Mansur Shomali
Research Subjects
Mansur Shomali has a total of 23 co-authors. Here are the top co-authors by number of shared publications:
Abhimanyu Kumbara
Welldoc, Inc., Columbia, MD, United States
4 shared publicationsAnand Iyer
Welldoc, Inc., Columbia, MD, United States
4 shared publications- Abhimanyu Kumbara
Welldoc Inc., Columbia, USA
3 shared publications - Anand K. Iyer
Welldoc Inc., Columbia, MD, USA
3 shared publications - Courtney R. Green
Medical Affairs Dexcom, Inc 6340 Sequence Dr 92121 San Diego San Diego CA CA USA
2 shared publications Janice MacLeod
Janice MacLeod Consulting, Glen Burnie, MD, United States
2 shared publications
Latest Publications
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 9510-9521
EventGlucose: A Benchmark for Continuous Glucose Monitoring Forecasting with Behavioral Event Context
Glucose forecasting is critical for diabetes management. Continuous glucose monitoring (CGM) provides real-time readings, but glucose values alone do not capture the behavioral events (meals, exercise, medication) that clinicians and patients use to make daily decisions. In spite of advances in time series foundation models and large language models (LLMs), systematic study of event-aware glucose forecasting remains limited, and how behavioral event context helps different model families remains unclear. We introduce EventGlucose, a benchmark where every task pairs glucose time series with behavioral event context; every task requires event context for accurate prediction. EventGlucose provides 60 task instances spanning demographics, event types, and 5 context conditions, from no context to detailed event descriptions. We evaluate 18 model configurations zero-shot, covering LLM-based forecasters, multimodal architectures, time series foundation models, and statistical baselines. We report Clarke Error Grid-weighted CRPS metrics that emphasize clinical risk zones. Our experiments reveal a Context Integration Gap : richer context can help, but most current models fail to use it. Only GPT-4o-mini improves with more context, gaining 9% from NoCtx to StandardEventInfo. Ten of twelve context-aware models degrade with detailed descriptions, by up to 78% in ct-CRPS. Context effects are also asymmetric: meal tasks benefit in 4 of 6 demographic subgroups, while exercise tasks benefit in none, suggesting failures depend on event type rather than on context generally. Larger models also tend to degrade more than smaller ones within each model family. Despite EventGlucose's balanced design, Type 1 diabetes and middle-aged patients remain harder to predict than other subgroups. EventGlucose is the first benchmark to evaluate time series foundation models and LLMs on event-aware glucose forecasting, and it shows where current models break down when behavioral context is available. The benchmark is available at https://github.com/JHU-CDHAI/EventGlucoseBench.
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 9510-9521
Open AccessEventGlucose: A Benchmark for Continuous Glucose Monitoring Forecasting with Behavioral Event Context
Glucose forecasting is critical for diabetes management. Continuous glucose monitoring (CGM) provides real-time readings, but glucose values alone do not capture the behavioral events (meals, exercise, medication) that clinicians and patients use to make daily decisions. In spite of advances in time series foundation models and large language models (LLMs), systematic study of event-aware glucose forecasting remains limited, and how behavioral event context helps different model families remains unclear. We introduce EventGlucose, a benchmark where every task pairs glucose time series with behavioral event context; every task requires event context for accurate prediction. EventGlucose provides 60 task instances spanning demographics, event types, and 5 context conditions, from no context to detailed event descriptions. We evaluate 18 model configurations zero-shot, covering LLM-based forecasters, multimodal architectures, time series foundation models, and statistical baselines. We report Clarke Error Grid-weighted CRPS metrics that emphasize clinical risk zones. Our experiments reveal a Context Integration Gap : richer context can help, but most current models fail to use it. Only GPT-4o-mini improves with more context, gaining 9% from NoCtx to StandardEventInfo. Ten of twelve context-aware models degrade with detailed descriptions, by up to 78% in ct-CRPS. Context effects are also asymmetric: meal tasks benefit in 4 of 6 demographic subgroups, while exercise tasks benefit in none, suggesting failures depend on event type rather than on context generally. Larger models also tend to degrade more than smaller ones within each model family. Despite EventGlucose's balanced design, Type 1 diabetes and middle-aged patients remain harder to predict than other subgroups. EventGlucose is the first benchmark to evaluate time series foundation models and LLMs on event-aware glucose forecasting, and it shows where current models break down when behavioral context is available. The benchmark is available at https://github.com/JHU-CDHAI/EventGlucoseBench.
