IT201800005512A1 - PROCEDURE FOR TREATING ELECTROPHYSIOLOGICAL SIGNALS, SYSTEM, VEHICLE AND CORRESPONDING IT PRODUCT - Google Patents

Description

DESCRIPTION of the industrial invention entitled: "Procedure for treating electrophysiological signals, corresponding system, vehicle and computer product"

TEXT OF THE DESCRIPTION

Technical field

The description refers to the processing of electrophysiological signals.

One or more embodiments can be applied to the processing of electrophysiological signals such as e.g. electroencephalography (EEG) and / or photoplethysmography (PPG) signals.

One or more embodiments can facilitate the obtaining of information (data, physical quantities) from the living animal or human body, eg. to support the diagnostic activity of a human in medical and veterinary activities or for other possible uses (eg. in the automotive sector).

Technological Background

Photoplethysmography (PPG) is a simple and inexpensive optical technique that can be used to detect changes in blood volume in the microvascular bed of tissue. It is often used non-invasively to make measurements at the surface of the skin.

A PPG waveform comprises a pulsatile physiological waveform ("AC") attributed to cardiac synchronous changes in blood volume with each heartbeat, and is superimposed on a slowly varying baseline ("DC") with various lower frequency components attributed to respiration, thermoregulation, skin tissues, etc.

For each cardiac cycle, the heart pumps blood to the periphery. Although this pressure impulse is somewhat dampened by the time it reaches the skin, this is enough to stretch the arteries and arterioles in the subcutaneous tissue. If a reflected light / transmission detector device is attached to the skin, a pressure pulse can also be seen from the venous plexus, as a small secondary peak. The change in volume caused by the pressure pulse is detected by illuminating the skin with light from a light emitting diode (LED) and subsequently measuring the amount of light either transmitted or reflected to a photodiode. Each cardiac cycle appears as a peak.

Since blood flow to the skin can be modulated through multiple other physiological systems, PPG can also be used to monitor breathing, hypovolemia, circulatory conditions as well as for subjective analysis. In addition, the shape of the PPG waveform differs from subject to subject, and varies with the position and manner in which the pulse oximeter is connected.

The following is a list of exemplary documents of an extensive activity dedicated to topics related to the PPG:

[1] Rohan Banerjee, et al: “Estimation of ECG parameters using photoplethysmography”, 13th IEEE International Conference on BioInformatics and BioEngineering, Year: 2013, Pages: 1-5;

[2] Vala Jeyhani, et al: “Comparison of HRV parameters derived from photoplethysmography and electrocardiography signals”, 2015 37th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Year: 2015, Pages: 5952–5955;

[3] K. Venu Madhav, et al: “Estimation of respiration rate from ECG, BP and PPG signals using empirical mode decomposition”, 2011 IEEE International Instrumentation and Measurement Technology Conference, Year: 2011, Pages: 1–4;

[4] M. Bolanos, et al: “Comparison of heart rate variability signal features derived from electrocardiography and photoplethysmography in healthy individuals”, in: Proc.

28th Annual Int. Conf. Of the IEEE EMBS, New York City, USA, 2006;

[5] N. Selvaraj, et al: “Assessment of heart rate variability derived from finger-tip photoplethysmography as compared to electrocardiography”, Journal of Medical Engineering & Technology 2008, pp. 479-484;

[6] Lu G. et al: “A comparison of photoplethysmography and ECG recording to analyze heart rate variability in healthy subjects”, Journal of Medical Engineering & Technology, vol. 33, 2009, pp. 634-641;

[7] Elgendi M .: “On the analysis of fingertip photoplethysmogram signals”, Current Cardiology Reviews, vol. 8, 2012 pp. 14-25;

[8] Mazomenos E. B .: “A Time-Domain Morphology and Gradient based logarithm for ECG Feature Extraction”, in: Proc: International Conference on Industrial Technology (ICIT), 2012, pp. 117-122; And

[9] J. Vicente and others: "Detection of Driver's Drowsiness by means of HRV Analysis", IEEE Computing in Cardiology 2011; 38: 89-92.

