Patent Application
20240337619
The present invention relates to a neurocomputational electrochemical sensing device, such as a transistor-based sensing device, for predicting dynamic (bio) chemical properties (i.e. dynamic properties which are chemical and that may relate to processes occurring within living organisms) of a substance. The device comprises an artificial neural network processor that is configured to be trained to compensate for sensor drift and cross-sensitivities. The invention equally relates to a method of operating the sensing device.
Conceived to sense and qualify biological and chemical information beyond human capabilities, biochemical analytical systems are in rising demand to assist societies in applications like health, food safety, and environmental protection. The need to prevent and control critical situations in situ/vivo and in real time has pushed transduction into the nanoscale, packing multiple sensors into miniaturised, fast lab-on-a-chip designs. As sensors' density increases, their outputs grow in redundancy, complementariness, and dimensionality. However, with downscaling, noise increases while stability decreases. Such effects add to selectivity restraints long endured by chemical detectors, impairing the reliability of latest microanalytical integrations.
Among sensor alternatives, microelectrodes and field-effect transistors (FETs) fabricated in semiconductor technologies present advantages such as miniaturisation, mass fabrication, and easy integration with readout electronic circuitry that make them particularly suitable for incorporation into arrays for on-site measurements. Even though these devices are sensitive enough for a wide range of healthcare (e.g. bio-fluid analysis), environmental (e.g. water quality monitoring), and industrial (e.g. food quality control) applications, their commercial application in such settings is conditioned on aging and cross-sensitivities that can severely impair the reliability of the device. Aging is mainly caused by fouling and drifts due either to substance migration into or out of the sensing surface, or to surface hydration. Cross-sensitivity to temperature and interfering species can also constrain the reliability of the measurements, especially when the sensors operate in the field, out of a controlled experimental setup.
In this context, chemometrics is a promising tool to boost accuracy of the measurements and to translate them into actual biochemical knowledge discovery. By modelling the underlying relationships and structure between the measured properties of the chemical system, cross-sensitivity in sensor measurements can be exploited to calibrate for non-idealities, such as temperature, aging or contamination by other chemical species. These same models can be further used to predict complex biochemical parameters, to classify future samples, and to identify critical episodes. Classical chemometric solutions run analytics offline, trained on data batches containing steady-state responses or complete measurement cycles of the sensor array. In order to facilitate machine learning, input data is usually transformed into an intermediate space by means of unsupervised (e.g principal component analysis (PCA), self-organising maps, t-distributed stochastic neighbour embedding) or supervised (e.g. partial least squares) approaches. Linear and non-linear algorithms, such as discriminant analysis, k-means, k-nearest neighbours, support vector machines and random forests, run in a succeeding step to predict categorical or real-valued variables in the big biochemical data. Deep artificial neural networks (ANNs) provide scalable multilayer representations to fit the complexity of the problem in a single model, at the cost of strong computational requirements.
Most chemometrics, and especially deep ANN solutions, run aside on high-speed clocked, synchronous computers. Because most current integrated measuring devices read out the level of every analyte separately, storage and processing bottlenecks arise at the posterior processing stages. This results in higher latencies, higher energy consumption and limited miniaturisation.
On- and off-chip compensation methods have been proposed to overcome some of the above challenges to some extent. However, none of the former approaches addresses, for example, sensor cross-sensitivity to chemical and thermal interferents on-chip when monitoring continuous readings from the sensor array. Thus, the wide adoption of these analytical systems for industrial applications is still limited mainly due to sensor non-idealities, and due to the restricted deployability of these devices in terms of compactness and energy autonomy.
Current online (bio) chemical monitoring devices are generally expensive, bulky, non-predictive, limited to a fix set of a few properties, and require periodical calibration in front of environmental changes. Measurements not feasible with such specific equipment are conducted in laboratories with high associated costs and delays. If predictive models are used, then they are deployed in the cloud. This translates into privacy concerns and, again, into latencies due to data transmission, and additional storage and processing bottlenecks that accentuate as the volume of information accumulates over time.