- Open Access
- 23 May 2025
- Published by Frontiers Media SA in Frontiers in Endocrinology
- Vol. 16, 1593321
Personalized cardiometabolic care powered by artificial intelligence
Advancements in artificial intelligence (AI) are providing a wealth of opportunities for improving clinical practice and healthcare delivery. It is predicted by AI experts that healthcare will change more in the next decade than it has in the previous century due to the volume and speed of these advancing capabilities. In this paper, we will illustrate the potential value of AI by sharing an example of an AI-powered digital health platform, designed to support people living with chronic cardiometabolic conditions and their care teams. The goal is to transform the care continuum from prevention through diagnosis, treatment, and ongoing management, including efficient acute care interventions when needed. The intent is to shift from reactive to proactive care including preventive-based guidance and interventions. AI-powered health technologies enable true person-centered care (i.e., for N=1), but for entire populations at scale (i.e., N >> 1), shifting the traditional mass generalization paradigm to one of mass customization. We demonstrate how an AI-powered digital health platform: 1) supports early detection and diagnosis of chronic conditions such as diabetes and related cardiometabolic conditions with data and insights; 2) optimizes personalized treatment; 3) tracks progress; 4) provides education; and 5) enables longitudinal behavior change to sustain health. We will present current AI capabilities as well as future considerations for the industry. We will also discuss principles that govern the responsible adoption of AI capabilities in healthcare to complement, not replace, the clinician.
- 23 May 2025
- Published by Frontiers Media SA in Frontiers in Endocrinology
- Vol. 16, 1593321
Open AccessPersonalized cardiometabolic care powered by artificial intelligence
Advancements in artificial intelligence (AI) are providing a wealth of opportunities for improving clinical practice and healthcare delivery. It is predicted by AI experts that healthcare will change more in the next decade than it has in the previous century due to the volume and speed of these advancing capabilities. In this paper, we will illustrate the potential value of AI by sharing an example of an AI-powered digital health platform, designed to support people living with chronic cardiometabolic conditions and their care teams. The goal is to transform the care continuum from prevention through diagnosis, treatment, and ongoing management, including efficient acute care interventions when needed. The intent is to shift from reactive to proactive care including preventive-based guidance and interventions. AI-powered health technologies enable true person-centered care (i.e., for N=1), but for entire populations at scale (i.e., N >> 1), shifting the traditional mass generalization paradigm to one of mass customization. We demonstrate how an AI-powered digital health platform: 1) supports early detection and diagnosis of chronic conditions such as diabetes and related cardiometabolic conditions with data and insights; 2) optimizes personalized treatment; 3) tracks progress; 4) provides education; and 5) enables longitudinal behavior change to sustain health. We will present current AI capabilities as well as future considerations for the industry. We will also discuss principles that govern the responsible adoption of AI capabilities in healthcare to complement, not replace, the clinician.