Other documents of interest include:

[10] N. N. Sari, et al: "A two-stage intelligent model to extract features from PPG for drowsiness detection," 2016 International Conference on System Science and Engineering (ICSSE), Puli, 2016, pp. 1-2;

[11] Li, G., et al: “Detection of Driver Drowsiness Using Wavelet Analysis of Heart Rate Variability and a Support Vector Machine Classifier”, Sensors (Basel, Switzerland), 13 (12), 16494–16511;

[12] Yo-Ping Huang, et al: "Early detection of driver drowsiness by WPT and FLFNN models", 2016 IEEE International Conference on Systems, Man, and Cybernetics (SMC), Budapest, 2016, pp. 000463-000468;

[13] Jaewon Lee, et al: "Correlation Analysis between Electrocardiography (ECG) and Photoplethysmogram (PPG) Data for Driver's Drowsiness Detection Using Noise Replacement Method", Procedia Computer Science, Volume 116, 2017, Pages 421-426, ISSN 1877- 0509;

[14] Saravanamoorthi, A., et al: “Prediction of Drowsy Drivers Fault Using Bio Signals Joint Stochastic FSD (BJSFSD) Algorithm”, (2016);

[15] Soltane, M., and others: “Artificial neural networks (ANN) approach to PPG signal classification”, International journal of computing & information sciences, 2 (1), pag.58;

[16] Lawoyin, S., 2014: “Novel technologies for the detection and mitigation of drowsy driving”, Virginia Commonwealth University;

[17] T. Hwang, et al: "Driver drowsiness detection using the in-ear EEG", 2016, 38th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Orlando, FL, 2016, pp. 4646 -4649;

[18] Liu, Y., et al: "Driver Drowsiness Detection System", https://3x10e8.files.wordpress.com/2015/08/ece4781final projectwrite-upteamorpheus.pdf

As noted, the use of PPG can be imagined in areas other than the medical field.

For example, PPG was considered for use in the automotive field, eg. in order to obtain useful information on the behavior and / or reaction of drivers and passengers in various situations that may occur in a vehicle.

In fact, there is extensive activity to address the technical problem of identifying a state of mental attention, eg. a drowsy state of a driver of a vehicle (both before and while driving), using PPG signals and / or other electrophysiological signals.

Purpose and summary

Despite the extensive activity in the area, improved solutions are desirable that facilitate, for example, the identification of a drowsy state of a vehicle driver.

According to one or more embodiments, such an object can be achieved by means of a method having the characteristics set out in the following claims.

One or more embodiments can refer to a corresponding system.

One or more embodiments can refer to a corresponding vehicle, such as, for example, a motor vehicle equipped with such a system.

One or more embodiments may comprise a computer product that can be loaded into the memory of at least one processing circuit (e.g., a computer) and comprising portions of software code for carrying out the steps of the method when the product is performed on at least one processing circuit. treatment. As used herein, it is understood that the reference to such a computer product is equivalent to the reference to a computer readable medium containing instructions for controlling the treatment system in order to coordinate the implementation of the procedure according to one or more embodiments. The reference to "at least one computer" is intended to highlight the eventuality that one or more embodiments are implemented in a modular and / or distributed form.

The claims are an integral part of the technical teaching provided herein with reference to the invention.

One or more embodiments may facilitate electroencephalography (EEG) signal processing (i.e. pattern recognition of EEG samples) applied in a PPG / EEG system comprising, e.g.

- PPG sensors (eg silicon photomultipliers - SiPM for PPG detection);

- a PPG signal pattern recognition method and / or system,

- a sample pattern recognition process and / or system.

One or more embodiments may involve a pipeline configured to process photoplethysmography (PPG) signals based on the use of detectors such as e.g. of silicon photomultiplier detectors (SiPM). Such probe sensors can provide advantages in terms of single photon sensitivity and high internal gain for relatively low reverse bias.

