Transcript Slide 1

Hierarchical Real-Time Filtering for Continuous Glucose
Sensor Data
J. G. Chase, X. Chen, H. Sirisena, G. Shaw, X. W. Wong, C. E. Hann, A. Le Compte, J. Lin, T. Lotz
INTRODUCTION
RESULTS
• The Solution: Point-of-care (POC) continuous glucose
sensors offer significant promise for real-time control and
artificial pancreas systems in general diabetes.
• The Problem with the Solution: Outlying errors up to 40%
with standard deviations of 10-17%1,2.
• The Actual Problem: Large errors inhibit model parameter
identification for real-time model-based control3,4. Such errors
can span several different clinical action ranges, especially if
glucose is tightly controlled to 75-140 mg/dL.
• A Solution to the Actual Problem with “The Solution”: An
array of filters that reduces mean absolute errors to less than
5% with minimal lag would enhance the potential for real-time
control using these types of continuous glucose sensors.
Moving Average of 2
Median Filter Results
30 Minutes Moving
Median
5% ~ 95% MAPE
1
1.6
[0.1 3.7]
2
1.4
[0.1 3.4]
3
1.3
[0.1 3.4]
4
1.4
[0.1 3.7]
5
1.4
[0.1 3.7]
6
1.4
[0.1 3.7]
7
1.4
[0.1 3.6]
8
1.3
[0.1 3.7]
9
1.5
[0.1 3.9]
10
1.5
[0.1 4.0]
1.4
[0.1 3.7]
1
2.8
[0.2 8.3]
2
3.4
[0.2 9.3]
3
3.4
[0.2 8.4]
4
3.2
[0.2 8.8]
5
2.8
[0.2 8.9]
6
3.0
[0.2 7.8]
7
3.2
[0.2 9.4]
8
3.1
[0.2 7.7]
9
2.9
[0.2 8.9]
10
2.6
[0.2 8.7]
Overall B
2.6
[0.2 8.4]
All Patients (A & B)
1.8
[0.1 5.4]
Overall A
Retrospective
Patients
SPRINT Patient Blood Glucose Profile
12
Blood Glucose Level(mmol/L)
30 Minutes Moving LMS
filter and use first 2/3 filter
results to avoid polynomial
peak at the edge
Mean Absolute % Error
(MAPE)
SPRINT Patients
METHOD: Filter Array Block Diagram
10 Minutes Moving
Median
PATIENT
True Blood Glucose
Noisy Blood Glucose
Filtered Blood Glucose
10
8
6
4
Retrospective Patient Blood Glucose Profile
2
14
True Blood Glucose
Noisy Blood Glucose
Filtered Blood Glucose
Clarke Error Grid Plot for 20 Patients
1000
2000
3000
4000
5000
6000
Time(Minutes)
20
A
E
C
15
Filtered
CGM
B
10
D
D
B
5
C
0
0
5
10
E
15
20
25
Noisy Blood Glucose Level(mmol/L)
• Filter removes outliers
• Outliers and variability
are too big for useful
model-based control
7000
12
8000
10
8
6
4
2
0
• Parallel non-linear median filters remove outliers using 3 and 7
consecutive measurements. The output is given to a least mean
squares (LMS) filter to further smooth the result.
• Monte Carlo simulations (n=10) using glucose data from 20 critical
care patients with a realistic continuous sensor error model.
• Error model has 20% outliers over 20% error with overall mean
absolute errors of 9-10%1-3.
• Data: 2410 hours (mean: 120.5 hours/patient, range: 34-502 hours).
• Mean absolute percentage error (MAPE) and its standard deviation
are reported after filtering, both overall and for each patient.
REFERENCES
[1] P. A. Goldberg, M. D. Siegel, R. R. Russell, R. S. Sherwin, J. I. Halickman, D. A. Cooper, J. D. Dziura, and S. E. Inzucchi,
"Experience with the continuous glucose monitoring system in a medical intensive care unit," /Diabetes Technol Ther/, vol.
6, pp. 339-47, 2004.
[2] B. Guerci, M. Floriot, P. Bohme, D. Durain, M. Benichou, S. Jellimann, and P. Drouin, "Clinical performance of CGMS in
type 1 diabetic patients treated by continuous subcutaneous insulin infusion using insulin analogs," Diabetes care, vol. 26,
pp. 582-9, 2003.
[3] J. G. Chase, G. M. Shaw, X. W. Wong, T. Lotz, J. Lin, and C. E. Hann, "Model-based Glycaemic Control in Critical Care - A
review of the state of the possible," /Biomedical Signal Processing & Control/, vol. 1, pp. 3-21, 2006.
[4] J. G. Chase, C. E. Hann, M. Jackson, J. Lin, T. Lotz, X. W. Wong, and G. M. Shaw, "Integral-based filtering of continuous
glucose sensor measurements for glycaemic control in critical care," Comput Methods Programs Biomed, vol. 82, pp.
238-47, 2006
9000
0
1000
2000
3000
4000
5000
6000
7000
8000
Time(Minutes)
Random Noise Model Results for 1 Patient
(7200 mins data)
1500
• Noise has ~80% in A-band
• MAPE
• Estimated for filter design
~10%1-3
%Outliers = 24% outside
A band
Number of Points (of 7200)
True Blood Glucose Level (mmol/L)
Unfiltered
CGM
0
Blood Glucose Level(mmol/L)
0
25
MAPE = 9.8%
1000
500
0
-0.5
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
Decimal Percentage Error
0.3
0.4
0.5
CONCLUSIONS
• Continuous glucose data with significant outliers can be
effectively filtered using a hierarchy of non-linear median
filters and smoothing via LMS or splines.
• More volatile data is more sensitive to error.
• Sensor lag is approximately 10 minutes.
• Future Work: this approach does not account for
calibration drift or bias – coming soon!