US2020/138344 A1 discloses an electrochemical sensor comprising a counter electrode, optionally a reference electrode, and an array of multiple working electrodes, wherein at least one of the working electrodes is a film-coated electrode, and wherein the film-forming material has repeat unit that comprises six-membered non-aromatic ring. A device for electrochemical detection that comprises the sensor (an electrochemical tongue) and a method for detecting analytes with the aid of the sensor are also disclosed.
US2018/217086 A1 discloses a three-electrode mixed-potential electrochemical sensor coupled with an artificial neural network data analysis approach, which can extract concentrations from voltages and identify gas streams consisting of single and binary mixtures of NO2, NO, CO, and C3H8. By using the data from the sensors in biased and unbiased mode, single and binary mixtures can be identified with >98% accuracy. The sensor is suitable for control and monitoring of diesel and gasoline engines, turbines, steam power plants, and other combustion technologies.
Publication “On the feasibility to apply a neural network processor for analyzing a gas response of a multisensor microarray” by A. A. Maschenko et al. describes an effort to implement a hardware neural network processor to carry out pattern recognition of signals generated by a multisensor microarray of electronic-nose type.
The objective of the present invention is thus to overcome at least some of the above limitations relating to bio (chemical) analytical systems.
According to a first aspect of the invention, there is provided a neurocomputational electrochemical sensing device for predicting properties of a substance as recited in claim 1.
The proposed device integrates a plurality of electrochemical sensors (e.g. solid-state electrochemical microsensors), which are sensitive to attributes advantageous to predict a set of properties of interest of the sensed substance, and which operate together with a readout circuit and an ANN processor. The readout circuit biases the sensors and converts their responses into signals conditioned to facilitate subsequent ANN processing into predictions of (bio) chemical properties of interest. The sensors, the readout circuit, and/or the ANN processor can be integrated in a common system substrate (e.g. silicon). The electrochemical sensing device may also integrate automated, miniaturised fluidics (e.g. micro-/nano-fluidics) to achieve precise manipulation and control of fluids and high-throughput analysis with low volume of fluids, chemicals, and reagents.
The proposed electrochemical-sensing analytical device provides at least some of the following advantages:
Reliability: The invention provides self-calibrated correction of multiple sensor non-idealities (e.g. aging, cross-sensitivities) in order to lengthen sensor lifetime when operating outside of controlled laboratory settings.
Miniaturisation: solid-state electrochemical (e.g. field-effect transistor (FET), microelectrodes) microsensors and processing circuits can be easily integrated in standard complementary metal-oxide-semiconductor (CMOS) very large-scale integration (VLSI) technologies.
Low cost: The present invention enables solid-state electrochemical monitoring, designed for cost-effective mass manufacturability at wafer level using standard CMOS electronic integration technologies.
Energy efficiency: The device is configured to carry out efficient, local readout encoding as early relevant features to facilitate prediction of target attributes. Signal processing in continuous time is also enabled to avoid redundant information and computation cycles, according to the dynamics of input signals and neural network, not at a fixed frequency provided by an external clock.
The present invention thus solves the technical problem of fabricating multiparametric analytical devices capable of precise sensing, capable of operating in tiny/mobile locations and in real time, and capable of predicting multiple ad-hoc parameters at low maintenance costs.
According to a second aspect of the invention, there is provided a method of operating the device as recited in claim 13.
Other aspects of the invention are recited in the dependent claims attached hereto.
Other features and advantages of the invention will become apparent from the following description of non-limiting example embodiments, with reference to the appended drawings, in which:
Some embodiments of the present invention will now be described in detail with reference to the attached figures. The non-limiting embodiments are described in the context of an electrochemical analysis (e.g. environmental/sanitary monitoring, health assessment from markers in biofluids, chemical process monitoring) of a substance, which in the present embodiments is a liquid. Identical or corresponding functional and structural elements which appear in different drawings are assigned the same reference numerals. It is to be noted that the use of words “first” and “second” may not imply any kind of particular order or hierarchy unless such order or hierarchy is explicitly or implicitly made clear in the context.