One or more embodiments can adopt (possibly in combination with SiPM detectors) a processing pipeline adapted to correct the signal distortion, for example by filtering and normalizing the signal.

One or more embodiments consequently facilitate the obtaining of information (data, physical quantities) from the living animal or human body, eg. to support the diagnostic activity of a human in medical and veterinary activities or for other possible uses. Obtaining information on the behavior and / or reaction of drivers and passengers in the automotive field is an example of such a possible use.

One or more embodiments may involve processing EEG signals that facilitate efficient segmentation of compliant EEG sample waveforms in a combined PPG / EEG system, which in turn facilitates robust sleep / alert monitoring. of a driver of a vehicle.

One or more embodiments may offer one or more of the following advantages:

- high-speed calculation facilitated by pattern recognition mechanisms based on the LM (Levenberg-Marquardt) algorithm and multi-layer neural network with Motor Map, which can be implemented in dedicated hardware;

- low complexity of data analysis and low CPU consumption;

- accuracy and robustness due to the correlation between PPG and EEG;

- facilitated continuous monitoring of the attention state of a driver of a vehicle;

- possibility of avoiding training algorithms or self-tuning of system parameters;

- simple implementation for the acquisition of EEG / PPG signal, for example, from detectors on the vehicle's steering wheel;

- high sensitivity / specificity ratio (eg.

98% / 98%) as opposed to low complexity design.

- reduction of data buffering requirements.

Brief description of the various views of the drawings One or more embodiments will now be described, by way of non-limiting example only, with reference to the attached Figures, in which:

Figure 1 is an exemplary diagram of a photoplethysmography (PPG) signal;

- Figure 2 is an example of a possible operation of PPG sensors;

Figure 3 is a diagram comprising four portions designated a), b), c) and d) exemplifying the possible behavior of electroencephalography (EEG) signals;

Figure 4 is an exemplary functional diagram of a possible signal processing in the embodiments;

Figure 5 is an exemplary block diagram of a possible arrangement of a portion of the pipeline of Figure 4;

Figures 6 and 7 are exemplary diagrams of the PPG signal processing in the embodiments;

Figure 8 is an exemplary block diagram of a possible arrangement on a portion of the pipeline of Figure 4;

- Figure 9 is an example of the possible treatment of the artificial neural network signal in the embodiments; And

- Figure 10 is an exemplary functional diagram of the possible signal processing in the embodiments.

Detailed description of exemplary embodiments

In the following description, one or more specific details are illustrated, aimed at providing a thorough understanding of examples of embodiments of this disclosure. Embodiments can be obtained without one or more of the specific details, or with other processes, components, materials, etc. In other cases, known structures, materials, or operations are not illustrated or described in detail so that certain aspects of the embodiments will not be made unclear.

Reference to "an embodiment" within the overall framework of the present disclosure is intended to indicate that a particular configuration, structure, or feature described in relation to the embodiment is included in at least one embodiment. Therefore, phrases such as, "in an embodiment" that may be present in one or more points of this description do not necessarily refer to a single and the same embodiment. Furthermore, particular conformations, structures or features can be combined in any suitable way in one or more embodiments.

The references used herein are provided merely for convenience and therefore do not define the scope of protection or the scope of the embodiments.

Figure 1 is an exemplary diagram of a photoplethysmography (PPG) signal.

As exemplified in Figure 1, a typical photoplethysmography signal / waveform (PPG for short) includes:

- a systolic peak SP at a peak value x,

- a DN dichrotic notch,

- a dystolic peak DP at a y value.

A pulse width W can also be defined at a given value of the PPG value.

PPG signals can be detected using PPG sensors / devices (e.g., PD sensor in Figure 2) comprising LED emitters operating at specific wavelengths (usually infrared at 940nm) and silicon or SiPM photomultipliers (see eg M. Mazzillo, and others: “Silicon Photomultiplier technology at STMicroelectronics”, IEEE Trans. Nucl. Sci, vol. 56, nr. 4, pages 2434–2442, 2009).

