549830
research-article2014
DSTXXX10.1177/1932296814549830Journal of Diabetes Science and TechnologyCichosz et al
Original Article
Combining Information of Autonomic Modulation and CGM Measurements Enables Prediction and Improves Detection of Spontaneous Hypoglycemic Events
Journal of Diabetes Science and Technology 1–6 © 2014 Diabetes Technology Society Reprints and permissions: sagepub.com/journalsPermissions.nav DOI: 10.1177/1932296814549830 dst.sagepub.com
Simon Lebech Cichosz, MSc1,2, Jan Frystyk, MD, PhD, DMSc1, Lise Tarnow, MD, DMSc3, and Jesper Fleischer, MSc, BME, PhD1
Abstract We have previously tested, in a laboratory setting, a novel algorithm that enables prediction of hypoglycemia. The algorithm integrates information of autonomic modulation, based on heart rate variability (HRV), and data based on a continuous glucose monitoring (CGM) device. Now, we investigate whether the algorithm is suitable for prediction of hypoglycemia and for improvement of hypoglycemic detection during normal daily activities. Twenty-one adults (13 men) with T1D prone to hypoglycemia were recruited and monitored with CGM and a Holter device while they performed normal daily activities. We used our developed algorithm (a pattern classification method) to predict spontaneous hypoglycemia based on CGM and HRV. We compared 3 different models; (i) a model containing raw data from the CGM device; (ii) a CGM* model containing data derived from the CGM device signal; and (iii) a CGM+HRV model-combining model (ii) with HRV data. A total of 12 hypoglycemic events (glucose levels < 3.9 mmol/L, 70 mg/dL) and 237 euglycemic measurements were included. For a 20-minute prediction, model (i) resulted in a ROC AUC of 0.69. If a high sensitivity of 100% was chosen, the corresponding specificity was 69%. (ii) The CGM* model yielded a ROC AUC of 0.92 with a corresponding sensitivity of 100% and specificity of 71%. (iii) The CGM+HRV model yielded a ROC AUC of 0.96 with a corresponding sensitivity of 100% and specificity of 91%. Data shows that adding information of autonomic modulation to CGM measurements enables prediction and improves the detection of hypoglycemia. Keywords continuous glucose monitoring, heart rate variability, hypoglycemia, diabetes, prediction Were it not for the risk of hypoglycemic episodes, patients with diabetes could have normal glucose levels over a lifetime of diabetes. Hypoglycemia often results in an increase physical as well as psychosocial morbidity, and is a risk factor for an increased mortality.1,2 Hypoglycemia is very common in patients with type 1 diabetes (T1D).3 Patients trying to improve or maintain a tight glycemic control suffer from innumerable episodes of asymptomatic hypoglycemia. Plasma glucose levels may be less than 60 mg/dl (3.3 mmol/l) 10% of the time, and on average, patients with T1D suffer from 2 weekly incidents of symptomatic hypoglycemia.1,4,5 Accordingly, patients with diabetes experience thousands of hypoglycemic events over a lifetime. In addition, these patients have a 4.7-fold excess mortality risk compared to healthy subjects.6 This explains why there is a considerable interest in using continuous glucose monitoring (CGM) devices to detect and warn diabetic patients about an
imminent hypoglycemic event.7-10 However, a proportion of false positive alarms must be tolerated. Koivikko et al studied 37 adults with T1D, who underwent CGM and ECG monitoring for 3 nights. The authors observed that spontaneous nocturnal hypoglycemia resulted in a reduction of the low-frequency component of the heart 1
Department of Endocrinology and Internal Medicine and Medical Research Laboratory, Aarhus University Hospital, Denmark 2 Department of Health Science and Technology, Aalborg University, Denmark 3 Steno Diabetes Center, Department of Clinical Epidemiology, Aarhus University and Nordsjaellands Hospitaler Hilleroed, Denmark Corresponding Author: Simon Lebech Cichosz, Aarhus University Hospital, Department of Endocrinology and Internal Medicine and Medical Research Laboratory, Norrebrogade 44, Building 3, 8000 Aarhus C, Denmark. Email: [email protected]
Downloaded from dst.sagepub.com at GEORGIAN COURT UNIV on February 22, 2015
2
Journal of Diabetes Science and Technology
rate variability (HRV).11 HRV, that is, the beat-to-beat fluctuations in the heart rate, is a noninvasive and extensively validated method to investigate the balance in the autonomic nervous system.12,13 Recently, we showed how a combination of CGM and HRV could be used to predict and detect hypoglycemia in bed-bound patients with T1D.14 Our data suggested that HRV could help improving specificity and thereby reduce false positive alarms. However, questions remain on how useful HRV information is during daily life living and during spontaneous hypoglycemia. HRV is influenced by many factors such as physical activity, diurnal variation, age and gender.15,16 Also, glucose excursions are much more diverse and complex under free-living conditions.17 Therefore, with the present study we aimed to investigate whether the developed algorithm14 was able to predict and improve the detection of hypoglycemia during normal daily activities.
