Continuous Glucose Monitoring sensors (CGMs) have revolutionized type 1 diabetes (T1D) management. In particular, in several cases, the retrospective analysis of CGM recordings allows clinicians to review and adjust patients' therapy. However, in this set-up, the artifacts that are often present in CGM data could lead to incorrect therapeutic actions. To mitigate this risk, we investigate how to detect one of the most common of these artifacts, the so-called pressure induced sensor attenuations, by means of anomaly detection algorithms. Specifically, these methods belong to the class of unsupervised techniques, which is particularly appealing since it does not require a labeled dataset, hardly available in practice. After having designed five features to highlight the anomalous state of the sensor, 8 different methods (e.g. Isolation Forest and Histogram-based Outlier Score) are assessed both in silico using the UVa/Padova Type 1 Diabetes Simulator and on real data of 36 subjects monitored for about 10 days. In the in silico scenario, the best results are achieved with Isolation Forest, which recognized the 74% of the failures generating on average only 2 false alerts during the whole monitoring time. In real data, Isolation Forest is confirmed to be effective in the detection of failures, achieving a recall of 55% and generating 3 false alarms in 10 days. By allowing to detect more than 50% of the artifacts while discarding only a few portions of correct data in several days of monitoring, the proposed approach could effectively improve the quality of CGM data used by clinicians to retrospectively evaluate and adjust T1D therapy.

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.

Type 1 diabetes (T1D) management in pediatric patients presents unique challenges due to the evolving physiology and the need for close monitoring and intervention. This article presents the system architecture of TWIN, a novel personalized Decision Support System (DSS) for managing multiple daily injections (MDI) therapy in pediatric T1D patients. TWIN incorporates advanced technologies such as continuous glucose monitoring (CGM) devices, smart insulin pens, and wearable physical activity trackers to collect large amounts of user-generated data. At its core, the system leverages an open-source digital twinning methodology called ReplayBG to create virtual clones of patients, representing their unique physiological characteristics. These virtual clones are then utilized in multiple simulation iterations to personalize and optimize therapy parameters. To promote the transparency of the generated results, TWIN employs a large-scale linguistic model based on a Generative Pre-trained Transformer (GPT). This model provides clear explanations and contextual information regarding the recommended therapy parameters. Finally, the proposed TWIN system architecture integrates these algorithms within a state-of-the-art digital platform. This platform offers a user-friendly interface for patients and healthcare providers, enabling effective management and tuning of MDI therapy. Future work will focus on testing and validating TWIN to assess its efficacy and usability in pediatric T1D management.

Continuous Glucose Monitors are minimally-invasive portable sensors that are revolutionizing the management of Type 1 Diabetes (T1D). A common issue encountered in their daily use is related to the presence of pressure-induced sensor attenuations (PISAs), temporary faults of the devices, resulting in false low blood glucose readings that can impact and compromise the reliability of CGMs. In this work, we explore the application of matched filters (MFs), a powerful pattern recognition technique, for the retrospective identification of PISAs failures. A MF is designed for the detection of a signal with a specific shape, associated with the occurrence of a PISA episode. The proposed algorithm is tested in-silico on a dataset generated with a state-of-art T1D patient simulator. MFs achieve a recall of 0.75 with about 1 false alarm every 5 days, outperforming other state-of-art algorithms proposed for the same purpose, including one based on a Random Forest classifier (RF). Moreover, when embedded as additional feature within a RF it improves the performance by granting a recall of 0.83 and 1 false alarm raised in 10 days. The encouraging outcomes in the simulated scenario pave the way for future investigations involving real-world data, as well as potential enhancements in detecting different types of sensors’ failures.

Type 1 diabetes (T1D) is a chronic condition characterized by elevated blood glucose (BG) levels resulting from the pancreas’ inability to secrete insulin. To keep their BG levels within the safe range of [70−180] mg/dL, individuals affected by T1D must adhere to a lifelong therapy, which involves multiple daily actions and may negatively impact patients’ quality of life.To address these challenges, decision support systems (DSSs) utilizing continuous glucose monitoring sensors have become vital in managing T1D. These tools assist individuals suggesting therapeutic actions like carbohydrate intake and insulin injections. In this context, accurate algorithms predicting future BG levels are essential for proactive interventions, improving glucose control and enhancing patient well-being.This paper presents a new algorithm for DSSs that recommends corrective insulin boluses (CIBs). The core of the proposed DSS is a BG predictive algorithm based on a long short-term memory (LSTM) neural network which has been trained on 12 patients with T1D monitored for 10 weeks. The proposed algorithm, named LSTM-CIB, has been retrospectively evaluated on an external test set composed by 30 T1D patients monitored in free-living conditions. Compared to a state-of-art heuristic-based strategy for hyperglycemia correction, LSTM-CIB significantly reduces time spent in hyperglycemia (33.26% vs. 39.42%) and increases time spent in euglycemia (65.29% vs. 59.16%), also improving patient safety.