These SiPMs can have a total area of 4.0x4.5 mm2 and 4871 square micro cells with 60 micron pitch (1 micron = 10-6 m). These devices have a geometric fill factor of 67.4% and are packaged in a surface mount housing (SMD) with 5.1x5.1mm2 total area (see e.g. M. Mazzillo, et al, cited above or M. Mazzillo, et al.: “Electro-optical performances of pon-n and n-on-p silicon photomultipliers”, IEEE Trans. Electron Devices, vol. 59, nr. 12, pp. 3419–3425, 2012).

A Pixelteq dichroic band pass filter with a passband centered at 542 nm with a Half Height (FWHM) of 70 nm (1 nm = 10-9 m) and an optical transmission greater than 90% in the passband range yes can bond on SMD package using Loctite® 352TM adhesive. With the dichroic filter at 3V-OV the SiPM has a maximum detection efficiency of about 29.4% at 565 nm and a PDE of about 27.4% at 540 nm (central wavelength in the band pass filter - 1 nm = 10-9 m). It has been noted that the dichroic filter can reduce the excess absorption of ambient light by 60% in the linear operating range of the detector operating in Geiger mode above its breakdown voltage (~ 27V). OSRAM LT M673 LEDs in the SMD package emit at 529 nm (1 nm = 10-9 m) and based on InGaN technology they have been used as optical light sources in exemplary embodiments. These LEDs have an area of 2.3x1.5mm2, viewing angle of 120 °, spectral bandwidth of 33nm (1nm = 10-9m) and typical power output of a few mW in the standard operating range .

The use of PPG probes including Silicon Photo Multiplier (SiPM) detectors can provide advantages in terms of single photon sensitivity and high internal gain for a relatively low reverse bias.

It was observed that (see for example D. Agrò, and others: "PPG embedded system for blood pressure monitoring," in AEIT Annual Conference - From Research to Industry: The Need for a More Effective Technology Transfer (AEIT), Trieste , 2014), that Silicon Photo Multipliers (SiPM) can provide advantages in PPG systems in terms of superior AC-to-DC ratio in the PPG pulse waveform, high repeatability in units of motion artifacts and environmental interference. One or more embodiments as discussed herein can detect PPG signals using SiPM (as available with ST group companies) as optical probe sensors, adapted for use in conjunction with hardware and software components to provide a signal processing pipeline.

Figure 2 is an example of a possible operation of PPG sensors.

The light emitted by the LEDs can be absorbed by the skin (DC component) and by the arteries, specifically, through oxygenated (and partially deoxygenated) hemoglobin (AC component).

The residual propagated / reflected (backscattered) light can be a function (proportional-differential) of the amount of light absorbed by hemoglobin in the blood in the various phases of the heart (systolic, diastolic, dicrotic, etc.). A SiPM photomultiplier can consequently detect the presence of photons in the propagated / reflected light by transducing an electrical signal that will be sampled by eg. a 24-bit ADC providing the PPG signal accordingly as discussed above.

Such PPG PD sensors can be applied to a vehicle steering device, in various arrangements as shown in Figure 2.

One or more embodiments may take advantage of the ability of PPG's PD sensors to both function in a transmission mode (see e.g. portion a) in Figure 2) which is with radiation from the LED propagating through the body (e.g. e.g. the body of a patient being clinically tested or of a driver), e.g. through a fingertip F, and in a reflection mode (see e.g. portion b) in Figure 2) which is with reflected radiation from the LED (backscattered) from the body, facilitating the reduction of requirements for the possible positioning of PPG's PD sensors / detectors relative to the body.

Electroencephalography (EEG) is an electrophysiological procedure to record the electrical activity of the brain (electrocortical activity or EEG). The acquisition of EEG signals can be performed as a non-invasive act, for example by means of a number of electrodes (e.g., from sixteen to twenty-four) placed on the skull by means of a conductive paste, to facilitate a connection to low resistance.

Scientific literature shows that EEG activity is due to synaptic currents generated by pyramidal cortical neurons, which follow signals arriving from other cortical areas or from the sensory thalamus.

