Thanks to these developments, the healthcare interface, such as the front-end integrated chip, is also working to leverage machine learning to deliver various solutions to consumers. The main idea of the article is front-end amplifiers that reduce noise and motion artifacts, wavelet filters that decompose the input signal for wavelet transformation and machine learning for gesture recognition.
Introduction
Bio signals in healthcare
With these properties, the process that analyzes the biosignal is developed and used to diagnose and research many diseases and physical disorders. 2(b), the biosignals have a small voltage amplitude, so the biosignal must be amplified with a large gain.
Data process of Bio signal
Fourier transform is well known as frequency domain analysis but it has several limitations to analyze in time domain. If the frequency of the signal changes at a certain time, Fourier transform cannot tell when a certain change occurs. Overall system includes front-end amplifiers, wavelet filters, 12bit sequential approximation analog to digital (SAR-ADC), and microcontroller (MCU).
The wavelet filters consist of several low pass filters (LPF) which can degrade the signal. The result of the chip is processed by MATLAB machine learning, so that the gestures of the fingers can be recognized in the frequency domain. There are some techniques in front-end readout integrated circuit (ROIC) to detect a biosignal.
Front-end ROICs mean integrated circuits designed to detect and control signals in a chip. To detect a specific biosignal, amplifiers used in front-end ROICs require specific gain and bandwidth because biosignals have their own frequency and voltage ranges.
Front-end circuit
- Chopper stabilization
This chapter explains some circuit techniques used to detect biosignals and reduce noise and offset in a chip. Noise in a chip is crucial for the quality of biosignal detection and for motion artifacts. Nowadays, to reduce noise, especially input noise, in a chip, chopping stabilization is commonly used.
AC signal is amplified by the amplifier and demodulated from AC to DC by the second chopper. However, output voltage present at the amplifier is not modulated from the first chopper. Since offset is DC signal, offset is modulated from DC to AC by the second chopper.
To remove the offset which is a high frequency component, a low pass filter (LPF) is used after second chopper.
LPFCHOPPER
CHOPPER
V offset
V OUTV1V2
DC Servo Loop (DSL)
The main problems with biological signals are high total interference and differential electrode offset caused by differential body potentials at each electrode. Most biological signal amplifiers have high gain because bio signals have a low voltage amplitude. The DC offset at the amplifiers input stage is amplified by high gain and then causes saturation at the output stage.
DC offset voltage appearing at the input stage of the capacitively coupled instrumentation amplifier (CCIA), CCIA is amplified and then integrated by the fully differential integrator.
V OUTVin
DSLCin
Ripple Reduction Loop (RRL)
The offset caused by the mismatch of the input transistor to the CCIA is modulated by the second chopper then becomes a ripple in the output stage. The ripple induced by the second chopper is demodulated to DC offset by the input chopper of the RRL. After integration, the integrated offset is modulated by the output chopper of the RRL and negative-returned to the input stage of the CCIA.
The difference between DSL and RRL is the position to which the loop output is fed back. Since the output of RRL is modulated and demodulated by two choppers, it is fed back to the first chopper of CCIA. However, the output of DSL is only modulated by one chopper and then fed back to the first chopper of CCIA.
RRLCin
Data process
- Data processing for bio signal
- Wavelet transform for bio signal
The analog bio signal is measured in a front-end amplifier with noise reduction circuit and then converted into digital data in ADC. With the amount of data being transmitted during the Fourth Industrial Revolution increasing exponentially, a way to analyze a biosignal has also been developed. Before the revolution, the only amplitude in the time domain was useful information, and it is difficult to analyze the biosignal in real time.
To analyze a biosignal such as EMG and EEG that are identified by frequency rather than specific waveforms, Fourier transform is the most commonly used method. In other words, the biosignal is usually constant and sometimes occurs when the specific gesture or state of the body changes. Because of these characteristics, time and frequency information is required to properly analyze a biosignal.
For this reason, Wavelet transform is good for analyzing sparse and non-stationary signal rather than Fourier transform. As mentioned in section 1.2, unlike Fourier transform, wavelet transform has both frequency and time information.
Filter bank
Limitation and Idea for sensing EMG
Movement artifact is the dominant limitation that reduces sensing quality as physical movements cause unstable electrode contact on human skin. The unstable contact changes an electrode impedance, then the moving object affects the transient signal. However, it is not easy to eliminate the motion object because the frequency range of the motion object is similar to that of the bio signal [10].
The advantages of wavelet transform are as follows; the wavelet transform gives frequency information in the time domain and can analyze the signal by zooming in on a certain frequency range [8]. This paper proposes a wavelet-based EMG detection system that consists of multiple front-end amplifiers, wavelet filters, and a 12-bit SAR-ADC.
