 60 without significant differences from the ARIMAX for PH ≥ 45. Although using less input information, the C-SARIMA achieves similar performance to other prediction methods such as the ARIMAX and NN-X and outperforming the CGM-only approaches on PH > 45min.