Lets talk about our impedance and the 'impedance' of Competitor 'B'.
Impedance is the electrical resistance for alternating current. So, an impedance spec is only meaningful when given at a specific frequency. A high input impedance is certainly an important feature for an EEG amplifier, see for example this peer reviewed paper by one of the BioSemi founders. From this and other publications it can be learned that what matters in particular is the input impedance at frequencies within the EEG bandwidth. The input resistance (the impedance at DC) is relatively unimportant so long as it is high enough to prevent high input bias currents (and this is ensured by all contemporary FET and CMOS operational amplifiers anyway).
BioSemi fully specs the amplifier input characteristics with an input resistance of 1000 MOhm and an input capacitance of 11 pF. These values lead to an input impedance at 50 Hz of 300 MOhm (and the impedance at every other frequency can be readily calculated from the resistance and capacitance values). The exceptionally low input capacitance (which determines the input impedance at higher frequencies) could be achieved by using active electrodes: the amplifier input is as close as possible to the actual electrode, so stray capacitances by long wires and amplifier cabinets are avoided.
Competitor 'B' on the other hand, specifies only the relatively unimportant and easy to achieve "input resistance" (Input impedance for DC). The all important input capacitance and following impedance at frequencies within the EEG bandwidth are not specified at all. It basically leaves the potential customers in the dark about the abilities of the amplifier to handle electrodes with a high impedance (e.g. electrodes on the unprepared skin with impedances higher than 5 kOhm at 10 Hz), and the abilities to reject 50 Hz interference.
BioSemi specifies an input impedance of 300 MOhm @ 50 Hz (10^12 Ohm // 11 pF) (https://www.biosemi.com/activetwo_full_specs.htm)
Competitor 'B' specifies an input impedance (for DC) of > 1000 MOhm
So, the input impedance at DC is exactly the same (1000 MOhm = 10^12 Ohm), whereas the important impedance at 50 Hz is 300 MOhm for the BioSemi, determined by the ultra low 11 pF input capacitance (Competitor 'B' does not provide the input capacitance and related AC input impedance at all)
Lets talk about EEG amplifier noise specifications.
Input noise is one of the key specifications of an EEG amplifier. It is therefore of utmost importance for a buyer, to compare the input noise between different EEG amplifiers. However, making a proper and fair comparison is often complicated in practice due to the fact that noise may be specified in different ways by different manufacturers. Below, the background of different noise specifications is explained, and methods are offered to convert specifications to allow a meaningful comparison:
1) Noise bandwidth
Noise is an interference signal with random amplitudes and random frequencies. When specifying the amount of noise of a measurement device, it is therefore always essential to define the bandwidth of the measured noise. In other words, a noise specification without mentioning the bandwidth is meaningless. A larger bandwidth always means more noise.
2) Noise spectrum
In EEG amplifiers, we see two main types of noise, each with a particular frequency spectrum: white noise and pink noise. White noise is noise with a flat frequency spectrum. This means that the noise power is equal for every frequency interval in the pass-band of the amplifier. The noise amplitude (square root of the noise power) is therefore proportional with the square root of the bandwidth. Pink noise -also called 1/f noise - is noise with the noise power inversely proportional to the frequency and the noise amplitude is inversely proportional with the square root of the frequency. So - contrary to white noise - the noise power of pink noise increases with decreasing frequency. We see pink noise in nearly all processes in nature. For example: the noise of the skin-gel-electrode interface is dominated by pink noise (high noise at low frequencies and very low noise at high frequencies).
3) Amplifier types
The noise of a properly designed EEG amplifiers is dominated by the input stage (and not for example by the quantization noise of the analog-to-digital converter). The input stage of a contemporary EEG amplifier is formed by an operational amplifier (Op-amp) executed as an integrated circuit (IC). Two types of Op-amp are used: linear types and chopper types. The noise of a linear Op-amp is a combination of pink and white noise: at lower frequencies pink noise dominates, whereas at higher frequencies white noise dominates. The transition point between pink and white noise (the "1/f corner frequency) is typically somewhere between 10 and 100 Hz. So, below the 1/f corner frequency, the noise density (the noise power or amplitude per frequency interval) increases with decreasing frequency, whereas the noise density above the the 1/f corner frequency remains constant. A chopper Op-amp on the other hand, has a sophisticated input circuit to eliminate pink noise. So, this type of amplifier has a white noise spectrum over its entire frequency band. The disadvantage is that usually the white noise level is a bit higher than the white noise that can be achieved in linear Op-amps. To summarize: when comparing linear and chopper Op-amps, the lowest noise at low frequencies (below approx. 10 Hz) is offered by chopper Op-amps, while linear Op-amps are preferred for the higher frequencies. Since the noise of the electrode-skin interface also has an pink characteristic, and since the electrode noise at low frequencies is typically well above typical linear Op-amp noise, a linear Op-amps seems to be the best choice as an input device of an EEG amplifier. Especially at high frequencies (e.g. ABR measurements) where electrode noise density may be very low, the lower noise of a linear Op-amp can be essential to achieve the best possible results. Nevertheless, other manufacturers of EEG amplifiers may still choose to use a chopper Op-amp as input device in an attempt to offer the lowest possible input noise in the lower part of the EEG spectrum.
4) BioSemi ActiveTwo
BioSemi uses linear Op-amps in the active electrodes and AD-box input stages. This is reflected in the noise specifications provided for different bandwidths (and different sample-rates): the noise in a 1600 Hz band is much less than double the noise in a 400 Hz band (square root of 1600/400 is 2).
5) Comparing noise specifications
The input noise of the BioSemi ActiveTwo system is specified as 8 uVpp for a 3200 Hz bandwidth (16 kHz sample rate). Now suppose that we want to compare the BioSemi specification with a hypothetical alternative based on a chopper Op-amp specified with a noise of 1.5 uVpp in a 20 Hz bandwidth. As explained above, noise figures can only be compared for equal bandwidth. Because we assume a chopper input for the alternative design, we can quite easily calculate the noise for a 3200 Hz bandwidth. As said: a chopper Op-amp only generates white noise which means that the total noise is proportional with the square root of the bandwidth: Noise = 1.5uVpp*(3200/20)^1/2 = 19 uVpp. So although the alternative design seems to have a lower noise level when bandwidth is not taken into account, the noise is actually 60% higher than the ActiveTwo when comparing for equal bandwidth. Note that we can not apply a similar calculation to reduce the noise bandwidth to 20 Hz for the BioSemi design, because the noise of the BioSemi has a pink spectrum at lower frequencies.
6) Practical implications
What does this factor of 1.6 higher noise mean in practice ? Well, the signal-to-noise ratio for an ERP average is proportional with the number of sweeps squared (to halve the noise level, you need a factor of 4 more sweeps). Since 1.6^2= 2.5 the alternative design will require 2.5 as much sweeps to achieve the same ERP result when compared with the BioSemi, provided that the total noise is dominated by amplifier noise (and not by electrode noise or noise of background EEG). Consequently, every ERP experiment will take 2.5 times as long.
February 9th, 2021.