Artificial intelligence to improve blood glucose control for people with type 1 diabetes
Authors
Navarathna, Pranesh
ORCID
Loading...
Other Contributors
Bequette, B. Wayne
Hahn, Juergen
Bennett, Kristin P.
Cramer, Steven
Hahn, Juergen
Bennett, Kristin P.
Cramer, Steven
Issue Date
2020-12
Keywords
Chemical engineering
Degree
PhD
Terms of Use
Attribution-NonCommercial-NoDerivs 3.0 United States
This electronic version is a licensed copy owned by Rensselaer Polytechnic Institute, Troy, NY. Copyright of original work retained by author.
This electronic version is a licensed copy owned by Rensselaer Polytechnic Institute, Troy, NY. Copyright of original work retained by author.
Full Citation
Abstract
Automatic activity detections and forecasts can allow a controller to prompt for meals and anticipate exercise. To this end, an automatic activity detection framework using a smartwatch and smartphone has been developed. Long Short-Term Memory (LSTM) networks, convolutional neural networks (CNN), and lookup tables are used to create minute-basedprobabilities of other, sleep, eating, exercise, and null using sensor data. A particle filter is used to detect activity events by aggregating minute-based probabilities. Population level time of day prior probabilities are used to forecast exercise and sleep. Using smartwatch sensor readings, the CNN achieves a class-weighted accuracy of 85% when personalized, versus 73% when unpersonalized. When personalized, the particle lter detects 94% of sleep, 98% of eating, and 97% of exercise events after 24 hours, while detecting 0.29 false sleep events/day, 3.6 false meal events/day and 0.7 false exercise event/day. When unpersonalized, the particle lter detects 93% of sleep, 97% of eating, and 90% of exercise events after 24 hours, while detecting 0.39 false sleep events/day, 4.1 false meal events/day and 1.2 false exercise event/day. The trade-off between using multiple watch sensors versus using a single watch sensor is also assessed. Time of day probabilities provide higher likelihoods of impending exercise and sleep when they truly occur compared to when they do not. Reliable activity detections, and forecasts of exercise and sleep can be achieved with daily patient
review.
Signals from the CGM and insulin pump can therefore be supplemented with activity detections/forecasts such that a controller may account for the activity of an individual with type 1 diabetes. The machine learning based PISA detection approach improves on the previous rules-based approach, and may be used to inform a controller when CGM readingscan be trusted. This algorithm can also be used to retrospectively analyze CGM data to understand one's susceptibility to PISA events. It is the hope that the techniques presented in this thesis can reduce patient burden and allow people with T1D to lead more normal lives.
For PISA detection, machine learning algorithms are compared on their ability to improve on a previously studied rules-based algorithm. Support vector machines, decision trees, neural networks, random forests, and gradient boosted machines are evaluated on data from an outpatient trial. A classification strategy for separately classifying the start of aPISA event and subsequent PISA CGM readings is outlined. The gradient boosted machine based approach has an event-based true positive rate of 92.9% while detecting 0.67 false positive PISA events per night. The machine learning model is interpreted by average PISA event analysis, false positive distribution estimation, permutation feature importance, and surrogate tree analysis. The gradient boosted machine based classifier is then evaluated on real-world data from 100 data donors. PISA detections are analyzed across different age groups, years since diagnosis, and between males and females. It is seen that the classifier behaves similarly on real-world data compared to trial data.
An individual with type 1 diabetes (T1D) must be vigilant in monitoring their blood glucose values while considering the impact of meals, exercise, and other critical events. More and more people with T1D are wearing continuous glucose monitors (CGM, providing blood glucose values at 5 min intervals) and insulin pumps that continuously deliver rapid-acting insulin. However, only a few commercial closed-loop systems, that automatically adjust insulin in response to glucose changes, are currently available. These systems require manual meal announcements to provide insulin boluses at mealtime (feedforward control) and automatically suspend insulin when glucose is low. Still, missed meal announcements and exercise-induced hypoglycemia are major causes for suboptimal glucose control in people with T1D. Another common problem is CGM pressure-induced sensor attenuation (PISAs) that typically occur overnight when a user sleeps on their CGM, resulting in inaccurate CGM signals. PISAs can cause the pump to shut off unnecessarily. Evaluating the likelihood of CGM readings being PISAs can allow a controller to ignore readings with a high PISA probability.
Signals from the CGM and insulin pump can therefore be supplemented with activity detections/forecasts such that a controller may account for the activity of an individual with type 1 diabetes. The machine learning based PISA detection approach improves on the previous rules-based approach, and may be used to inform a controller when CGM readingscan be trusted. This algorithm can also be used to retrospectively analyze CGM data to understand one's susceptibility to PISA events. It is the hope that the techniques presented in this thesis can reduce patient burden and allow people with T1D to lead more normal lives.
For PISA detection, machine learning algorithms are compared on their ability to improve on a previously studied rules-based algorithm. Support vector machines, decision trees, neural networks, random forests, and gradient boosted machines are evaluated on data from an outpatient trial. A classification strategy for separately classifying the start of aPISA event and subsequent PISA CGM readings is outlined. The gradient boosted machine based approach has an event-based true positive rate of 92.9% while detecting 0.67 false positive PISA events per night. The machine learning model is interpreted by average PISA event analysis, false positive distribution estimation, permutation feature importance, and surrogate tree analysis. The gradient boosted machine based classifier is then evaluated on real-world data from 100 data donors. PISA detections are analyzed across different age groups, years since diagnosis, and between males and females. It is seen that the classifier behaves similarly on real-world data compared to trial data.
An individual with type 1 diabetes (T1D) must be vigilant in monitoring their blood glucose values while considering the impact of meals, exercise, and other critical events. More and more people with T1D are wearing continuous glucose monitors (CGM, providing blood glucose values at 5 min intervals) and insulin pumps that continuously deliver rapid-acting insulin. However, only a few commercial closed-loop systems, that automatically adjust insulin in response to glucose changes, are currently available. These systems require manual meal announcements to provide insulin boluses at mealtime (feedforward control) and automatically suspend insulin when glucose is low. Still, missed meal announcements and exercise-induced hypoglycemia are major causes for suboptimal glucose control in people with T1D. Another common problem is CGM pressure-induced sensor attenuation (PISAs) that typically occur overnight when a user sleeps on their CGM, resulting in inaccurate CGM signals. PISAs can cause the pump to shut off unnecessarily. Evaluating the likelihood of CGM readings being PISAs can allow a controller to ignore readings with a high PISA probability.
Description
December 2020
School of Engineering
School of Engineering
Department
Dept. of Chemical and Biological Engineering
Publisher
Rensselaer Polytechnic Institute, Troy, NY
Relationships
Rensselaer Theses and Dissertations Online Collection
Access
CC BY-NC-ND. Users may download and share copies with attribution in accordance with a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 License. No commercial use or derivatives are permitted without the explicit approval of the author.