Advanced Diabetes Management Using Artificial Intelligence and Continuous Glucose Monitoring Sensors

May combine two claims.

CGM sensors allow tracking glucose dynamics and prompt detection of asymptomatic high/low glucose events, which SMBG cannot.

“Indeed, in contrast to SMBG, CGM sensors allow to track glucose dynamics and they ensure prompt detection of even asymptomatic high/low glucose events, which are difficult to capture with SMBG.”

• cites [10] Cappon, 2019, Continuous glucose monitoring sensors for diabetes management: A review of technologies and applications, Diabetes Metab. J. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

DSSs for patient use receive input about glucose concentration, diet, insulin administrations and physical activity, and primarily perform three tasks: personalized insulin bolus calculation, adaptive tuning of BC parameters, and glucose prediction.

“The DSSs for patient use proposed in the literature receive input information about glucose concentration, diet, insulin administrations and physical activity and primarily perform three tasks: Personalized insulin bolus calculation, adaptive tuning of BC parameters and glucose prediction (Figure 1).”

May combine two claims.

The standard formula for bolus computation was designed when SMBG was the only way to measure glucose at home, and it only uses current glucose concentration, ignoring glucose rate-of-change available with CGM.

“The SF was designed when SMBG was the only way to measure glucose concentration at home. In fact, the SF only uses information about the current glucose concentration (Gc), ignoring the information on glucose rate-of-change (ROC) that is now available with modern CGM sensors.”

May combine two claims.

Machine learning methods can be used to design new data-driven formulae for insulin dosing that naturally account for CGM information including glucose rate-of-change and could improve glycemic control in type 1 diabetes.

“In conclusion, machine learning methods can be used to design new data-driven formulae for insulin dosing that are able to naturally account for CGM information (including glucose ROC) and which could improve glycemic control in T1D.”

May combine two claims.

Although on average the NN-based bolus calculator improved glycemic control compared to the standard formula, in a non-negligible number of cases the dose provided by the NN-based bolus calculator led to worse glycemic control than the standard formula.

“Although, on average, the NN-based BC did improve BG control when compared to the SF, in a non-negligible number of cases the dose provided by the NN-based BC led to worse glycemic control than did the SF.”

• cites [37] Cappon, 2018, A Neural-network-based approach to personalize insulin bolus calculation using continuous glucose monitoring, J. Diabetes Sci. Technol. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

DSSs for patients offer an alternative to closed-loop systems because many users do not feel confident using the artificial pancreas.

“These DSSs for patients offer an alternative to closed-loop systems. Indeed, many users do not feel confident about using the artificial pancreas [28,29], as they are concerned about possible errors occurring in the insulin pump and so prefer an open-loop therapy (i.e., a therapy in which the patient manually sets the insulin doses), which can be assisted by a DSS.”

• cites [28] Hendrieckx, 2017, “It is definitely a game changer”: A qualitative study of experiences with in-home overnight closed-loop technology among adults with type 1 diabetes, Diabetes Technol. Ther. sourceNot checked yet

  Not checked in the free run.

• cites [29] Kropff, 2016, Psychological outcomes of evening and night closed-loop insulin delivery under free living conditions in people with Type 1 diabetes: A 2-month randomized crossover trial, Diabet. Med. sourceNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

Results from a simulated experiment suggested that run-to-run offers a promising approach to optimizing standard formula parameters.

“The results obtained from a simulated experiment suggested that R2R offers a promising approach to optimizing SF parameters [41].”

• cites [41] Herrero, 2015, Method for automatic adjustment of an insulin bolus calculator: In silico robustness evaluation under intra-day variability, Comput. Methods Programs Biomed. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Since 2015, some rtCGM devices have received regulatory approval for therapeutic use, meaning rtCGM measurements may be used for making therapeutic decisions such as insulin dosing.

“Since 2015, some rtCGM devices have also received regulatory approval for therapeutic use, which means that rtCGM measurements may be used for making therapeutic decisions, such as insulin dosing [19].”

• cites [19] Edelman, 2016, Regulation Catches Up to Reality, J. Diabetes Sci. Technol. sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Herrero et al. proposed an approach to automatically adjust carbohydrate-to-insulin ratio and correction factor using CGM measurements, updating parameter values daily in a run-to-run fashion.

“Herrero et al. [41] proposed an approach to automatically adjust CR and CF using CGM measurements. This method updates the parameters values every day in a run-to-run (R2R) fashion.”