The organisation of care at diabetes outpatient clinics is typically different from that delivered by general practitioners, it is thus of interest to assess whether there is also a difference in the predictive power of heart failure hospitalisation risk scores developed independently for each subpopulation. To such a purpose, a diabetes outpatient clinic dataset and a primary care dataset were considered. A Cox proportional hazard model, an accelerated failure time model, a logistic regression, a random forest, and a K-nearest neighbours model were trained in each dataset and tested on both. The UK Prospective Diabetes Study (UKPDS) risk engine was used as benchmark. Results show that models developed using primary care data performed well on the corresponding test set but poorly when used in the diabetes outpatient clinic setting (best C-Index = 0.759vs. 0.615, best AUROC = 0.757vs. 0.598). Models trained on the diabetes outpatient clinic data performed well on the corresponding test set, and their predictive power in the primary care setting was not statistically different from the one of models developed using primary care data (best C-Index = 0.814 vs 0.740, best AUROC = 0.812 vs 0.750). In both settings UKPDS had lower predictive power than the best newly-developed models. Different care setting led to a difference in the predictive power of heart failure hospitalisation risk scores that depended on both the data used for training and the methodological approach chosen. This suggests the need to consider these factors when applying risk scores to a target population where the expected incidence of the outcome and the distribution of baseline covariates differ from those of the population for which scores were proposed.

People with diabetes require lifelong access to healthcare services to delay the onset of complications. Their disease management processes generate great volumes of data across several domains, from clinical to administrative. Difficulties in accessing and processing these data hinder their secondary use in an institutional setting, even for highly desirable applications, such as the prediction of cardiovascular disease, the main driver of excess mortality in diabetes. Hence, in the present work, we propose a deep learning model for the prediction of major adverse cardiovascular events (MACE), developed and validated using the administrative claims of 214,676 diabetic patients of the Veneto region, in North East Italy. Specifically, we use a year of pharmacy and hospitalisation claims, together with basic patient's information, to predict the 4P-MACE composite endpoint, i.e., the first occurrence of death, heart failure, myocardial infarction, or stroke, with a variable prediction horizon of 1 to 5 years. Adapting to the time-to-event nature of this task, we cast our problem as a multi-outcome (4P-MACE and components), multi-label (1 to 5 years) classification task with a custom loss to account for the effect of censoring. Our model, purposefully specified to minimise data preparation costs, exhibits satisfactory performance in predicting 4P-MACE at all prediction horizons: AUROC from 0.812 (C.I.: 0.797 - 0.827) to 0.792 (C.I.: 0.781 - 0.802); C-index from 0.802 (C.I.: 0.788 - 0.816) to 0.770 (C.I.: 0.761 - 0.779). Components' prediction performance is also adequate, ranging from death's 0.877 1-year AUROC to stroke's 0.689 5-year AUROC.

Kinetic modeling is used to indirectly measure physiological parameters from dynamic positron emission tomography (PET) data. Usually, the unknown parameters of the model are estimated, in any given region of interest (ROI), by least squares (LS). However, when the signal-to-noise ratio (SNR) of PET data is too low, LS does not allow reliable parameter estimation. To overcome this problem, we study in this paper the applicability of approaches originally developed in the pharmacokinetic/pharmacodynamic literature and referred to as "population approaches." In particular, we consider the iterative two stage (ITS) method, which, given a set of M ROIs drawn on PET images of a given individual, estimates the unknown model parameters of each ROI by exploiting the information contained in all the M ROIs. After having revised the theory behind ITS, we assess its performance versus LS by using Monte Carlo simulations which allow us to evaluate the bias of the two methods in a variety of situations. Then, we compare the performance of LS and ITS in two case studies on [18F]FDG kinetics in human skeletal muscle. Both simulated and real case studies results show that a population approach is of potential in modeling PET images since it allows to reliably estimate model parameters also in those ROIs where either a bad SNR or a poor sampling (e.g., infrequent scanning and/or short experiment duration) make the use of LS unsuccessful.