- 11 November 2024
- Published by American Diabetes Association in Diabetes Spectrum
- Vol. 38 (2) , 153-160
Safety of a Novel Continuous Glucose Monitoring–Informed Insulin Bolus Calculator Mobile Application for People With Type 1 or Type 2 Diabetes
BACKGROUND: Managing bolus insulin dosing can be a significant burden for people with diabetes, many of whom have limited numeracy skills. Insulin bolus calculators (IBCs) may improve glycemia as well as treatment satisfaction. OBJECTIVE: The purpose of this study was to demonstrate the safety of a novel, continuous glucose monitoring (CGM)-informed IBC mobile device app that applies trend arrow adjustments to bolus insulin dose recommendations. RESEARCH DESIGN AND METHODS: This clinical trial was an open-label, industry-sponsored single-arm study conducted at two sites. Fifty-four participants with type 1 or type 2 diabetes were enrolled and used the IBC app on their mobile device for 30 days. Study participants were adults who were already using CGM and dosing bolus insulin. The analysis examined both noninferiority and superiority of time in range (TIR) during the study period compared with baseline. Other important end points included hypoglycemia, glucose variability, nocturnal and diurnal TIR, and diabetes distress. The per-protocol (PP) group was defined as participants who used the IBC >30 times during the study. RESULTS: Mean TIR improved by 3.8% (95% CI 0.7–6.9%) from 69.2 to 73.0% (P = 0.017) in the PP group. This TIR corresponds to a mean of 0.9 more hours per day spent in range, and the improvement was driven by those with type 2 diabetes. There was no increase in measures of hypoglycemia or diabetes distress. Exploratory analysis revealed a reduction in measures of glucose variability. In addition, individuals with type 1 diabetes had greater improvements in diurnal TIR than in nocturnal TIR. CONCLUSION: A CGM-informed IBC app that applies trend arrow adjustments to bolus insulin dose recommendations improved TIR without increasing hypoglycemia or diabetes distress in individuals with type 1 or type 2 diabetes.
- 11 November 2024
- Published by American Diabetes Association in Diabetes Spectrum
- Vol. 38 (2) , 153-160
Safety of a Novel Continuous Glucose Monitoring–Informed Insulin Bolus Calculator Mobile Application for People With Type 1 or Type 2 Diabetes
BACKGROUND: Managing bolus insulin dosing can be a significant burden for people with diabetes, many of whom have limited numeracy skills. Insulin bolus calculators (IBCs) may improve glycemia as well as treatment satisfaction. OBJECTIVE: The purpose of this study was to demonstrate the safety of a novel, continuous glucose monitoring (CGM)-informed IBC mobile device app that applies trend arrow adjustments to bolus insulin dose recommendations. RESEARCH DESIGN AND METHODS: This clinical trial was an open-label, industry-sponsored single-arm study conducted at two sites. Fifty-four participants with type 1 or type 2 diabetes were enrolled and used the IBC app on their mobile device for 30 days. Study participants were adults who were already using CGM and dosing bolus insulin. The analysis examined both noninferiority and superiority of time in range (TIR) during the study period compared with baseline. Other important end points included hypoglycemia, glucose variability, nocturnal and diurnal TIR, and diabetes distress. The per-protocol (PP) group was defined as participants who used the IBC >30 times during the study. RESULTS: Mean TIR improved by 3.8% (95% CI 0.7–6.9%) from 69.2 to 73.0% (P = 0.017) in the PP group. This TIR corresponds to a mean of 0.9 more hours per day spent in range, and the improvement was driven by those with type 2 diabetes. There was no increase in measures of hypoglycemia or diabetes distress. Exploratory analysis revealed a reduction in measures of glucose variability. In addition, individuals with type 1 diabetes had greater improvements in diurnal TIR than in nocturnal TIR. CONCLUSION: A CGM-informed IBC app that applies trend arrow adjustments to bolus insulin dose recommendations improved TIR without increasing hypoglycemia or diabetes distress in individuals with type 1 or type 2 diabetes.
- Open Access
- 16 September 2024
- Published by Frontiers Media SA in Frontiers in Endocrinology
- Vol. 15, 1469471
The critical elements of digital health in diabetes and cardiometabolic care
Digital innovations provide novel opportunities to individualize a person’s care to best match their lifestyle needs and circumstances and to support them as they live their daily lives with diabetes. These innovations also serve to provide actionable data and insights for the care team giving them a “Webb telescope-like” view into their individual self-management journey, allowing them to see what cannot be seen during infrequent and limited office visits, thereby facilitating collaboration and communication to optimize the care plan on a timely basis. Technology advances are enabling diabetes care to transition from episodic, synchronous, primarily in-person care to include synchronous virtual care options and to continuous, on-demand, data-informed, asynchronous digital care better matching the demands of living with a relentless 24/7 chronic condition. In this paper we will discuss the critical elements and considerations in designing and implementing successful diabetes digital health tools in clinical practice.
