Organic building blocks. The choice of materials to implement the proposed signal sorting architecture is oriented to achieve prompt translatability, using only polyimide substrates, gold leads and poly(3,4-ethylenedioxythiophene) polystyrene sulfonate as organic resistor, as semi-conductive channel and gate electrode material in EGOTs and as interfacial layer in organic electrolyte capacitors (Fig. 1a). Application-specific fabrication details are provided in the Methods. Passive low-pass (Fig. 1b) and high-pass (Fig. 1d) filters are built by connecting in series an organic electrolyte capacitor and an organic resistor, and connecting to ground the capacitor or the resistor, respectively. In this work, organic low-pass filters feature 6Ω resistor and 295µF capacitor, while organic high-pass filters are built on 14 Ω resistance and 220µF capacitor. The input voltage, VIN, is supplied at the non-grounded terminal of the series composed by the organic resistor and the organic electrolyte capacitor and the output voltage - termed VL for the low-pass filter and VH for the high-pass one - is collected at the shared node. The frequency profiles of organic filters are evaluated in response to monochromatic sinusoidal VIN waves with peak-to-peak amplitude of 200mV and frequency values ranging from 1Hz to 1kHz. Figure 1c and e show the responses of low- and high-pass filters, respectively, in the form of Bode plots obtained calculating the gain of low-pass filters, gainLP= 20 log10(VL/Vin) and the gain of the high-pass filter, gainHP = 20 log10(VH/Vin). Cut-off frequencies, fc, are estimated as the frequency values for which gainLP = -3dB for low-pass filters and gainHP = -19dB for high-pass ones (i.e., 3 dB less than the maximum gainHP =-16 dB value achieved in the investigated frequency range). The resulting cut-off frequencies are fc ≈ 5Hz for the proposed organic low-pass filters and fc ≈ 30Hz for the high-pass ones, with slopes equal to -6dB/decade and 9dB/decade, respectively.
Interestingly, the use of organic analogues of passive circuit elements causes two main effects with respect to inorganic solid-state counterparts, namely a “red-shift” of fc with respect to theoretically expected values (both for high- and low-pass filters) and a substantially negative gain (around − 16dB) of the passed high-frequency band in the case of high-pass filters. Both these effects arise from the fact that, here, an electrolyte compartment is used as a capacitive element in RC filters. If, on the one hand, this enables the engineering of flexible architectures and the exploitation of the high equivalent capacitance values and of the fine tunability of a PEDOT/PSS-electrolyte interface 25, on the other it brings into play substantial deviations from ideal response, which should be taken into account. The main deviation of the proposed electrolyte organic capacitor from an ideal one is that the former features a non-negligible resistance contributed by the electrolyte solution, RS, which is in series with the equivalent capacitor of the electrical double layer, CDL. It is well known that, for standard RC passive filters, equations (1) and (2) hold, for high-pass and low-pass configurations, respectively.
$$\frac{{V}_{out}}{{V}_{in}}=\frac{\left|{Z}_{R}\right|}{{|Z}_{R}|+{|Z}_{C}|}\approx \frac{\omega RC}{\omega RC+1} \left(1\right)$$
$$\frac{{V}_{out}}{{V}_{in}}=\frac{\left|{Z}_{C}\right|}{{|Z}_{R}|+{|Z}_{C}|}\approx \frac{1}{\omega RC+1} \left(2\right)$$
Where ZC and ZR are the impedances of the resistor and of the capacitor, ω is the angular frequency of the input wave, ω = 2πν, R is the resistance of the resistor, and C is the capacitance of the capacitor. The attenuation effect of RS, observed solely in high-pass organic filters, is manifest by re-casting Eq. (1) into Eq. (3):
$$\frac{{V}_{out}}{{V}_{in}}=\frac{\omega R{C}_{DL}}{\omega (R+{R}_{S}){C}_{DL}+1} \left(3\right)$$
Despite following the expected trend of a high-pass filter, namely being 0-valued at null frequency and growing linearly with ω when \(\omega \left(R+{R}_{S}\right){C}_{DL}\ll 1,\) the \({V}_{out}/{V}_{in}\)ratio approaches \(R/(R+{R}_{S})\) at the limit of infinite frequency, which necessarily is smaller than one. The entity of this attenuation is indeed a measure of the solution resistance. In the proposed high-pass filter (Fig. 1d and e), where R=14 Ω, the plateau at -16dB corresponds to \(\frac{R}{R+{R}_{S}}\approx 0.156\) leading to an estimated RS≈75 Ω, which is coherent with the structure of the proposed capacitive element of organic RC filters.