The sum of the activity of a plurality of pyramidal neurons gives rise to detectable EEG signals.

A pyramidal neuron receives various inputs; if these inputs excite a group of adjacent neurons sufficiently simultaneously (synchronization), the EEG activity can present wide and slow waves.

Figure 3 is an exemplary diagram of electroencephalography signals (EEG for short).

As exemplified in Figure 3, an EEG signal can comprise waves of different frequency and amplitude, often called "rhythms" and labeled with Greek letters: α, β, δ, θ. The amplitude variation of these waves is specifically related to:

- physiological events (e.g., sensory stimulation, sleep, etc.)

- pathological (eg, epilepsy, coma, etc.).

In the EEG, each "rhythm" can perform certain frequency and amplitude characteristics. For example: - the β rhythm (portion a) in Figure 3) can have a frequency above 13 Hz and amplitude below 40 μV,

- the rhythm α (portion b) in Figure 3) can have a frequency from 8 to 13 Hz and an amplitude from 40 to 50 μV,

- the rhythm θ (portion c) in Figure 3) can have a frequency from 4 to 7 Hz and an amplitude from 50 to 75 μV,

- the rhythm δ (portion d) in Figure 3) can have a frequency lower than 4 Hz and an amplitude greater than 75 μV.

The synchronization of the EEG “rhythms” therefore reflects the collective behavior of the neurons involved.

As is known from the literature, EEG recordings - for example α and β waves - can be (directly) indicative of a person's sleepy and alert condition, respectively. Other electrophysiological signals, such as electrocardiogram (ECG) signals, are only indirect measurements.

Figure 4 is an example of a possible signal processing pipeline in a system 40 according to embodiments.

In one or more embodiments, a PPG sensor PD (shown in dashed lines, as far as possible as this may represent a distinct element from the embodiments) may be coupled to a first processing circuit block or stage 42 to provide to this a "raw" signal S of untreated PPG.

In one or more embodiments as discussed herein, the PPG signal can be detected in a known manner at a position of the body of a driver D of a vehicle V. For example, the signal can be detected via one or more PPG sensors. (for example, of the type discussed above) arranged in correspondence with a steering wheel SW of vehicle V.

One or more embodiments of the processing step 42 may comprise filtering steps, mathematical analysis steps, and artificial neural network circuits (trained with sample EEG signals / waveforms), the functions of which will be discussed below, also making reference to figures such as Figures 5 to 10.

In one or more embodiments, the exit from the treatment stage 42 can be coupled to the decision stage 44.

In one or more embodiments, the decision stage 44 may comprise neural network circuits and / or comparator circuits, as discussed below.

One or more embodiments of the decision stage 44 can be configured to evaluate a state of the driver of a vehicle D, for example by providing a signal DS indicative of the attention level of the driver D which can be fed to an interface A (for example a display unit, a sound and / or light generator, and so on). This can facilitate, for example, making driver D aware of the reduced level of attention, possibly due to sleepiness or other reasons.

In one or more embodiments, the decision stage 44 may also provide signals (e.g., the DS signal) to an error monitoring stage 46.

In one or more embodiments, the error monitoring stage 46 may in turn operate, via a feedback loop path 48, on the processing stage 42 or on the decision stage 44 as a signal function from the decision stage 44 and / or input from user D (as provided via interface A, for example).

For example, the error monitoring stage 46 can trigger the activation of the feedback loop path 48 to facilitate the retraining of neural network circuits included in the treatment stage 42 and / or in the decision stage 44.

As exemplified in Figure 5, in one or more embodiments, the processing stage 42 may comprise substeps 42a, 42b which may substantially comprise similar artificial neural network circuits and designed to operate different sets of training data and realdrowsy , e_real_wakeful, as discussed below.

Due to the substantial similarity of the sub-stages 42a, 42b, for the sake of brevity, a detailed description will be provided below with reference mainly to the previous one (i.e. 42a), it being otherwise understood that the same description also applies, mutatis mutandis, at the last (i.e. 42b).