- 16 September 2024
- Published by Frontiers Media SA in Frontiers in Endocrinology
- Vol. 15, 1469471
Open AccessThe critical elements of digital health in diabetes and cardiometabolic care
Digital innovations provide novel opportunities to individualize a person’s care to best match their lifestyle needs and circumstances and to support them as they live their daily lives with diabetes. These innovations also serve to provide actionable data and insights for the care team giving them a “Webb telescope-like” view into their individual self-management journey, allowing them to see what cannot be seen during infrequent and limited office visits, thereby facilitating collaboration and communication to optimize the care plan on a timely basis. Technology advances are enabling diabetes care to transition from episodic, synchronous, primarily in-person care to include synchronous virtual care options and to continuous, on-demand, data-informed, asynchronous digital care better matching the demands of living with a relentless 24/7 chronic condition. In this paper we will discuss the critical elements and considerations in designing and implementing successful diabetes digital health tools in clinical practice.
- 19 February 2024
- Published by SAGE Publications in Journal of Diabetes Science and Technology
- Vol. 19 (3) , 658-665
The Development and Potential Applications of an Automated Method for Detecting and Classifying Continuous Glucose Monitoring Patterns
Introduction: Continuous glucose monitoring (CGM) is emerging as a transformative tool for helping people with diabetes self-manage their glucose and supporting clinicians in effective treatment. Unfortunately, many CGM users, and clinicians, find interpreting the large volume of CGM data to be overwhelming and complex. To address this challenge, an efficient, intelligent method for detecting and classifying discernable patterns in CGM data was desired. Methods: We developed an automated artificial intelligence (AI)-driven method to detect and classify different discernable CGM patterns which called “CGM events.” We trained different models using 60 days of CGM data from 27 individuals with diabetes from a publicly available data set and then evaluated model performance using separate test data from the same group. Each event is classified according to clinical significance based on three parameters: (1) glucose category at or near the beginning of the CGM event; (2) a calculated severity score that encompasses both signal shape and temporal characteristics (e.g., how high the CGM curve goes (measured in mg/dL) and how long it stays above target (as established by published consensus guidelines); and (3) the glucose category at or near the end of the event. Results: The system accurately detected and classified events from actual CGM data. This was also validated with expert diabetes clinicians. Conclusions: Advanced pattern recognition methods can be used to detect and classify CGM events of interest and may be used to provide actionable insights and self-management support to CGM users and decision support to the clinicians caring for them.
- 19 February 2024
- Published by SAGE Publications in Journal of Diabetes Science and Technology
- Vol. 19 (3) , 658-665
The Development and Potential Applications of an Automated Method for Detecting and Classifying Continuous Glucose Monitoring Patterns
Introduction: Continuous glucose monitoring (CGM) is emerging as a transformative tool for helping people with diabetes self-manage their glucose and supporting clinicians in effective treatment. Unfortunately, many CGM users, and clinicians, find interpreting the large volume of CGM data to be overwhelming and complex. To address this challenge, an efficient, intelligent method for detecting and classifying discernable patterns in CGM data was desired. Methods: We developed an automated artificial intelligence (AI)-driven method to detect and classify different discernable CGM patterns which called “CGM events.” We trained different models using 60 days of CGM data from 27 individuals with diabetes from a publicly available data set and then evaluated model performance using separate test data from the same group. Each event is classified according to clinical significance based on three parameters: (1) glucose category at or near the beginning of the CGM event; (2) a calculated severity score that encompasses both signal shape and temporal characteristics (e.g., how high the CGM curve goes (measured in mg/dL) and how long it stays above target (as established by published consensus guidelines); and (3) the glucose category at or near the end of the event. Results: The system accurately detected and classified events from actual CGM data. This was also validated with expert diabetes clinicians. Conclusions: Advanced pattern recognition methods can be used to detect and classify CGM events of interest and may be used to provide actionable insights and self-management support to CGM users and decision support to the clinicians caring for them.