Conversely, for the low-pass filter, the addition of RS into Eq. (2) yields Eq. (4):
$$\frac{{V}_{out}}{{V}_{in}}=\frac{1}{\omega \left(R+{R}_{S}\right){C}_{DL}+1} \left(4\right)$$
which predicts the ideal behaviour of a low-pass filter (i.e., \(\frac{{V}_{out}}{{V}_{in}}=1\) at \(\omega =0\) and \(\frac{{V}_{out}}{{V}_{in}}=0\) at \(\omega =\infty\)).
The shift towards lower frequencies – i.e. the red-shift effect – observed on the cut-off frequency, fc, arises from the equality between Eq. (1) and Eq. (3) or between Eq. (2) and Eq. (4) which implies that the response of the ideal filter to the input wave with angular frequency ω1, is equal to the response of (real) organic filter to the input wave with angular frequency ω2. By assuming C = CDL, it turns out:
$${\text{ω}}_{\text{2}}\text{=}\frac{\text{R}}{\text{R+}{\text{R}}_{\text{s}}}{\text{ω}}_{\text{1}}$$
Since \(R/(R+{R}_{S})\) is always less than 1, this means that ω2 will be always lower than ω1, quantitatively accounting for the red-shift in terms of solution resistance.
Electrolyte-Gated Organic Transistor. In addition to passive filters, the third building block for the proposed basic sorting-platform is the electrolyte-gated organic transistor. EGOTs were fabricated and characterized, as described in the Methods. Briefly, EGOTs with aspect ratio (W/L) equal to 4 were used to implement the sorting platform and to record EMG signals. A PEDOT:PSS (Clevios PH1000, 5% v/v DMSO, 0.2% v/v GOPS) formulation, diluted ten times in MilliQ, was drop-cast in a well which exposed only the terminal portion of the source and drain leads (area = 0.8 mm × 0.9 mm). A PEDOT:PSS planar gate was obtained by following the same protocol. EGOTs were characterized in common-source/common-ground configuration (Fig. 2a) by acquiring transfer curves (Fig. 2b) and evaluating their bandwidth. This is done by applying monochromatic sinusoidal waves, at different DC offset values, at the gate electrode while recording the channel current (Fig. 2b and d), as described in the Methods. Organic transistors generally act as low-pass filter, and their frequency band can be modulated not only acting on the device geometrical parameters23, 26–28 but also varying the gate potential. Thus, increasing the density of mobile charges in the channel induces a slight shift of the cut-off frequency from higher to lower values 29 (Fig. 2d). The frequency responses of the presented EGOTs show the expected trend, with a low-pass behaviour and a gain set by the gate potential which also show a minor influence on the cut-off frequency, especially for positive values. The gate DC-offset sets the device cut-off frequency and its (quasi-static) transconductance, gm, value, i.e., the EGOT potentiometric sensitivity in DC operation. Starting from negative to positive VGS, the cut-off frequency of our devices blue shifts from 180 Hz to 220 Hz (Supplementary Fig. 1).
Transduction and sorting of electrophysiological measurements. To benchmark EGOT VGS-driven signal filtering and sorting capabilities in real time in the context of electrophysiological acquisitions, EGOTs were integrated in the processing chain of an EMG data recording session. Namely, we quantified the spectral content of muscle activity during steady isometric contraction at a moderate force level.
Figure 3a and b depict the setup for EMG signal acquisition. Data are collected as described in the Methods. Briefly, EMG signals (VEMG vs time, Supplementary Fig. 2) from the biceps brachii (acting as the prime mover in the current setting) were acquired during isometric contraction required to statically hold a weight of10 kg on the hand for 10s; EGOT-mediated EMGs are collected as IDS vs time traces. Figure 3c highlights the differences between rest and contraction (weight holding) conditions. As already mentioned, VGS sets the current DC-offset11,29, as shown in Fig. 3d, which reports the average value of current EMGs vs the correspondent applied VGS. The < IDS> vs VGS profile closely mirrors EGOT transfer characteristics (see Fig. 2b). Each < IDS> value is averaged over time and over trials between seven IDS vs time traces acquired for each applied VGS, see (Supplementary Fig. 3). The spectral content of VEMG vs time EMG acquisition is reported in Fig. 3e. The spectral amplitude is not affected by the gate voltage, as shown by the individual spectral amplitude vs time EMG acquisition for each VGS applied in Supplementary Fig. 4. Spectrogram (Fig. 3e) shows, as expected, that the EMG spectral content is mainly concentrated between 10 and 100Hz, regardless of VGS. Muscle activity that is spectrally confined in the beta-band (~ 15–30 Hz) is of particular physiological relevance because it most likely has a central neurogenic origin, contributing a major component of the descending neural drive to the spinal motor neuron pool 30–32, and could function as a useful control signal in human-machine interfaces and motor augmentation platforms 33. Therefore, we focused on the relative spectral amplitude of muscle activity within the beta-band (centered at 20 Hz) with respect to both lower- and higher frequency content.