Methods Participants Data for this study were obtained from a trial performed at Steno Diabetes Center (Gentofte, Denmark).18 A total of 21 (13 men and 8 women) adults with long lasting T1D were recruited. The patients were 58 ± 10 years old, had a diabetes duration of 34 ± 12 years and a HbA1c 7.9 ± 0.7%, and 11 participants had peripheral neuropathy measured by biothesiometer. All participants were prone to hypoglycemia, that is, they had experienced at least 2 episodes of severe hypoglycemia within the last year. None of the patients had a history of cardiovascular disease or were taking drugs affecting the cardiovascular system. All patients had a normal electrocardiogram. The study protocols were approved by the local ethics committee and the study conducted according to the principles of the Helsinki Declaration II. All patients gave their written informed consent.
Study Design Participants were admitted to the Clinical Research Unit at the Steno Diabetes Center Thursday at 8 pm. ECG was measured from lead II using a digital Holter monitor (SpiderView Plus, ELA Medical, Montrouge, France). At the same time, CGM was monitored using a Guardian Real-Time Continuous Glucose Monitoring System (Medtronic MiniMed, Northridge, CA, USA) with the prevailing glucose level blinded. At 11 pm a cannula was placed into an antecubital arm vein. Blood glucose samples were taken at hourly intervals until 7 am the next morning. At 8 am, participants were sent home with the monitoring equipment and were instructed to calibrate the CGM at least 4 times a day. Monitoring ended on Sunday at 8 pm. A total of 72 hours of continuous CGM and ECG data were available for each participant.
Data Processing The ECG was analyzed using custom analysis software developed in MatLab (Version R2014a; MathWorks, Natick, MA, USA). ECG QRS detection was implemented based on the methods of Pan and Tompkins with (1) bandpass filter, (2) differentiating, (3) squaring, and (4) movingwindow integration.19 Initial R-peaks were identified with a threshold and a minimum time distance of 250 ms from the moving-window integration output. R detections were then found as the highest point in the original signal within the timeframe of the initial detected peak. Interbeat intervals were derived from the R detections and interpolation was used to remove outliers based on 2.StdHRV. The filtered HRV signal was inspected manually and periods with substantial noisy signals were labeled for excluding glucose measurements with appertaining corrupted HRV. The HRV signal was analyzed with a 5 minutes, 90% overlapping sliding window calculating typical derived measures describing HRV; heart rate, SDNN (standard deviation of all NN intervals), SDANN (standard deviation of the averages of NN intervals in all 5-minute segments of the entire recording), pNNx (proportion of pairs of adjacent NN intervals differing by more than 50 ms), RMSSD (the square root of the mean of the sum of the squares of differences between adjacent NN intervals), VLF (power in very low frequency range,
research-article2014
DSTXXX10.1177/1932296814549830Journal of Diabetes Science and TechnologyCichosz et al
Original Article
Combining Information of Autonomic Modulation and CGM Measurements Enables Prediction and Improves Detection of Spontaneous Hypoglycemic Events
Journal of Diabetes Science and Technology 1–6 © 2014 Diabetes Technology Society Reprints and permissions: sagepub.com/journalsPermissions.nav DOI: 10.1177/1932296814549830 dst.sagepub.com
Simon Lebech Cichosz, MSc1,2, Jan Frystyk, MD, PhD, DMSc1, Lise Tarnow, MD, DMSc3, and Jesper Fleischer, MSc, BME, PhD1
