際際滷

際際滷Share a Scribd company logo
ECG Arrhythmia Classification on an
Ultra-Low-Power Microcontroller
 Wearable biomedical systems allow doctors to continuously monitor their patients over longer periods,
which is especially useful to detect rarely occurring events such as cardiac arrhythmias.
 This work is the first to document a complete beat to-beat arrhythmia classification system implemented
on a custom ultra-low-power microcontroller
 It includes a single-channel analog front-end (AFE) circuit for electrocardiogram (ECG) signal
acquisition, and a digital back-end (DBE) processor to execute the support vector machine (SVM)
classification software with a Cortex-M4 CPU.
 The proposed system has been prototyped on the SleepRider SoC, a 28-nm fully-depleted silicon on
insulator (FD-SOI) 3.1-mm2 chip. It consumes 13.1 亮Won average from a 1.8-V supply.
 CARDIOVASCULAR diseases (CVD) are the leading cause of death and disability worldwide, especially
in developed countries.
 Among those, cardiac arrhythmias are associated with an increased risk of stroke and sudden cardiac death.
 In order to diagnose heart diseases, one of the most effective tools is the electrocardiogram (ECG) which
measures the electrical activity of the heart.
 However, due to the infrequent nature of arrhythmia events, long-term monitoring using a wearable device
is often necessary to observe them.
 Nevertheless,being able to use monitoring devices during the patients daily activities imposes stringent
requirements on their signal quality, size, and power consumption.
 To avoid the need for transmitting or storing very large amounts of data, real-time processing of the ECG
signal to detect cardiac arrhythmias has emerged as a solution to extract only the relevant information when
abnormal events occur.
INTRODUCTION
SYSTEM
ARCHITE
CTURE
 The proposed ECG arrhythmia classification system is embedded in the SleepRider MCU SoC
as shown in Fig.A pair of chest electrodes are connected to the differential input of the
instrumentation amplifier (IA) in the AFE.
 A third electrode is used to set the patient reference voltage through an off-chip buffer. The IA
uses a current-balancing topology with chopper modulation .
 The differential output of the IA is digitized at 250 Hz using two single-ended time-based ADCs
with 14-bit resolution.
 The acquired samples are temporarily stored in the SRAM data memory by the direct memory
access (DMA) module while the system is in sleep mode and the DBE is inactive. When 128
samples have been stored in the memory (approximately every 0.5 s), the ARM Cortex M4 CPU
inside the DBE wakes up and processes the data.
 A clock and power management unit (CPMU) provides a reference current, a low-frequency
peripheral clock at 4 MHz and a high-frequency clock for the DBE at 64 MHz. Several low-
voltage supplies for the different blocks are also generated from a single 1.8-V supply: 0.4 V for
the ADC and the digital logic, 0.5 V and 0.8 V for memories and auxiliary blocks.
SYSTEM ARCHITECTURE
ANALOG FRONT-END DESIGN Low-Noise Instrumentation Amplifier
The fully-differential current-balancing topology of the IA shown in Fig.
 The fully-differential current-balancing topology of the IA shown in Fig. 2 relies on a transconductance (Gm) and a transimpedance (Zm)
stage. In the Gm stage, the input differential voltage vin is copied by the flipped voltage follower (FVF) structure with transistors P1 and P2
across resistance Ri, which induces a current difference between the left and right branches.
 The FVF provides a unity gain with a large input range, while readily acting as a current sensor [18]. The P3-P5 structure allows to lower the
drain voltage of P1 and improves the input dynamic range of the IA. The current difference is copied to the Zm stage thanks to the P2-P6
current mirror with current gain A and goes through resistor Ro to generate a differential voltage.
 The total gain Av of the IA is defined by the product of the Gm with gm and gd the transistor transconductance and output conductance,
respectively.
Time-Based ADC
The differential signal at the output of the IA is digitized by two
single-ended time-based ADCs. As shown in Fig. 5, a voltage-
controlled ring oscillator (VCRO) performs the voltageto-time
domain conversion. The phase of the oscillator is integrated by
a counter and successive samples from the coare differentiated
to obtain the frequency information. The main advantage of
time-based ADCs is that they can be implemented as digital-
only modules using foundry standard cells resulting in a limited
design effort and low area
 The main bottleneck of VCRO-based ADCs is the
nonlinearity of the voltage-to-frequency conversion.
 A typical current-starved ring oscillator (CSRO) has a
strongly non-linear transfer function which strongly limits
the input dynamic range as illustrated in Fig. 6.
 Previous works have dealt with this issue by adding a
feedback loop to reduce the input swing of the signal.