Implementation of integrated circuit
12bit SAR-ADC
External devices(Machine Learning)
Front-end Amplifiers
Front-end Amplifiers
After the CCIA, a programmable gain amplifier (PGA) amplifies the output of the CCIA and is implemented as an LPF. This section will explain the CCIA structure and additional circuitry to eliminate noise from various sources. The proposed front-end amplifiers have a CCIA structure, which consists of a noise amplifier (LNA), a chopper, a DC servo.
Unlike the conventional CCIA, the proposed CCIA of the front-end amplifier has no resistance in the feedback loop, because the programmed gain amplifier (PGA) is installed behind the CCIA, then the additional LPF in the CCIA is not required. Instead of the feedback capacitor, a duty cycle resistor (DCR) is added to give the total voltage at the input of the LNA.
RRL IBL
CCIA
Front-end Amplifier
Wavelet Filters
As mentioned earlier in this section, the motion artifact should be eliminated to improve the quality of monitoring and analysis. To eliminate the motion artifact, this paper adopts the wavelet transform for frequency domain analysis. The proposed wavelet filters perform the same decomposition as a filter bank in DWT, but use analog circuits unlike DWT.
The main goals of the proposed wavelet filters are as follows: removing motion artifacts and decomposing the input signal into different frequency ranges. To achieve these goals, the proposed wavelet filters used buffer and subtractor instead of HPF and its block diagram is shown in Figure 17. Since the motion artifact occurs when a subject moves the body part with the electrode, the signal contains motion artifacts in some frequency ranges.
The buffered signal is subtracted from the low-pass filtered signal, then the motion object contained in the buffered and filtered signal is removed. Also, when the low-frequency component is removed from the primary signal, the subtracted signal can be a high-pass filtered output because only the high-frequency component remains.
Wavelet filters
Machine learning algorithm for EMG recognition
Serial Peripheral Interface (SPI) and Universal Asynchronous Receiver/Transmitter (UART) transmits the output of 12 bit SAR-ADC to the Bluetooth module, then the Bluetooth modules transmit the output to the PC. In machine learning, the number of hidden layer should be chosen to optimize the accuracy. As the number of hidden layers increases, the motion object is also equipped, then it causes an overfitting problem.
To avoid this problem, the optimization is performed after extracting features with frequency information. As a result, the movement of each finger is recognized with frequency domain features and an optimized number of hidden layers.
Software & embedded system Architecture
C SETUP
- Fabrication results
The embedded system to control the whole module is implemented by a low power MCU (STM32L4, STmicroelectronis) and the flowchart and block diagram are shown in Figure. The chip is implemented in a CMOS 0.18 µm process and the chip area is 2 mm x 2 mm. The other parts are for the other article because the 12-bit SAR ADC is shared with the EEG analysis circuit.
SAR-ADC
Measurement results .1 DSL
- Wavelet Filters
The result of CCIA when the input signal is sine wave with 170Hz, 10mVpp, 0.9V common mode voltage and no offset voltage is shown in Fig. The difference of common mode voltage decreases when the DSL is on with the same input signal. 22 Measurement of RRL with input signal that is 5mVpp and 170Hz. a) RRL off and (b) RRL on with 75KHz chopper frequency.
Then the amplitude of output voltage of each stage is larger in the frequency range of the stage than others. The wavelet filtered data has a high frequency component due to the reset phase in subtractor. Due to the above graph, this paper chose the number of hidden layer as 25. 29 shows the confusion matrices of this paper.
The percentage of each element represents the ratio between the number of the element and the total number of all elements. The green letters of the gray-colored element indicate the accuracy of each row and column.
Conclusion
MULTIBRIEFS: EXCLUSIVE, http://exclusive.multibriefs.com/content/printed-electronics-allow-technological-leap-in-wearable-devices/science-technology, May 2017, Web. 3] https://en.wikipedia.org/wiki/Electroencephalography [4] https://en.wikipedia.org/wiki/Electrooculography [5] https://en.wikipedia.org/wiki/Electrocardiography [6] https://en.wikipedia.org/wiki/Photoplethysmogram [7] https://en.wikipedia.org/wiki/Electromyography. Chun-Lin, A Tutorial of the Wavelet Transform., Taipei, Taiwan, 2010 [9] J, Olkkonen, “Discrete Wavelet Transforms – Theory and Applications”, 2011.
The virtual trackpad: an electromyography-based, wireless, real-time, low-power, embedded hand gesture recognition system using an event-driven artificial neural network.