- Open Access
- 11 September 2023
- Published by JMIR Publications Inc. in JMIR Diabetes
- Vol. 8 (1) , e47638
Impact of a Combined Continuous Glucose Monitoring–Digital Health Solution on Glucose Metrics and Self-Management Behavior for Adults With Type 2 Diabetes: Real-World, Observational Study
Background: The BlueStar (Welldoc) digital health solution for people with diabetes incorporates data from multiple devices and generates coaching messages using artificial intelligence. The BlueStar app syncs glucose data from the G6 (Dexcom) real-time continuous glucose monitoring (RT-CGM) system, which provides a glucose measurement every 5 minutes. Objective: The objective of this real-world study of people with type 2 diabetes (T2D) using the digital health solution and RT-CGM was to evaluate change in glycemic control and engagement with the program over 3 months. Methods: Participants were current or former enrollees in an employer-sponsored health plan, were aged 18 years or older, had a T2D diagnosis, and were not using prandial insulin. Outcomes included CGM-based glycemic metrics and engagement with the BlueStar app, including logging medications taken, exercise, food details, blood pressure, weight, and hours of sleep. Results: Participants in the program that met our analysis criteria (n=52) were aged a mean of 53 (SD 9) years; 37% (19/52) were female and approximately 50% (25/52) were taking diabetes medications. The RT-CGM system was worn 90% (SD 8%) of the time over 3 months. Among individuals with suboptimal glycemic control at baseline, defined as mean glucose >180 mg/dL, clinically meaningful improvements in glycemic control were observed, including reductions in a glucose management indicator (–0.8 percentage points), time above range 181-250 mg/dL (–4.4 percentage points) and time above range >250 mg/dL (–14 percentage points; all P<.05). Time in range 70-180 mg/dL also increased by 15 percentage points (P=.016) in this population, which corresponds to an increase of approximately 3.5 hours per day in the target range. Over the 3-month study, 29% (15/52) of participants completed at least one engagement activity per week. Medication logging was completed most often by participants (23/52, 44%) at a rate of 12.1 (SD 0.8) events/week, and this was closely followed by exercise and food logging. Conclusions: The combination of an artificial intelligence–powered digital health solution and RT-CGM helped people with T2D improve their glycemic outcomes and diabetes self-management behaviors.
- 11 September 2023
- Published by JMIR Publications Inc. in JMIR Diabetes
- Vol. 8 (1) , e47638
Open AccessImpact of a Combined Continuous Glucose Monitoring–Digital Health Solution on Glucose Metrics and Self-Management Behavior for Adults With Type 2 Diabetes: Real-World, Observational Study
Background: The BlueStar (Welldoc) digital health solution for people with diabetes incorporates data from multiple devices and generates coaching messages using artificial intelligence. The BlueStar app syncs glucose data from the G6 (Dexcom) real-time continuous glucose monitoring (RT-CGM) system, which provides a glucose measurement every 5 minutes. Objective: The objective of this real-world study of people with type 2 diabetes (T2D) using the digital health solution and RT-CGM was to evaluate change in glycemic control and engagement with the program over 3 months. Methods: Participants were current or former enrollees in an employer-sponsored health plan, were aged 18 years or older, had a T2D diagnosis, and were not using prandial insulin. Outcomes included CGM-based glycemic metrics and engagement with the BlueStar app, including logging medications taken, exercise, food details, blood pressure, weight, and hours of sleep. Results: Participants in the program that met our analysis criteria (n=52) were aged a mean of 53 (SD 9) years; 37% (19/52) were female and approximately 50% (25/52) were taking diabetes medications. The RT-CGM system was worn 90% (SD 8%) of the time over 3 months. Among individuals with suboptimal glycemic control at baseline, defined as mean glucose >180 mg/dL, clinically meaningful improvements in glycemic control were observed, including reductions in a glucose management indicator (–0.8 percentage points), time above range 181-250 mg/dL (–4.4 percentage points) and time above range >250 mg/dL (–14 percentage points; all P<.05). Time in range 70-180 mg/dL also increased by 15 percentage points (P=.016) in this population, which corresponds to an increase of approximately 3.5 hours per day in the target range. Over the 3-month study, 29% (15/52) of participants completed at least one engagement activity per week. Medication logging was completed most often by participants (23/52, 44%) at a rate of 12.1 (SD 0.8) events/week, and this was closely followed by exercise and food logging. Conclusions: The combination of an artificial intelligence–powered digital health solution and RT-CGM helped people with T2D improve their glycemic outcomes and diabetes self-management behaviors.