 In this work, we take advantage of the strong back-gate
effect in FD-SOI technology to create a back-bias
controlled oscillator (BBCO) that offers a much larger
input dynamic range and significantly improves linearity.
Finally, the resolution and quantization noise of the ADC are set
by the frequency of the BBCO, which depends on the number
of stages, and the sampling frequency. A smaller BBCO with
higher frequency improves the resolution but increases the
power consumption of the counter and has more thermal noise
due to poorer averaging as illustrated in Fig. 8.
By connecting the back-gate bias to the input voltage of the
ADC, the threshold voltage Vth of the transistor can be reduced,
which changes the current in each stage and the frequency of the
RO. The STMicroelectronics 28-nm FD-SOI technology used in
this work offers different Vth options that can be operated in
forward back biasing (FBB, for LVT transistors) or reverse back
biasing (RBB, for RVT transistors). Note that LVT PMOS and
RVT NMOS devices take negative back-bias voltages [22]
while LVT NMOS and RVT PMOS take positive ones. As the
input voltage ranges from 0 to 1.8 V, only the LVT NMOS and
RVT PMOS back-gates can be driven while the other transistors
are kept at their nominal back-gate voltage
DIGITAL BACK-END DESIGN
The data samples retrieved from the AFE are processed sequentially in the DBE. First, the heartbeats are
detected by the algorithm from Pan and Tompkins which consists in a band-pass filter, a derivation, a moving
average and an adaptive threshold. This allows to detect the sharp peaks in the signal that likely correspond
to the QRS complexes of the ECG signal shown in Fig. 9
When using an MCU as a basis for the arrhythmia classification system, the detection and classification
algorithm is programmed in the SRAM and executed in software. Sections IV-A and IV-B describe the algorithm
design with the feature selection process and SVM classifier, respectively. Section IV-C details the hardware
implementation of the DBE and the real-time execution constraints
The 3 classes considered in this work are the normal heartbeats with sinus rhythm (N), supraventricular
ectopic beats (S), and ventricular ectopic beats (V). Examples of these three classes are shown in Fig. 9.
As observed in the examples of Fig. 9, while both arrhythmia types show a disruption in the normal sinus
rhythm, their characteristics are quite different. The V beats differ mostly by the shape of their QRS complex.
The S beats are morphologically close to the N beats with only a missing P-wave, which is often hard to observe
in the baseline noise. For classification purposes, the features used to best distinguish those classes must
therefore be different in each case, which is rarely done when using traditional one-against-one or one-against-
all SVM classifiers
A. Feature Selection
The feature selection process is performed by a sequential forward
floating search (SFFS) algorithm which iteratively adds or removes
a feature from the set based on the improvement of the cross-
validation inference performance [16]. This performance is
measured by the j index (Ij ) equal to the mean of the sensitivities
(Sen) and positive predictive values (PPV ) for the S and V classes,
i.e. Ij = (SenS + SenV + PPVS + PPVV )/4 [16]. This arbitrary
figure of merit, normalized here to 100%, avoids the bias due to the
large proportion of N beats and focuses on the discriminative
performance for the S and V arrhythmia classes. Fig. 10 shows the
performance evolution with the number of features, where it can be
observed that a low number of features below 10 is already
sufficient to achieve peak performance. As expected, the selected
feature types are different for each classifier as shown in Table I.
Despite feature selection, a significant overlap between classes is
still present due to inter-patient variability, as seen in Fig. 11 for the
two most discriminative features in each classifier.
B. SVM Classifier
 Software execution is performed
on an ARM Cortex-M4 CPU. In order
to minimize the energy required for
performing the task, ultra-low-
power implementation techniques
are used [11]. TheCPUlogic operates
at ultra-low-voltage (0.4 V) to
reduce its power consumption, at
the expense of reduced clock
frequency and increased sensitivity
to PVT variations
Hardware Implementation and Real-Time Execution
Hardware Implementation and Real-Time Execution
RESULTS
 The SleepRider SoC was prototyped in 28-nm FD-SOI on a 3.1-mm2 chip with the core taking
up 0.94 mm2 Measurements are performed on a custom test PCB with off-chip components
such as the input HPF and a patient reference circuit.
 The setup uses a Keysight N6715B power supply, a Tektronix MSO56 oscilloscope, and a
laptop for JTAG programming and UART communication.
 In-vivo trials are performed with single-use adhesiveAg/AgCl electrodes with solid gel.
 Measurements were performed on one healthy male subject to test the functionality of the
acquisition and detection
 When performing arrhythmia classification, the SleepRider SoC consumes 13.1 亮W including
DC/DC conversion losses from the 1.8-V supply.