Abstract We have previously tested, in a laboratory setting, a novel algorithm that enables prediction of hypoglycemia. The algorithm integrates information of autonomic modulation, based on heart rate variability (HRV), and data based on a continuous glucose monitoring (CGM) device. Now, we investigate whether the algorithm is suitable for prediction of hypoglycemia and for improvement of hypoglycemic detection during normal daily activities. Twenty-one adults (13 men) with T1D prone to hypoglycemia were recruited and monitored with CGM and a Holter device while they performed normal daily activities. We used our developed algorithm (a pattern classification method) to predict spontaneous hypoglycemia based on CGM and HRV. We compared 3 different models; (i) a model containing raw data from the CGM device; (ii) a CGM* model containing data derived from the CGM device signal; and (iii) a CGM+HRV model-combining model (ii) with HRV data. A total of 12 hypoglycemic events (glucose levels < 3.9 mmol/L, 70 mg/dL) and 237 euglycemic measurements were included. For a 20-minute prediction, model (i) resulted in a ROC AUC of 0.69. If a high sensitivity of 100% was chosen, the corresponding specificity was 69%. (ii) The CGM* model yielded a ROC AUC of 0.92 with a corresponding sensitivity of 100% and specificity of 71%. (iii) The CGM+HRV model yielded a ROC AUC of 0.96 with a corresponding sensitivity of 100% and specificity of 91%. Data shows that adding information of autonomic modulation to CGM measurements enables prediction and improves the detection of hypoglycemia. Keywords continuous glucose monitoring, heart rate variability, hypoglycemia, diabetes, prediction Were it not for the risk of hypoglycemic episodes, patients with diabetes could have normal glucose levels over a lifetime of diabetes. Hypoglycemia often results in an increase physical as well as psychosocial morbidity, and is a risk factor for an increased mortality.1,2 Hypoglycemia is very common in patients with type 1 diabetes (T1D).3 Patients trying to improve or maintain a tight glycemic control suffer from innumerable episodes of asymptomatic hypoglycemia. Plasma glucose levels may be less than 60 mg/dl (3.3 mmol/l) 10% of the time, and on average, patients with T1D suffer from 2 weekly incidents of symptomatic hypoglycemia.1,4,5 Accordingly, patients with diabetes experience thousands of hypoglycemic events over a lifetime. In addition, these patients have a 4.7-fold excess mortality risk compared to healthy subjects.6 This explains why there is a considerable interest in using continuous glucose monitoring (CGM) devices to detect and warn diabetic patients about an
imminent hypoglycemic event.7-10 However, a proportion of false positive alarms must be tolerated. Koivikko et al studied 37 adults with T1D, who underwent CGM and ECG monitoring for 3 nights. The authors observed that spontaneous nocturnal hypoglycemia resulted in a reduction of the low-frequency component of the heart 1
Department of Endocrinology and Internal Medicine and Medical Research Laboratory, Aarhus University Hospital, Denmark 2 Department of Health Science and Technology, Aalborg University, Denmark 3 Steno Diabetes Center, Department of Clinical Epidemiology, Aarhus University and Nordsjaellands Hospitaler Hilleroed, Denmark Corresponding Author: Simon Lebech Cichosz, Aarhus University Hospital, Department of Endocrinology and Internal Medicine and Medical Research Laboratory, Norrebrogade 44, Building 3, 8000 Aarhus C, Denmark. Email: [email protected]
Downloaded from dst.sagepub.com at GEORGIAN COURT UNIV on February 22, 2015
2
Journal of Diabetes Science and Technology
rate variability (HRV).11 HRV, that is, the beat-to-beat fluctuations in the heart rate, is a noninvasive and extensively validated method to investigate the balance in the autonomic nervous system.12,13 Recently, we showed how a combination of CGM and HRV could be used to predict and detect hypoglycemia in bed-bound patients with T1D.14 Our data suggested that HRV could help improving specificity and thereby reduce false positive alarms. However, questions remain on how useful HRV information is during daily life living and during spontaneous hypoglycemia. HRV is influenced by many factors such as physical activity, diurnal variation, age and gender.15,16 Also, glucose excursions are much more diverse and complex under free-living conditions.17 Therefore, with the present study we aimed to investigate whether the developed algorithm14 was able to predict and improve the detection of hypoglycemia during normal daily activities.