The sensor array 3 comprises a plurality of sensors that, together with the readout circuit 5, perform as an artificial neural network pre-processor 11 mimicking the functioning or operation of (artificial) synapses and neurons. In
Each sensor may be, to a certain extent, selective (i.e. mainly or predominantly sensitive) to a given property of a substance, but may also present some degree of cross-response to other properties sensed in the array. In
The artificial neural network processor 7 and the artificial neural network pre-processor 11 comprise two types of parts: multiple-input, single-output neurons N, connected by single-input, single-output synapses W, where the synapses have a direction associated with them. Readout input synapses 13 can be understood to be multi-property single-input, single-output synapses. All the neurons and synapses may comprise one or more integrator compartments, which are considered to integrate their inputs obeying dynamic equations (e.g. they act as leaky integrator compartments). Thus, in the present description, the synapses and neurons are configured to process their input signals to generate output signals.
Accordingly, when a signal passes through synapses and/or neurons, the resulting output signal may be considered to be a processed input signal. However, the verb “process” is to be understood broadly in the present description. In other words, the meaning of this word is understood to encompass any kind of operation applied to a signal. The neurons may transmit their outputs in the digital domain by performing analogue-to-digital conversion through a spike generator (i.e. asynchronous pulse modulation), or in the analogue domain by an analogue representation of their instantaneous firing rate. For the former purpose, the neurons may comprise spike generators 19 as shown in
In the neurons and synapses of the artificial neural network processor 7 and the artificial neural network pre-processor 11, learning can be automated by using optimisation methods like gradient descent that minimise a loss computed from a loss function (e.g. Kullback-Leibler divergence, or possible simplified functional equivalents such as cross-entropy, negative log-likelihood, or (root) mean squared error) representing how well observed properties are modelled. Upon generating the artificial neural network processor output signals ŷ(t), sensor drift may be compensated by modelling the temporal evolution of some given teaching signals y(t)—representing the signals to predict by the ANN processor—with respect to sensor output signals subject to temporal drift; cross-sensitivities may be compensated by modelling the underlying temporal relationships between (i) output signals from sensors selective to properties correlated with the desired properties to predict, (ii) output signals from sensors sensitive primarily to the main interferents, and/or (iii) the teaching signals y(t). The teaching signals may be used together with the output signals from the sensors sensitive primarily to the main interferents (ii) when the teaching signals (iii) are available. It is to be noted that the sensor output signals are also used for training, but they are not referred to as teaching signals in the present description. The teachings signals in the present description are only the signals teaching how the outputs should evolve in time. An embodiment of the invention may train neurons and synapses on a dataset collected offline via backpropagation to compute the gradients (or the surrogate gradients in the case of employing non-differentiable spikes as neuron outputs) of the loss function with respect to the circuit parameters. Both parts (i.e. the neurons and synapses) may be fine-trained online by using local plasticity rules to learn their corresponding parameters. For instance, any of the parameters modifying the overall synaptic weights in W could be learned combining homosynaptic Hebbian and heterosynaptic non-Hebbian rules. In periods when the temporal data of the output properties to predict is made available via teaching signals y(t), local plasticity rules could additionally be gated by error signals indicative of the contributions of each dynamical part to a prediction error depending on a difference between the teaching signals y(t), and the predicted ANN processor output signals ŷ (t). Local plasticity rules update the parameters of synapses and neurons locally in space and time according to mathematical equations or learning rules which only use variables available in the neurons (e.g. integrated variables equivalent to postsynaptic currents or membrane potentials in biological neurons, pre- and post-synaptic traces, error signals, etc.). In this case, error backpropagation can be approximated by including biologically plausible neural networks with both feedforward and feedback synapses that allow the prediction errors made by the network to propagate in space and time through the network in the form of the aforementioned error signals.
As can be seen in
The configuration of
The configuration of
The configuration of
Like the ANN pre-processor, the ANN processor feedforward synapses 33 and the ANN processor recurrent synapses 35 comprise one or more integrator compartments, which each are composed of an integrator filter element 25 with gain A 27, and optionally also summation elements 29. The ANN processor neurons 31, 37 also comprise one or more integrator compartments, which each are composed of a summation element 29 to sum up the neural input signals and/or neuron biases, an integrator filter element 25 with gain A 27, and optionally also of a spike generator 19 and/or additional summation elements 29. In the present embodiments, the compartment of the ANN processor neurons 31, 37 always includes a summation element 29 to sum up the neural input signals.