 This is lower than the previously reported value of 19.8 亮Wthanks to the reducedwake-up
overhead by the DMA and reduced power in the AFE
A. AFE Signal Acquisition
The measured IA shows a differential gain of 37 dB and a maximum -3-dB bandwidth of 1 kHz. The transfer
function in Fig. 15 shows a flat response in the ECG signal bandwidth. The common-mode attenuation
measured with a 300-mV pp input signal stays below 40 dB over the whole bandwidth and increases down
to 90 dB at lower frequency.
The resulting IRN is 0.75 袖 V RMS integrated over the 1-to-100-Hz frequency range. A sample signal
measured in-vivo through chest electrodes is shown in Fig. 17. An electrode distance of 8 cm is
chosen as it is the limit before QRS amplitude starts to decrease
B. DBE Inference
Given the long duration of the MIT-BIH arrhythmia database inference performance on the full dataset can hardly
be evaluated on the SoC and is thus emulated off-line. In order to take into account the impact of analog non-
idealities on the inference performance, a Python model simulates the noise and transfer response characteristics
of the IA and ADC before performing the classification. The characteristics of the AFE model were matched to the
measured hardware performance. Execution directly on the SoC is validated using shorter signal samples
This work is the first implementing an inter-patient arrhythmia classification algorithm on an MCU. The
performance obtained is comparable to other works using the inter-patient training scheme, as shown in Table III
Table V compares several mixed-signal ECG SoCs from the state of the art. The SleepRider SoC integrates all
necessary components for ECG signal acquisition and processing, as well as the power and clock management. In
addition to being more versatile, arrhythmia classification performed on such an ULP MCU achieves an energy
efficiency close to other works while performing a much more complex task.
CONCLUSION
 This work is the first tackling end-to-end ECG beat-to-beat arrhythmia classification implemented on a ULP
MCU.
 It aims at meeting the constraints of power consumption,signal quality and inference performance inherent
to wearablemonitoring devices while providing enough versatility to perform complex tasks. A prototype
chip has been fabricated in 28-nm FD-SOI technology and tested with the classification software.
 The low-power and low-noise AFE acquires the ECG signal with 2.1 亮W. The software-based DBE
performs beat-to-beat arrhythmia classification with a high energy efficiency of 2.1亮J/beat (including beat
detection and classification)and an accuracy of 91.4%.
 The system has an average power consumption of 13.1 亮Wand has been tested in-vivo to validate the
functionality

More Related Content

ARRHYTHMIA.pptx

  • 1. ECG Arrhythmia Classification on an Ultra-Low-Power Microcontroller Wearable biomedical systems allow doctors to continuously monitor their patients over longer periods, which is especially useful to detect rarely occurring events such as cardiac arrhythmias. This work is the first to document a complete beat to-beat arrhythmia classification system implemented on a custom ultra-low-power microcontroller It includes a single-channel analog front-end (AFE) circuit for electrocardiogram (ECG) signal acquisition, and a digital back-end (DBE) processor to execute the support vector machine (SVM) classification software with a Cortex-M4 CPU. The proposed system has been prototyped on the SleepRider SoC, a 28-nm fully-depleted silicon on insulator (FD-SOI) 3.1-mm2 chip. It consumes 13.1 亮Won average from a 1.8-V supply.
  • 2. CARDIOVASCULAR diseases (CVD) are the leading cause of death and disability worldwide, especially in developed countries. Among those, cardiac arrhythmias are associated with an increased risk of stroke and sudden cardiac death. In order to diagnose heart diseases, one of the most effective tools is the electrocardiogram (ECG) which measures the electrical activity of the heart. However, due to the infrequent nature of arrhythmia events, long-term monitoring using a wearable device is often necessary to observe them. Nevertheless,being able to use monitoring devices during the patients daily activities imposes stringent requirements on their signal quality, size, and power consumption. To avoid the need for transmitting or storing very large amounts of data, real-time processing of the ECG signal to detect cardiac arrhythmias has emerged as a solution to extract only the relevant information when abnormal events occur. INTRODUCTION
  • 4. The proposed ECG arrhythmia classification system is embedded in the SleepRider MCU SoC as shown in Fig.A pair of chest electrodes are connected to the differential input of the instrumentation amplifier (IA) in the AFE. A third electrode is used to set the patient reference voltage through an off-chip buffer. The IA uses a current-balancing topology with chopper modulation . The differential output of the IA is digitized at 250 Hz using two single-ended time-based ADCs with 14-bit resolution. The acquired samples are temporarily stored in the SRAM data memory by the direct memory access (DMA) module while the system is in sleep mode and the DBE is inactive. When 128 samples have been stored in the memory (approximately every 0.5 s), the ARM Cortex M4 CPU inside the DBE wakes up and processes the data. A clock and power management unit (CPMU) provides a reference current, a low-frequency peripheral clock at 4 MHz and a high-frequency clock for the DBE at 64 MHz. Several low- voltage supplies for the different blocks are also generated from a single 1.8-V supply: 0.4 V for the ADC and the digital logic, 0.5 V and 0.8 V for memories and auxiliary blocks. SYSTEM ARCHITECTURE
  • 5. ANALOG FRONT-END DESIGN Low-Noise Instrumentation Amplifier The fully-differential current-balancing topology of the IA shown in Fig.