The transduction efficiency of EGOTs can be expressed by the quantity Ψ, corresponding to the ratio between the spectral amplitudes of IDS vs time traces and that of VEMG voltage, for each investigated frequency band and VGS value. Ψ vs f vs VGS spectrogram is reported in the Supplementary Fig. 5, and shows an analogous trend with respect to that observed for transconductance in Fig. 2d.
Defining ϕ as the ratio between the Ψ value in the 19-21Hz range, Ψ 20, and the Ψ value at each considered frequency band, Ψfb, and expressing this ratio in dB (see Methods), it is possible to obtain a sorting efficiency map, Fig. 3f, which quantitatively expresses how much the applied VGS favours the transduction of the frequency band of interest with respect to all the other spectral components, acting as a frequency-specific transduction efficiency selector. Figure 3f provides a map of operational conditions for EGOT-transduced EMG, to be selected accordingly to what frequency contribution must be suppressed with respect to the 19-21Hz band (ϕ voltage and current map are reported in Supplementary Figs. 6 and 7). For electrophysiological evaluations which could benefit from relative enhancing of beta-band activity and attenuation of higher-frequency components, positive VGS should be applied, as testified by the strongly positive ϕ values in the top-right corner of Fig. 3f. This gain with respect to high frequency spectral content is lost at negative VGS. On the other hand, to increase the transduction efficiency at 20 Hz with respect to that at lower frequencies, which would be naturally boosted by EGOTs, Fig. 3f shows that EMG evaluations should be carried out with VGS around 0V. Additionally, the sorting efficiency map in Fig. 3f turns out to be a valuable tool also for applications which demand frequency-invariant transduction efficiency, by identifying the proper VGS regime (VGS≈-0.4V, ϕ ≈ 0 for all the investigated frequency bands).
Coupling EGOTs with organic RC filters. As already mentioned, EGOTs act as low pass filters, albeit with rather wide bandwidth. This means that, every time EGOTs are used for transduction, the entire information content which falls in the EGOT passing band is collected and amplified and this ultimately results in poor frequency-selectivity. As discussed, frequency band selectivity is highly desirable in closed-loop BMIs since it would limit transduction to specific pathologically relevant events, which may serve as triggers for the real-time loco-regional treatment, be it chemical or electrical 34–36. This would result in improved compliance and durability and in the minimization of side effects due to overuse. Bandwidth tuning in EGOTs is usually addressed by acting on geometrical device parameters. On the one hand, it is possible to redshift EGOT’s cut-off frequency by engineering large area devices37, at the price of losing some spatial resolution when aiming at their translation to epicortical arrays. Conversely, it is possible to widen their frequency range (i.e., to blue-shift – i.e. to shift towards higher frequency – their cut-off frequency) by miniaturization 26,27, at the price of significantly lowering the transconductance. Both these approaches fail in yielding band-selective devices with both high spatial resolution and good amplification capability. At this purpose, by making advantage of the inherent filtering properties of EGOTs and coupling them with organic passive building blocks, it is possible to tighten the bandwidth toward lower frequencies or to implement band-pass filters. The first scenario is achieved by coupling the low-pass filter with an EGOT (Fig. 4a). Conversely, the second one arises from the coupling between an EGOT and a high-pass filter (Fig. 4c). In these configurations the filters’ output voltages, termed VL and VH, are used as input gate voltages for the EGOTs, keeping a constant channel bias, VDS = – 0.7 V. The resulting channel currents are termed IL and IH, for the configurations in Fig. 4a and Fig. 4c, respectively. Transconductance vs frequency profiles are obtained by dividing the amplitudes of IL and IH by the amplitude of the input voltages. The configuration in Fig. 4a exhibits low-pass behaviour with a cut-off frequency as low as 5 Hz (Fig. 4b), while the configuration in Fig. 4b exhibits a band-pass profile from 5 Hz to 90 Hz (Fig. 4d). These two independent architectures can be paired to foster a platform to sort and amplify a common VIN in two differently factorized output signals, as discussed in the next section.