• cites [41] Herrero, 2015, Method for automatic adjustment of an insulin bolus calculator: In silico robustness evaluation under intra-day variability, Comput. Methods Programs Biomed. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

An in silico head-to-head comparison of empirical methods that adjust the standard formula according to glucose rate-of-change showed that their efficacy depends on the pre-meal scenario and no method globally outperforms the others.

“However, an in silico head-to-head comparison of these methods has shown that their efficacy depends on the pre-meal scenario and that not one of these methodologies globally outperforms the others [36].”

• cites [36] Cappon, 2018, In Silico Assessment of literature insulin bolus calculation methods accounting for glucose rate of change, J. Diabetes Sci. Technol. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

rtCGM devices provide real-time glucose concentration and trend information, which can be used for making more informed therapeutic decisions.

“rtCGM devices are able to provide, in real-time, not only the glucose concentration present but also information about the glucose trend (typically displayed to the user by an arrow), which can be used for making more informed therapeutic decisions.”

May combine two claims.

The standard formula is intrinsically suboptimal, and using it as a starting point to be corrected might prove to be an obstacle in the training process of the neural network.

“Indeed, the SF is intrinsically suboptimal, and using it as a starting point to be corrected might prove to be an obstacle in the training process of NN [36].”

• cites [36] Cappon, 2018, In Silico Assessment of literature insulin bolus calculation methods accounting for glucose rate of change, J. Diabetes Sci. Technol. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

The PEPPER project developed a DSS for personalized T1D management including an adaptive bolus advisor based on case-based reasoning and a carbohydrate recommendation system based on a predictive model.

“For example, PEPPER (Patient Empowerment through Predictive PERsonalised decision support), a project funded by the European Commission under the Horizon 2020 program, recently developed a DSS for personalized T1D management that includes an adaptive bolus advisor, based on case-based reasoning, as well as a system for carbohydrate recommendations based on a predictive model [31].”

• cites [31] Herrero, P., Lopez, B., and Martin, C. (2016, January 30). PEPPER: Patient empowerment through predictive personalized decision support. Proceedings of the AID: Artificial Intelligence for Diabetes: 1st ECAI Workshop on Artificial Intelligence for Diabetes at the 22nd European Conference on Artificial Intelligence (ECAI 2016), The Hague, The Netherlands.Not checked yet

  Not checked in the free run.

May combine two claims.

In a cohort of 20 virtual subjects, the R2R+CBR algorithm significantly improved glycemic control, achieving better results than the stand-alone R2R algorithm.

“The results obtained in a cohort of 20 virtual subjects showed that the adoption of R2R+CBR significantly improved glycemic control, achieving better results than the stand-alone R2R algorithm.”

• cites [42] Herrero, 2014, Advanced insulin bolus advisor based on run-to-run control and case-based reasoning, IEEE J. Biomed. Health Inform.Not checked yet

  Not checked in the free run.

May combine two claims.

The study demonstrated the safety and effectiveness of ABC4D, which made it possible to reduce the occurrence of postprandial hypoglycemia.

“The study demonstrated the safety and effectiveness of ABC4D, which made it possible to reduce the occurrence of postprandial hypoglycemia.”

• cites [45] Reddy, 2016, Clinical safety and feasibility of the advanced bolus calculator for type 1 diabetes based on case-based reasoning: A 6-week nonrandomized single-arm pilot study, Diabetes Technol. Ther. sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Current research in diabetes is putting considerable effort into developing algorithms and software applications to enhance type 1 diabetes therapy by exploiting the latest technologies and their interoperability.

“Current research in diabetes is putting considerable effort into developing algorithms and software applications to enhance T1D therapy by exploiting the latest technologies and their interoperability.”

May combine two claims.

Sun et al.'s reinforcement learning algorithm for insulin therapy optimization showed encouraging results in a simulated study, suggesting it could personalize insulin therapy parameters and improve glycemic control.

“The encouraging results obtained in a simulated study suggest that reinforcement learning could be a promising approach, one able to both personalize insulin therapy parameters and improve glycemic control.”

• cites Sun et al. [48] Sun, 2019, A dual mode adaptive basal-bolus advisor based on reinforcement learning, IEEE J. Biomed. Health Inform. sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The neural network hyper-parameters were tuned using a five-fold cross-validation procedure.

“The NN hyper-parameters (i.e., the number of neurons per layer and the activation function type) were tuned using a five-fold CV procedure, in other words, the training set was divided into five groups of data, four of which were used for training the NN and the remaining one used as a validation set to compute the mean squared error (MSE).”

• cites [37] Cappon, 2018, A Neural-network-based approach to personalize insulin bolus calculation using continuous glucose monitoring, J. Diabetes Sci. Technol. sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based).