  • 6. The fully-differential current-balancing topology of the IA shown in Fig. 2 relies on a transconductance (Gm) and a transimpedance (Zm) stage. In the Gm stage, the input differential voltage vin is copied by the flipped voltage follower (FVF) structure with transistors P1 and P2 across resistance Ri, which induces a current difference between the left and right branches. The FVF provides a unity gain with a large input range, while readily acting as a current sensor [18]. The P3-P5 structure allows to lower the drain voltage of P1 and improves the input dynamic range of the IA. The current difference is copied to the Zm stage thanks to the P2-P6 current mirror with current gain A and goes through resistor Ro to generate a differential voltage. The total gain Av of the IA is defined by the product of the Gm with gm and gd the transistor transconductance and output conductance, respectively.
  • 7. Time-Based ADC The differential signal at the output of the IA is digitized by two single-ended time-based ADCs. As shown in Fig. 5, a voltage- controlled ring oscillator (VCRO) performs the voltageto-time domain conversion. The phase of the oscillator is integrated by a counter and successive samples from the coare differentiated to obtain the frequency information. The main advantage of time-based ADCs is that they can be implemented as digital- only modules using foundry standard cells resulting in a limited design effort and low area
  • 8. The main bottleneck of VCRO-based ADCs is the nonlinearity of the voltage-to-frequency conversion. A typical current-starved ring oscillator (CSRO) has a strongly non-linear transfer function which strongly limits the input dynamic range as illustrated in Fig. 6. Previous works have dealt with this issue by adding a feedback loop to reduce the input swing of the signal. In this work, we take advantage of the strong back-gate effect in FD-SOI technology to create a back-bias controlled oscillator (BBCO) that offers a much larger input dynamic range and significantly improves linearity.
  • 9. Finally, the resolution and quantization noise of the ADC are set by the frequency of the BBCO, which depends on the number of stages, and the sampling frequency. A smaller BBCO with higher frequency improves the resolution but increases the power consumption of the counter and has more thermal noise due to poorer averaging as illustrated in Fig. 8. By connecting the back-gate bias to the input voltage of the ADC, the threshold voltage Vth of the transistor can be reduced, which changes the current in each stage and the frequency of the RO. The STMicroelectronics 28-nm FD-SOI technology used in this work offers different Vth options that can be operated in forward back biasing (FBB, for LVT transistors) or reverse back biasing (RBB, for RVT transistors). Note that LVT PMOS and RVT NMOS devices take negative back-bias voltages [22] while LVT NMOS and RVT PMOS take positive ones. As the input voltage ranges from 0 to 1.8 V, only the LVT NMOS and RVT PMOS back-gates can be driven while the other transistors are kept at their nominal back-gate voltage
  • 10. DIGITAL BACK-END DESIGN The data samples retrieved from the AFE are processed sequentially in the DBE. First, the heartbeats are detected by the algorithm from Pan and Tompkins which consists in a band-pass filter, a derivation, a moving average and an adaptive threshold. This allows to detect the sharp peaks in the signal that likely correspond to the QRS complexes of the ECG signal shown in Fig. 9 When using an MCU as a basis for the arrhythmia classification system, the detection and classification algorithm is programmed in the SRAM and executed in software. Sections IV-A and IV-B describe the algorithm design with the feature selection process and SVM classifier, respectively. Section IV-C details the hardware implementation of the DBE and the real-time execution constraints
  • 11. The 3 classes considered in this work are the normal heartbeats with sinus rhythm (N), supraventricular ectopic beats (S), and ventricular ectopic beats (V). Examples of these three classes are shown in Fig. 9.