Methods Participants Data for this study were obtained from a trial performed at Steno Diabetes Center (Gentofte, Denmark).18 A total of 21 (13 men and 8 women) adults with long lasting T1D were recruited. The patients were 58 ± 10 years old, had a diabetes duration of 34 ± 12 years and a HbA1c 7.9 ± 0.7%, and 11 participants had peripheral neuropathy measured by biothesiometer. All participants were prone to hypoglycemia, that is, they had experienced at least 2 episodes of severe hypoglycemia within the last year. None of the patients had a history of cardiovascular disease or were taking drugs affecting the cardiovascular system. All patients had a normal electrocardiogram. The study protocols were approved by the local ethics committee and the study conducted according to the principles of the Helsinki Declaration II. All patients gave their written informed consent.
Study Design Participants were admitted to the Clinical Research Unit at the Steno Diabetes Center Thursday at 8 pm. ECG was measured from lead II using a digital Holter monitor (SpiderView Plus, ELA Medical, Montrouge, France). At the same time, CGM was monitored using a Guardian Real-Time Continuous Glucose Monitoring System (Medtronic MiniMed, Northridge, CA, USA) with the prevailing glucose level blinded. At 11 pm a cannula was placed into an antecubital arm vein. Blood glucose samples were taken at hourly intervals until 7 am the next morning. At 8 am, participants were sent home with the monitoring equipment and were instructed to calibrate the CGM at least 4 times a day. Monitoring ended on Sunday at 8 pm. A total of 72 hours of continuous CGM and ECG data were available for each participant.
Data Processing The ECG was analyzed using custom analysis software developed in MatLab (Version R2014a; MathWorks, Natick, MA, USA). ECG QRS detection was implemented based on the methods of Pan and Tompkins with (1) bandpass filter, (2) differentiating, (3) squaring, and (4) movingwindow integration.19 Initial R-peaks were identified with a threshold and a minimum time distance of 250 ms from the moving-window integration output. R detections were then found as the highest point in the original signal within the timeframe of the initial detected peak. Interbeat intervals were derived from the R detections and interpolation was used to remove outliers based on 2.StdHRV. The filtered HRV signal was inspected manually and periods with substantial noisy signals were labeled for excluding glucose measurements with appertaining corrupted HRV. The HRV signal was analyzed with a 5 minutes, 90% overlapping sliding window calculating typical derived measures describing HRV; heart rate, SDNN (standard deviation of all NN intervals), SDANN (standard deviation of the averages of NN intervals in all 5-minute segments of the entire recording), pNNx (proportion of pairs of adjacent NN intervals differing by more than 50 ms), RMSSD (the square root of the mean of the sum of the squares of differences between adjacent NN intervals), VLF (power in very low frequency range,