- Open Access
- 28 March 2023
- Published by JMIR Publications Inc. in JMIR Preprints
Impact of a Combined Continuous Glucose Monitoring–Digital Health Solution on Glucose Metrics and Self-Management Behavior for Adults With Type 2 Diabetes: Real-World, Observational Study (Preprint)
BACKGROUND The BlueStar (Welldoc) digital health solution for people with diabetes incorporates data from multiple devices and generates coaching messages using artificial intelligence. The BlueStar app syncs glucose data from the G6 (Dexcom) real-time continuous glucose monitoring (RT-CGM) system, which provides a glucose measurement every 5 minutes. OBJECTIVE The objective of this real-world study of people with type 2 diabetes (T2D) using the digital health solution and RT-CGM was to evaluate change in glycemic control and engagement with the program over 3 months. METHODS Participants were current or former enrollees in an employer-sponsored health plan, were aged 18 years or older, had a T2D diagnosis, and were not using prandial insulin. Outcomes included CGM-based glycemic metrics and engagement with the BlueStar app, including logging medications taken, exercise, food details, blood pressure, weight, and hours of sleep. RESULTS Participants in the program that met our analysis criteria (n=52) were aged a mean of 53 (SD 9) years; 37% (19/52) were female and approximately 50% (25/52) were taking diabetes medications. The RT-CGM system was worn 90% (SD 8%) of the time over 3 months. Among individuals with suboptimal glycemic control at baseline, defined as mean glucose >180 mg/dL, clinically meaningful improvements in glycemic control were observed, including reductions in a glucose management indicator (–0.8 percentage points), time above range 181-250 mg/dL (–4.4 percentage points) and time above range >250 mg/dL (–14 percentage points; all P<.05). Time in range 70-180 mg/dL also increased by 15 percentage points (P=.016) in this population, which corresponds to an increase of approximately 3.5 hours per day in the target range. Over the 3-month study, 29% (15/52) of participants completed at least one engagement activity per week. Medication logging was completed most often by participants (23/52, 44%) at a rate of 12.1 (SD 0.8) events/week, and this was closely followed by exercise and food logging. CONCLUSIONS The combination of an artificial intelligence–powered digital health solution and RT-CGM helped people with T2D improve their glycemic outcomes and diabetes self-management behaviors.