Sorting platform. In Fig. 5a the sorting platform architecture sketch is reported. A common VIN is driven through the filter inputs and the resulting VL and VH are connected to two EGOTs gate electrodes. The channel currents are recorded by fixing VDS equal to -0.7V, as discussed in Methods. To validate our architecture a first characterisation was performed by applying sinusoidal waves with a peak-to-peak amplitude of 200 mV and frequencies spanning from 1 to 1000 Hz as common input signals, while collecting IL and IH. The current amplitude vs frequency trends are reported in Fig. 5c, highlighting the sorting platform frequency-response, closely mirroring the transconductance trends in Fig. 4. After validating its operation with monochromatic waves, a sweep wave with 200mV peak-to-peak amplitude and frequency interval from 1 to 200 Hz was used as input signal. The IL and IH currents versus time are reported in Fig. 5b (bottom panel), showing that the designed basic circuit enables to discriminate both in amplitude and in frequency the output responses in real-time. To benchmark the sorting efficiency of the proposed architecture against a more complex signal, a pre-recorded in vivo somatosensory evoked potential (SEP), which is a well-defined potential modulation in the rat barrel cortex, is digitalized 38,39,Fig. 5d, and driven (see Methods) as input signal. In a translational perspective, we want to test its sensitivity in discriminate some of the relevant frequency bands that are usually evaluated in SEP characterization: alfa, α (5 Hz – 15Hz),beta, β (15 Hz – 30 Hz), low gamma, γL (30 Hz- 80Hz), and high gamma, γH (80 Hz – 150 Hz), band9,40. Digitalised SEP amplitude was increased to 200mVpp to use it as a VIN model signal. IL and IH are recorded by keeping drain-source voltage at -0.7 V.
Defining η as the percentage of total power in the selected frequency interval, the analysis of the frequency content in relevant electrophysiological bands was performed, as discussed in Methods. Briefly, η is evaluated as the ratio between the IL (IH) power in a specific frequency band and the IL (IH) broadband power, Fig. 5e. Even with complex event related signals, the sorting architecture successfully discriminates its different frequency domains. The low-pass branch of this architecture selects α and β frequency bands, meanwhile the band-pass branch preferentially transmits frequencies higher than 30 Hz.
Since the EGOT implemented in this design has a cut-off frequency of roughly 120 Hz and, as a consequence, the gain of the two branches is comparable at high frequencies, the high-gamma η−values of both branches are almost equal.
To mark the frequency gain switch in the sorting platform branches, the sorting factor is estimated, Fig. 5f. The sorting factor is computed by dividing the IL and IH output signals in 13 equally spaced intervals, as described in Methods, and calculating the power of each frequency interval. The ratio between the IL power in a specific frequency band and the IH power in the same frequency band results in the sorting factor, which is a quantitative estimate of the preferential routing of that frequency among the two branches of the sorting architecture. If, for a given frequency interval the sorting factor exceeds unity, the signal content in that frequency interval will be preferentially transduced as IL. Conversely, frequency intervals which yield sorting factors minor than one will be preferentially transduced as IH. In the proposed architecture, the sorting factor equals one in the 20–25 Hz interval, Fig. 5f, meaning that spectral content in this frequency interval remains unsorted and can be found with equal power in IL and in IH.
Signal factorization and independent routing with organic RC filters. As a last example of the versatility of our material/device platform, we show the design and operation of a standalone sorting circuitry by pairing RC organic filters. We demonstrate it on the equalization of an audio signal. Audio signals were chosen as archetypes of frequency-encoded information, with relevant spectral content across a wide frequency range, namely between 20Hz and 20kHz. Since our architecture is tailored to operate in the 0.1Hz -1500Hz range relevant to electrophysiological signals, we had to adopt an out of the box modification of the input signal to adapt it to the doable processing window of the circuitry. For this, by time-stretching an audio trace of a factor 25, as described in Methods, we bring it in the desired operational range, while preserving its information content. In Fig. 6 the signal processing chain of an audio sorting experiment is shown. A common audio signal input to the paired organic RC filters is split in real time into two differently equalized output signals. In detail, an audio track, namely 7 seconds of the introduction of “Sweet Child of Mine” by Guns ‘n Roses, is acquired as voltage vs time, as discussed in Methods. The voltage track is used as VIN for the filter architecture and VL and VH are collected. Figure 6 (top center) depicts the spectra of the untreated input signal, the low-passed and the high-passed signal spectra are shown in Fig. 6 (bottom). By time stretching these signals of a factor 1/25 it is possible to hear the different equalization of the two branches, their related audio files can be downloaded as Supplementary Audio1 and Supplementary Audio2. From the spectra it is possible to notice how the normalized amplitude of the low-pass and high-pass spectra strongly differ in terms of frequency contribution. A major frequency content between 120 Hz and 250 Hz can be observed in the low-passed spectrum with respect to the high-passed one, in which the dominant spectral contributions are above 250 Hz.