1. Ayan Banerjee and Sandeep K.S. Gupta
Clinical Evaluation of Generative Model Based Monitoring and Comparison with Compressive Sensing
Ayan Banerjee and Sandeep K.S. Gupta
- Funded by NIH NIBIB R21 EB019202

2. Long term Cardiac Monitoring
Continuous Long Term Monitoring
Behavior monitoring
Physiological signal monitoring
Actuation or feedback
Fitbit - Activity
Zeo - Sleep
iMec - ECG
Glucose
Infusion pump
Ventilators
Emotiv - EEG
Ethlife - Stress
Sports
Social
Neuro-feedback
Problems:
High sampling rate
For ECG Hz – 512 Hz (3 leads)
High memory requirements
2 GB SD card consumed in 8 days
High bandwidth requirements
82 kbps
Pervasiveness, ambulatory,
Clinical motivation, quantify properties related to their health, wearable computing, pervasive health monitor not only ECG, but also brain, movement, quantified self.
From ICU to home, power (Holter monitor only 48 hours) bulky and inconvenient
Renewed focus on data compression and resource efficiency
Long term Cardiac Monitoring
ICU
At Home
Preterm Infants

3. Two competing technologies
Compressive Sensing
Aim:
Reduction of sensing needs
Source of compression
Sparsity in some linear domain
Characteristics
Generic simple sensors
Complex recovery algorithms
Recovery Hypothesis
Point by point signal comparison
Outcome:
Increased battery life, reduce bandwidth, reduce storage
Reduce sensing power
Generative Model Based Monitoring
Aim:
Reduction of communication
Source of compression
Periodic shape properties
Characteristics
Individualized complex sensors
Simple recovery algorithm
Recovery hypothesis
Diagnostic Equivalence
Outcome:
Increased battery life, reduced bandwidth and storage
Automated annotation
Which parameters are reducing effort for sensor and base station?
System level differences between GemREM and cs

4. Generative Model Based Monitoring (GeMREM)
Why does it work?:
A doctor’s Perspective
Doctor’s perspective
Learn model install model
Sparsity in model
Diagnostic feature based signal recovery
Signals dont have to match point by point
Shape is important for diagnosis
Temporal parameters have high error margin
GeMREM supports Diagnostic Feature based signal recovery

5. Monitoring Regimen Study Device Patient population Monitoring duration
Shimmer 2R wearable sensors
Nexus One Smartphone
Holter monitor for raw data
Patient population
25 patients in ICU (limited ambulation)
14 men, 11 women
Monitoring duration
24 hour monitoring
Data collection method
Model learning phase
Sensors and smartphone programmed with model
Start continuous monitoring
GeMREM and Raw signal deployment
Compressive sensing tested through simulation

6. Monitoring artifacts Stitching artifact
False Premature Atrial Complex (PAC)
Recovery algorithm should avoid stitching
Mobility and Bluetooth interference
Presence of medical equipment with magnetic properties can disrupt communication
Motion artifacts
Limited ambulation
Device uninstalled during procedures such as MRI

7. Temporal Parameters Error in heart rate2 4 6 8 10 12 14 16 5 15 20 25
Error in heart rate
Percentage of error in heart rate
Number of subjects
GeMREM
Compressive Sensing
Error in lf-hf ratio of R-R intervals
Percentage error in lf-hf ratio of R-R intervals
Number of subjects 2 4 6 8 10 12 14 16 18
>20 5 15 2 4 6 8 10 12 14 16 18
>20 5 15
GeMREM
Compressive Sensing
Number of subjects
Percentage error in lf-hf ratio of R-R intervals
Errors in standard deviation of R-R intervals
Percentage error in standard deviation of R-R intervals
Number of subjects 2 4 6 8 10 12 14 16 18
>20 5 15 2 4 6 8 10 12 14 16 18
>20 5 15
GeMREM
Compressive Sensing
Number of subjects
Percentage error in standard deviation of R-R intervals