The artificial neural network processor 7 may thus be one of the following networks: a multi-layer perceptron, a convolutional neural network, a causal convolutional neural network, a dilated convolutional neural network, a recurrent neural network, or a combination of a convolutional neural network and a recurrent neural network. The output of at least one layer of the artificial neural network processor 7 may skip one or more following layers in the artificial neural network processor 7, and the output may be directly connected to the input of at least another succeeding layer of the artificial neural network processor 7 using skip feedforward connections or synapses. The readout circuit 5, on the other hand, may be understood as an artificial neural network layer according to any one of the following networks: a multi-layer perceptron, a convolutional neural network, a causal convolutional neural network, a dilated convolutional neural network, or a recurrent neural network.
The process of operating the neurocomputational electrochemical sensing device is next explained with reference to the flow chart of
The proposed device is particularly advantageous to correct sensor non-idealities and to predict (bio) chemical properties of a substance using multimodal raw measurements collected continuously from a sensor array with sensors selective to properties correlated with target properties to predict as per resemblance of the teaching signals y(t), and with sensors sensitive primarily to the main interfering analytes. The teaching signals y(t) can be used to calculate a prediction loss jointly with the predicted ANN processor output signals ŷ(t). From a prediction loss including a prediction error term depending on a difference between the teaching signals y(t) and the predicted ANN processor output signals ŷ(t), error backpropagation could be approximated by including both feedforward and feedback synapses. Such synapses would allow information about the prediction errors made by the network to propagate through the network in the form of the error signals. Drift and matrix effects severely impair both performance and lifetime of electrochemical sensors such as FETs, but according to the present invention, an artificial neural network can be embedded in the device to cross-compensate such effects by fusing interrelated readout data from multiple sensors.
The proposed device may serve as a portable, self-calibrated electrochemical monitor to provide cost-effective and reliable assessment of analytes detectable by FET-based sensors and/or microelectrode sensors, for instance. In some embodiments, the device comprises an array of FET sensors with electrolyte-gate interfaces functionalised to be selective to properties correlated with the analytes of interest, or sensitive primarily to the main interferents of the measurements. All sensors are inherently sensitive to temperature due to the joint electrochemical and solid-state nature of these devices. The sensor array is coupled to an ANN trainable with sensor readings collected in the field to model both the temporal characteristics of sensor responses and the underlying relationships between sensor input data to predict analyte concentrations while compensating for sensor drift and matrix effects. The artificial neural network processor synapses 33, 35 and/or the artificial neural network pre-processor synapses 13, 17 form a first-type dynamical part, and the artificial neural network processor neurons 31, 37 and/or the artificial neural network pre-processor neurons 15 form a different, second-type dynamical part. Instances of the first-type dynamical part are configured as single-input, single-output synapses, and instances of the second-type dynamical part are configured as multiple-input, single-output neurons. The artificial neural network processor synapses 33, 35 and/or the artificial neural network pre-processor synapses 13, 17 and/or the artificial neural network processor neurons 31, 37 and/or the artificial neural network pre-processor neurons 15 comprise one or more integrator compartments, the synapses and the neurons forming a directed neural network exhibiting continuous-time dynamical behaviour.
While the invention has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive, the invention being not limited to the disclosed embodiments. Other embodiments and variants are understood, and can be achieved by those skilled in the art when carrying out the claimed invention, based on a study of the drawings, the disclosure and the appended claims. Further variants of the invention may be obtained by combining any of the above teachings.
In the claims, the word “comprising” does not exclude other elements or steps, and the indefinite article “a” or “an” does not exclude a plurality. The mere fact that different features are recited in mutually different dependent claims does not indicate that a combination of these features cannot be advantageously used. Any reference signs in the claims should not be construed as limiting the scope of the invention.
| Number | Date | Country | Kind |
|---|---|---|---|
| 21188804.5 | Jul 2021 | EP | regional |
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/EP2022/069999 | 7/18/2022 | WO |