  • 12. As observed in the examples of Fig. 9, while both arrhythmia types show a disruption in the normal sinus rhythm, their characteristics are quite different. The V beats differ mostly by the shape of their QRS complex. The S beats are morphologically close to the N beats with only a missing P-wave, which is often hard to observe in the baseline noise. For classification purposes, the features used to best distinguish those classes must therefore be different in each case, which is rarely done when using traditional one-against-one or one-against- all SVM classifiers A. Feature Selection
  • 13. The feature selection process is performed by a sequential forward floating search (SFFS) algorithm which iteratively adds or removes a feature from the set based on the improvement of the cross- validation inference performance [16]. This performance is measured by the j index (Ij ) equal to the mean of the sensitivities (Sen) and positive predictive values (PPV ) for the S and V classes, i.e. Ij = (SenS + SenV + PPVS + PPVV )/4 [16]. This arbitrary figure of merit, normalized here to 100%, avoids the bias due to the large proportion of N beats and focuses on the discriminative performance for the S and V arrhythmia classes. Fig. 10 shows the performance evolution with the number of features, where it can be observed that a low number of features below 10 is already sufficient to achieve peak performance. As expected, the selected feature types are different for each classifier as shown in Table I. Despite feature selection, a significant overlap between classes is still present due to inter-patient variability, as seen in Fig. 11 for the two most discriminative features in each classifier. B. SVM Classifier
  • 14. Software execution is performed on an ARM Cortex-M4 CPU. In order to minimize the energy required for performing the task, ultra-low- power implementation techniques are used [11]. TheCPUlogic operates at ultra-low-voltage (0.4 V) to reduce its power consumption, at the expense of reduced clock frequency and increased sensitivity to PVT variations Hardware Implementation and Real-Time Execution
  • 15. Hardware Implementation and Real-Time Execution
  • 16. RESULTS The SleepRider SoC was prototyped in 28-nm FD-SOI on a 3.1-mm2 chip with the core taking up 0.94 mm2 Measurements are performed on a custom test PCB with off-chip components such as the input HPF and a patient reference circuit. The setup uses a Keysight N6715B power supply, a Tektronix MSO56 oscilloscope, and a laptop for JTAG programming and UART communication. In-vivo trials are performed with single-use adhesiveAg/AgCl electrodes with solid gel. Measurements were performed on one healthy male subject to test the functionality of the acquisition and detection When performing arrhythmia classification, the SleepRider SoC consumes 13.1 亮W including DC/DC conversion losses from the 1.8-V supply. This is lower than the previously reported value of 19.8 亮Wthanks to the reducedwake-up overhead by the DMA and reduced power in the AFE
  • 17. A. AFE Signal Acquisition The measured IA shows a differential gain of 37 dB and a maximum -3-dB bandwidth of 1 kHz. The transfer function in Fig. 15 shows a flat response in the ECG signal bandwidth. The common-mode attenuation measured with a 300-mV pp input signal stays below 40 dB over the whole bandwidth and increases down to 90 dB at lower frequency.
  • 18. The resulting IRN is 0.75 袖 V RMS integrated over the 1-to-100-Hz frequency range. A sample signal measured in-vivo through chest electrodes is shown in Fig. 17. An electrode distance of 8 cm is chosen as it is the limit before QRS amplitude starts to decrease
  • 19. B. DBE Inference Given the long duration of the MIT-BIH arrhythmia database inference performance on the full dataset can hardly be evaluated on the SoC and is thus emulated off-line. In order to take into account the impact of analog non- idealities on the inference performance, a Python model simulates the noise and transfer response characteristics of the IA and ADC before performing the classification. The characteristics of the AFE model were matched to the measured hardware performance. Execution directly on the SoC is validated using shorter signal samples This work is the first implementing an inter-patient arrhythmia classification algorithm on an MCU. The performance obtained is comparable to other works using the inter-patient training scheme, as shown in Table III
  • 20. Table V compares several mixed-signal ECG SoCs from the state of the art. The SleepRider SoC integrates all necessary components for ECG signal acquisition and processing, as well as the power and clock management. In addition to being more versatile, arrhythmia classification performed on such an ULP MCU achieves an energy efficiency close to other works while performing a much more complex task.
  • 21. CONCLUSION This work is the first tackling end-to-end ECG beat-to-beat arrhythmia classification implemented on a ULP MCU. It aims at meeting the constraints of power consumption,signal quality and inference performance inherent to wearablemonitoring devices while providing enough versatility to perform complex tasks. A prototype chip has been fabricated in 28-nm FD-SOI technology and tested with the classification software. The low-power and low-noise AFE acquires the ECG signal with 2.1 亮W. The software-based DBE performs beat-to-beat arrhythmia classification with a high energy efficiency of 2.1亮J/beat (including beat detection and classification)and an accuracy of 91.4%. The system has an average power consumption of 13.1 亮Wand has been tested in-vivo to validate the functionality