- 28 March 2023
- Published by JMIR Publications Inc. in JMIR Preprints
Open AccessImpact of a Combined Continuous Glucose Monitoring–Digital Health Solution on Glucose Metrics and Self-Management Behavior for Adults With Type 2 Diabetes: Real-World, Observational Study (Preprint)
BACKGROUND The BlueStar (Welldoc) digital health solution for people with diabetes incorporates data from multiple devices and generates coaching messages using artificial intelligence. The BlueStar app syncs glucose data from the G6 (Dexcom) real-time continuous glucose monitoring (RT-CGM) system, which provides a glucose measurement every 5 minutes. OBJECTIVE The objective of this real-world study of people with type 2 diabetes (T2D) using the digital health solution and RT-CGM was to evaluate change in glycemic control and engagement with the program over 3 months. METHODS Participants were current or former enrollees in an employer-sponsored health plan, were aged 18 years or older, had a T2D diagnosis, and were not using prandial insulin. Outcomes included CGM-based glycemic metrics and engagement with the BlueStar app, including logging medications taken, exercise, food details, blood pressure, weight, and hours of sleep. RESULTS Participants in the program that met our analysis criteria (n=52) were aged a mean of 53 (SD 9) years; 37% (19/52) were female and approximately 50% (25/52) were taking diabetes medications. The RT-CGM system was worn 90% (SD 8%) of the time over 3 months. Among individuals with suboptimal glycemic control at baseline, defined as mean glucose >180 mg/dL, clinically meaningful improvements in glycemic control were observed, including reductions in a glucose management indicator (–0.8 percentage points), time above range 181-250 mg/dL (–4.4 percentage points) and time above range >250 mg/dL (–14 percentage points; all P<.05). Time in range 70-180 mg/dL also increased by 15 percentage points (P=.016) in this population, which corresponds to an increase of approximately 3.5 hours per day in the target range. Over the 3-month study, 29% (15/52) of participants completed at least one engagement activity per week. Medication logging was completed most often by participants (23/52, 44%) at a rate of 12.1 (SD 0.8) events/week, and this was closely followed by exercise and food logging. CONCLUSIONS The combination of an artificial intelligence–powered digital health solution and RT-CGM helped people with T2D improve their glycemic outcomes and diabetes self-management behaviors.
- Open Access
- 14 October 2014
- Published by SAGE Publications in Journal of Diabetes Science and Technology
- Vol. 9 (1) , 86-90
Hypoglycemia Prediction Using Machine Learning Models for Patients With Type 2 Diabetes
Background: Minimizing the occurrence of hypoglycemia in patients with type 2 diabetes is a challenging task since these patients typically check only 1 to 2 self-monitored blood glucose (SMBG) readings per day. Method: We trained a probabilistic model using machine learning algorithms and SMBG values from real patients. Hypoglycemia was defined as a SMBG value < 70 mg/dL. We validated our model using multiple data sets. In addition, we trained a second model, which used patient SMBG values and information about patient medication administration. Results: The optimal number of SMBG values needed by the model was approximately 10 per week. The sensitivity of the model for predicting a hypoglycemia event in the next 24 hours was 92% and the specificity was 70%. In the model that incorporated medication information, the prediction window was for the hour of hypoglycemia, and the specificity improved to 90%. Conclusions: Our machine learning models can predict hypoglycemia events with a high degree of sensitivity and specificity. These models—which have been validated retrospectively and if implemented in real time—could be useful tools for reducing hypoglycemia in vulnerable patients.
- 14 October 2014
- Published by SAGE Publications in Journal of Diabetes Science and Technology
- Vol. 9 (1) , 86-90
Open AccessHypoglycemia Prediction Using Machine Learning Models for Patients With Type 2 Diabetes
Background: Minimizing the occurrence of hypoglycemia in patients with type 2 diabetes is a challenging task since these patients typically check only 1 to 2 self-monitored blood glucose (SMBG) readings per day. Method: We trained a probabilistic model using machine learning algorithms and SMBG values from real patients. Hypoglycemia was defined as a SMBG value < 70 mg/dL. We validated our model using multiple data sets. In addition, we trained a second model, which used patient SMBG values and information about patient medication administration. Results: The optimal number of SMBG values needed by the model was approximately 10 per week. The sensitivity of the model for predicting a hypoglycemia event in the next 24 hours was 92% and the specificity was 70%. In the model that incorporated medication information, the prediction window was for the hour of hypoglycemia, and the specificity improved to 90%. Conclusions: Our machine learning models can predict hypoglycemia events with a high degree of sensitivity and specificity. These models—which have been validated retrospectively and if implemented in real time—could be useful tools for reducing hypoglycemia in vulnerable patients.