8. Morphological Parameters
Error in width of QRS complex
Error in maximum amplitude of QRS
Percentage error in width of QRS complex
Number of subjects 2 4 6 8 10 12 14 16 18 5 15 20 25
GeMREM 6
Number of subjects 5
GeMREM
Number of subjects 4 3 2 1
Percentage error in width of QRS complex 2 4 6 8 10 12 14
Percentage error in maximum amplitude of QRS complex 2 4 6 8 10 12 14 16 18 5 15 20 25 6
Compressive Sensing 5
Compressive Sensing 4
Number of subjects 3 2 1 2 4 6 8 10 12 14
Percentage error in maximum amplitude of QRS complex
Error in minimum amplitude of QRS
Error in maximum amplitude of P wave 6 10 5
GeMREM 8
GeMREM
Number of subjects 4 6 3
Number of subjects 2 4 1 2 2 4 6 8 10 12 14 16 18
>20 5 10 15
>20
Percentage error in minimum amplitude of QRS complex
Percentage of error in maximum amplitude of P wave 6 10 5
Compressive Sensing 8
Compressive Sensing 4 6
Number of subjects 3
Number of subjects 2 4 1 2 2 4 6 8 10 12 14 16 18
>20 5 10 15
>20
Percentage error in minimum amplitude of QRS complex
Percentage of error in maximum amplitude of P wave

9. Morphological Parameters Continued
Error in duration of P wave
Error in maximum amplitude of T wave
Percentage error in duration of P wave
Number of subjects 2 4 6 8 10 12 14 16 18
>20 3 9 15 10
GeMREM 8
GeMREM
Number of subjects 6
Number of subjects 4 2 2 4 6 8 10 12 14 16 18
Percentage error in duration of P wave
Percentage error in maximum amplitude of T wave 15 10 12
Compressive Sensing 8 9
Compressive Sensing 6 6
Number of subjects 4 3 2 2 4 6 8 10 12 14 16 18
>20 2 4 6 8 10 12 14 16 18
Percentage error in maximum amplitude of T wave
GeMREM better than CS is terms of morphology
Histogram of error in duration of T waves 20 16
GeM-REM
Number of subjects 12
CS better than GeMREM in terms of temporal parameters 8 4 2 4 6 8 10 12 14 16 18
>20
Percentage of error in duration of T waves
GeMREM communication compression 20
33.6 16
Compressive Sensing
Number of subjects 12 8
CS sensing compression 4
1.5 2 4 6 8 10 12 14 16 18
>20
Percentage of error in duration of T waves

10. Recovery Execution Time

11. Lesson Learned & Future Works
GeMREM focusses on shape
Non linear transformation
Complex sensors easier reconstruction
CS focusses on time domain
Linear domain sparsity
Lightweight sensors
Complex reconstruction
Performance evaluation on different scenarios
Ambulatory ECG monitoring at home (ongoing study on 100 patients)
Pre-mature baby monitoring

12. Thank You and Questions

13. Compressive Sensing X Z Y X = A = Φ
Reduction in sensing frequency below Nyquist rate
Original Signal
Sparse Coefficients
Sensed Signal
Original Signal X Z Y X = A = Φ
Random Sparse Sensing Matrix
Minimize – mean square magnitude of 𝑍
Such that - 𝑌=Φ𝐴𝑍
Simpler and generic sensors
Signal recovery requires solving optimization
Formulae simple form
Sparsity domain?
How to find a transform domain?
Is it really sparse?
Why does it work?
Need a sparsity domain
Only linear transforms
General signal sampling theory

14. Comparison With Compressive Sensing
Linear DWT Transform Basis
DWT coefficients for 30s of ECG data, coefficients less than 0.01 were ignored, sparsity 𝒌 𝒏 =𝟐.𝟓
5s snippet from 30 seconds of ECG data
Nonlinear generative model based transform
Generative model coefficients for recovering 30 s of ECG data, sparsity 𝑘 𝑛 =36.2
𝒍 𝒑 norm error metric based recovery algorithm
Diagnostic equivalence based signal recovery
Very low error for temporal properties, however, high error in morphology recovery
Very low error for morphological properties, error in temporal parameters under diagnostic requirements
Sparsity used in reducing sensing frequency
Sparsity used in reducing only communication bandwidth
Compressive Sensing
GeMREM
Signal Shape
Shape distortions
Shape Preservation
Advantages
Disadvantages
Legend: P R Q S T
Fidelity in practice?
Clinical Study on ICU patients