The increasing availability of information offered by wearable devices such as CGM sensors, insulin pumps, smart pens, and activity trackers is making the decision-making process increasingly complex.

“The increasing availability of information offered by wearable devices, such as CGM sensors, insulin pumps, smart pens, and activity trackers, is making the decision-making process increasingly complex.”

May combine two claims.

Run-to-run assumes the process to be controlled (blood glucose level) is repetitive, which is somewhat unrealistic in many cases and limits run-to-run applicability.

“One fundamental issue of R2R is that it assumes that the process that is to be controlled, in our case the BG level, is repetitive. This assumption is somewhat unrealistic in many cases, thus it limits R2R applicability.”

May combine two claims.

The most recent CGM sensors adopt simple prediction algorithms, such as linear extrapolation, to predict future interstitial glucose concentrations 15-30 min in advance in real-time and generate predictive alerts when a hypo/hyperglycemic event is foreseen.

“The most recent CGM sensors adopt very simple prediction algorithms, e.g., linear extrapolation, to predict future interstitial glucose concentrations (e.g., 15–30 min in advance) in real-time and generate predictive alerts when a hypo/hyperglycemic event is foreseen [49].”

• cites [49] Abraham, S.B., Arunachalam, S., Zhong, A., Agrawal, P., Cohen, O., and McMahon, C.M. (2019). Improved real-world glycemic control with continuous glucose monitoring system predictive alerts. J. Diabetes Sci. Technol. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

The R2R+CBR algorithm was able to tackle intrasubject variability and external perturbations, and was robust to uncertainty in carbohydrate input and noise in CGM measurements.

“The proposed R2R+CBR algorithm was able to tackle intrasubject variability and external perturbations, and it was shown to be robust to uncertainty in carbohydrate input (i.e., carb-counting error) and noise in CGM measurements.”

• cites [42] Herrero, 2014, Advanced insulin bolus advisor based on run-to-run control and case-based reasoning, IEEE J. Biomed. Health Inform.Not checked yet

  Not checked in the free run.

May combine two claims.

In a virtual cohort of 100 adults, Cappon et al.'s XGB-based bolus calculator, compared to the standard formula, significantly improved time in euglycemia from 61.98% to 67.00% without increasing time in hypoglycemia.

“The results obtained in a virtual cohort of 100 adults, showed that, compared to the SF, the XGB-based BC significantly improved time in euglycemia (from 61.98% to 67.00%) without increasing time in hypoglycemia.”

• cites Cappon et al. [59] Cappon, G., Facchinetti, A., Sparacino., G., Georgiou, P., and Herrero, P. (2019). Classification of postprandial glycemic status with application to insulin dosing in type 1 diabetes—An in silico proof-of-concept. Sensors, 19. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Most glucose prediction algorithms in the literature approach glucose prediction as a regression problem, seeking to predict interstitial glucose concentration with a prediction horizon typically below 60 min, with 30 min the most frequent.

“Most of the prediction algorithms in the literature approach the problem of glucose prediction as a regression problem, in other words they seek to predict the value of interstitial glucose concentration with a prediction horizon (PH) typically below 60 min (30 min is the most frequent PH value) [52].”

• cites [52] Oviedo, 2016, A review of personalized blood glucose prediction strategies for T1DM patients, Int. J. Numer. Methods Biomed. Eng. sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based).

Most studies on glucose prediction have focused on forecasting future glucose concentration levels using a regression approach, usually with a short prediction horizon, such as 30 min.

“Most studies on glucose prediction have focused on forecasting future glucose concentration levels using a regression approach, usually with a short PH (e.g., 30 min).”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Although each module of Vehì et al.'s Patient Safety System has been independently validated on retrospective data, a clinical validation of the proposed DSS is needed to prove its efficacy as a tool for improving glycemic control in subjects with T1D.

“Although each module has been independently validated on retrospective data, a clinical validation of the proposed DSS is needed to prove its efficacy as a tool for improving glycemic control in subjects with T1D.”

• cites [61] Vehi, 2019, Prediction and prevention of hypoglycaemic events in type-1 diabetic patients using machine learning, Health Inform. J. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

AI techniques, especially supervised machine learning techniques, are extensively used for building black-box models for glucose prediction, and neural networks are the most used method.

“As reported in recent systematic reviews [25,54], AI techniques, especially supervised machine learning techniques, are extensively used for building black-box models for glucose prediction. The most used method is NN, but other machine learning methodologies, such as random forest, support vector regression, regression algorithms and deep-learning techniques.”

