Search

Abstract Machine learning has become a popular tool for learning models of complex dynamics from biomedical data. In Type 1 Diabetes (T1D) management, these models are increasingly been integrated in decision support systems (DSS) to forecast glucose levels and provide preventive therapeutic suggestions, like corrective insulin boluses (CIB), accordingly. Typically, models are chosen based on their prediction accuracy. However, since patient safety is a concern in this application, the algorithm should also be physiologically sound and its outcome should be explainable. This paper aims to discuss the importance of using tools to interpret the output of black-box models in T1D management by presenting a case-of-study on the selection of the best prediction algorithm to integrate in a DSS for CIB suggestion. By retrospectively “replaying” real patient data, we show that two long-short term memory neural networks (LSTM) (named p-LSTM and np-LSTM) with similar prediction accuracy could lead to different therapeutic decisions. An analysis with SHAP—a tool for explaining black-box models’ output—unambiguously shows that only p-LSTM learnt the physiological relationship between inputs and glucose prediction, and should therefore be preferred. This is verified by showing that, when embedded in the DSS, only p-LSTM can improve patients’ glycemic control.

Postbariatric hypoglycemia (PBH) is an increasingly recognized late metabolic complication of bariatric surgery, characterized by low blood glucose levels 1–3 h after a meal, particularly if the meal contains rapid-acting carbohydrates. PBH can often be effectively managed through appropriate nutritional measures, which remain the cornerstone treatment today. However, their implementation in daily life continues to challenge both patients and health care providers. Emerging digital technologies may allow for more informed and improved decision-making through better access to relevant data to manage glucose levels in PBH. Examples include applications for automated food analysis from meal images, digital receipts of purchased food items or integrated platforms allowing the connection of continuously measured glucose with food and other health-related data. The resulting multi-dimensional data can be processed with artificial intelligence systems to develop prediction algorithms and decision support systems with the aim of improving glucose control, safety, and quality of life of PBH patients. Digital innovations, however, face trade-offs between user burden vs. amount and quality of data. Further challenges to their development are regulatory non-compliance regarding data ownership of the platforms acquiring the required data, as well as user privacy concerns and compliance with regulatory requirements. Through navigating these trade-offs, digital solutions could significantly contribute to improving the management of PBH.

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.

In type 1 diabetes management, the availability of algorithms capable of accurately forecasting future blood glucose (BG) concentrations and hypoglycemic episodes could enable proactive therapeutic actions, e.g., the consumption of carbohydrates to mitigate, or even avoid, an impending critical event. The only input of this kind of algorithm is often continuous glucose monitoring (CGM) sensor data, because other signals (such as injected insulin, ingested carbs, and physical activity) are frequently unavailable. Several predictive algorithms fed by CGM data only have been proposed in the literature, but they were assessed using datasets originated by different experimental protocols, making a comparison of their relative merits difficult. The aim of the present work was to perform a head-to-head comparison of thirty different linear and nonlinear predictive algorithms using the same dataset, given by 124 CGM traces collected over 10 days with the newest Dexcom G6 sensor available on the market and considering a 30-min prediction horizon. We considered the state-of-the art methods, investigating, in particular, linear black-box methods (autoregressive; autoregressive moving-average; and autoregressive integrated moving-average, ARIMA) and nonlinear machine-learning methods (support vector regression, SVR; regression random forest; feed-forward neural network, fNN; and long short-term memory neural network). For each method, the prediction accuracy and hypoglycemia detection capabilities were assessed using either population or individualized model parameters. As far as prediction accuracy is concerned, the results show that the best linear algorithm (individualized ARIMA) provides accuracy comparable to that of the best nonlinear algorithm (individualized fNN), with root mean square errors of 22.15 and 21.52 mg/dL, respectively. As far as hypoglycemia detection is concerned, the best linear algorithm (individualized ARIMA) provided precision = 64%, recall = 82%, and one false alarm/day, comparable to the best nonlinear technique (population SVR): precision = 63%, recall = 69%, and 0.5 false alarms/day. In general, the head-to-head comparison of the thirty algorithms fed by CGM data only made using a wide dataset shows that individualized linear models are more effective than population ones, while no significant advantages seem to emerge when employing nonlinear methodologies.