Reservoir computing with biocompatible organic electrochemical networks for brain-inspired biosignal classification

Abstract

Early detection of malign patterns in patients' biological signals can save millions of lives. Despite the steady improvement of artificial intelligence-based techniques, the practical clinical application of these methods is mostly constrained to an offline evaluation of the patients' data. Previous studies have identified organic electrochemical devices as ideal candidates for biosignal monitoring. However, their use for pattern recognition in real time was never demonstrated. Here, we produce and characterize brain-inspired networks composed of organic electrochemical transistors and use them for time-series predictions and classification tasks using the reservoir computing approach. To show their potential use for biofluid monitoring and biosignal analysis, we classify four classes of arrhythmic heartbeats with an accuracy of 88%. The results of this study introduce a previously unexplored paradigm for biocompatible computational platforms and may enable development of ultralow-power consumption hardware-based artificial neural networks capable of interacting with body fluids and biological tissues.

Fig. 1. Nonlinear behavior of the networks.

(A) Optical microscope picture of a network, with four input channels and four output channels labeled (scale bar, 100 μm). (B) Sketch of a network with E/I balance with highlighted excitatory and inhibitory nodes. (C) Input signals injected to the four labeled channels and (D) readout of the reservoir states measured at the four output channels. (E) and (F) report the Fourier transforms of (C) and (D), proof that the transfer function of the network is nonlinear: A multitude of new frequencies appear, proving the nonlinear projection performed by the reservoir. (G) Short-term memory featured by the network, key in RC. ΔG is the change in conductance and τ is the response time of the fiber.

Fig. 2. E/I balance of the network at different stages of the network growth.

Output signals read out with a network grown to be (A) totally inhibitory, (B) E/I balanced, and (C) totally excitatory. The input channels are marked with a red “*,” and the outputs are read out at the electrodes marked with a green “+.” As proven by the respective Fourier transforms (bottom row), only a balanced network is able to project the input nonlinearly onto the output layer.

Fig. 3. Heartbeat classification using PEDOT networks.

(A) Schematics of the steps used for carrying out information processing with the networks. The use of the delay line is optional. (B) Confusion plot of the measurement performed with the delay line to classify heartbeats: A = 88%. (C) Reservoir patterned and grown onto a conformable substrate of polyimide. In the inset, a magnification of the network grown on polyimide (photo credit: Matteo Cucchi, Technische Universität Dresden). The different surface does not affect the growth or the lithographic process.