• cites [25] Contreras, 2018, Artificial intelligence for diabetes management and decision support: Literature review, J. Med. Internet Res. sourceNot checked yet

  Not checked in the free run.

• cites [54] Woldaregay, 2019, Data-driven modeling and prediction of blood glucose dynamics: Machine learning applications in type 1 diabetes, Artif. Intell. Med. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

In Xie and Wang's comparison, the models showing best performance in terms of RMSE were ARX, Ridge regression, linear SVR, LSTM, and TCN, with RMSE between 19 and 20 mg/dL.

“The models showing best performance in terms RMSE were ARX, Ridge regression, linear SVR, LSTM, and TCN, with RMSE between 19 and 20 mg/dL.”

• cites [58] Xie, J., and Wang, Q. (2020). Benchmarking machine learning algorithms on blood glucose prediction for Type 1 Diabetes in comparison with classical time-series models. IEEE Trans. Biomed. Eng., 1. sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based).

Predicting the exact glucose concentration level over a long prediction horizon, such as 2-6 h, would be a very challenging task.

“On the other hand, predicting the exact glucose concentration level over such a long PH would be a very challenging task.”

May combine two claims.

In Xie and Wang's comparison, considering both RMSE and time gain, linear SVR and TCN performed better, with RMSE around 20 mg/dL and a time gain of about 6 min.

“Considering both of the metrics, the models that performed better were linear SVR and TCN with RMSE around 20 mg/dL and a TG of about 6 min.”

• cites [58] Xie, J., and Wang, Q. (2020). Benchmarking machine learning algorithms on blood glucose prediction for Type 1 Diabetes in comparison with classical time-series models. IEEE Trans. Biomed. Eng., 1. sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Subject-specific glucose prediction algorithms train a different model for each subject to account for the large inter-individual variability characterizing individuals with T1D.

“Another option is to develop subject-specific algorithms, where a different model is trained for each subject, in order to take into account the large inter-individual variability characterizing individuals with T1D [52].”

• cites [52] Oviedo, 2016, A review of personalized blood glucose prediction strategies for T1DM patients, Int. J. Numer. Methods Biomed. Eng. sourceNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim. May combine two claims.

In Reddy et al.'s study, all data were collected in a controlled inpatient environment, which is a limitation; further validation with data gathered under free-living conditions and with different types of exercise is required.

“However, one limitation of this study is that all the data were collected in a controlled inpatient environment. Further validation of the approach with data gathered under free-living conditions, and with different types of exercise, is required.”

• cites Reddy et al. [60] Reddy, 2019, Prediction of hypoglycemia during aerobic exercise in adults with type 1 diabetes, J. Diabetes Sci. Technol. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

In Xie and Wang's comparison, the maximum time gain was achieved for Gradient boosting trees and LSTM, which predicted CGM data on average 7 min in advance.

“The maximum TG was achieved for Gradient boosting trees and LSTM, which made it possible to predict the CGM data, on average, 7 min in advance.”

• cites [58] Xie, J., and Wang, Q. (2020). Benchmarking machine learning algorithms on blood glucose prediction for Type 1 Diabetes in comparison with classical time-series models. IEEE Trans. Biomed. Eng., 1. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Data collected by activity trackers can be used to automatically detect exercise sessions, predict exercise-induced hypoglycemia, and if necessary recommend the consumption of carbohydrates to avoid hypoglycemia.

“The data collected by activity trackers can be used to automatically detect exercise sessions [65], predict exercise-induced hypoglycemia [60], and if necessary, recommend the consumption of carbohydrates to avoid hypoglycemia.”

• cites [65] Dasanayake, 2015, Early detection of physical activity for people with type 1 diabetes mellitus, J. Diabetes Sci. Technol. sourceNot checked yet

  Not checked in the free run.

• cites [60] Reddy, 2019, Prediction of hypoglycemia during aerobic exercise in adults with type 1 diabetes, J. Diabetes Sci. Technol. sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

AI-enabled DSSs are promising tools that could potentially reduce the patient burden in diabetes management and improve glucose control.

“Amongst the possible applications, AI-enabled DSSs are promising tools that could potentially reduce the patient burden in diabetes management and improve glucose control.”

Probably not a substantial claim: may not pass S2 (checkable).

A prospective clinical validation proving the safety and effectiveness of AI-based DSS applications is required to support their adoption in T1D management.

“A prospective clinical validation proving the safety and effectiveness of such techniques is required to support the adoption of AI-based DSS applications in T1D management.”