At least in principle, the sub-stages 42a, 42b could also be implemented as a single circuit intended to alternatively perform the role of the sub-stage 42a (data set e_real_drowsy) and the role of the sub-stage 42b (data set e_real_wakeful) .

In one or more embodiments, stage 42a may comprise a first filter circuit 420, which receives the "raw" PPG signal S from the PD sensor and which provides a "clean" PPG signal Sclean at a normalization stage 422 and a mathematical analysis stage 800 (see Figure 8).

In order to produce, from a "raw" PPG signal S, detected by the PD sensor, a "clean" Sclean PPG signal that facilitates processing, the first filter stage 420 - and similarly to one or more forms of realization as exemplified here - they can adopt a solution as described in the Italian Patent Application No. 102017000081018 and also discussed in:

- F. Rundo and others: "Progresses towards a Processing Pipeline in Photoplethysmogram (PPG) based on SiPMs", IEEE Proceedings of 23 European Conference on Circuit Theory and Design, Catania (Italy) 4-6 September 2017;

- Rundo, F .; Conoci, S .; Ortis, A .; Battiato, S .: “An Advanced Bio-Inspired PhotoPlethysmoGraphy (PPG) and ECG Pattern Recognition System for Medical Assessment”, Sensors 2018, 18, 405.

Otherwise it will be appreciated that:

- although desirable, such "cleaning" of the PPG signals from the PPG probe section PD may not be mandatory, so that, at least in certain embodiments, the first filtering stage 420 can be dispensed with or at least simplified, e.g. in the form of a filter;

- in one or more embodiments, the PPG signals from the PD probe section of PPG can be "cleaned" by resorting to solutions other than those described in the Italian Patent Application No. 102017000081018 and in the Rundo et al documents discussed above .

In one or more embodiments as exemplified herein, the normalization stage 422 can receive a "clean" PPG Sclean signal and provide a normalized PPG signal to a first downstream artificial neural network circuit (for short, ANN) 424. For example, the normalization stage 422 can process the "clean" PPG Sclean signal at a unit interval [0,1] before segmentation of the PPG signal. The segmented waveform of PPG obtained can be further normalized and resized eg. through a closer algorithm (see for example F. Rundo, and others: "Adaptive Learning for Zooming Digital Images" - ICCE 2007. Digest of Technical Papers. International Conference on Consumer Electronics, 2007) in order to make it comparable (in terms of value and regarding the time axis) with other waveforms of PPG.

In one or more embodiments, the normalized Sclean signal of PPG (photoplethysmography) can be received via the first downstream ANN 424 which can process the signal to provide a "reconstructed" EEG (electroencephalography, electroencephalogram) e_rec signal at second stage ANN 426 and mathematical analysis stage 800.

In one or more embodiments, the first ANN 424 may include a storage area for a collection of sample EEG waveforms e_real.

As noted, in one or more embodiments the substeps 42a and 42b may be provided to operate different e_real_drowsy, e_real_wakeful training data sets.

Consequentially:

- in the substage 42a, the first ANN 424 can store a collection of EEG and_real_drowsy waveforms of drivers --- in a sleepy state;

- in the 42b substage, the first ANN 424 can store a collection of EEG and_real_wakeful waveforms of drivers --- in an alert state.

In one or more embodiments, the "reconstructed" EEG signals e_rec can be calculated in the first ANN's 424 of the two stages 42a, 42b as a function of:

- the PPG signal from the PD probe of PPG, e

- the respective collection of EEG waveforms, i.e. e_real_drowsy (stage 42a) or e_real_wakeful (stage 42b).

This processing will consequently result in two reconstructed EEG signals, namely a “sleepy” e_rec_drowsy EEG signal (stage 42a) and an “alert” e_rec_wakeful EEG signal (in stage 42a, with the e_rec_wakeful signal not visible in Figure 5).

In one or more embodiments the two signals e_rec_drowsy and e_rec_wakeful can be subsequently processed (to some extent, compared) in order to evaluate - for example in block 44) whether the PPG signal S as detected (currently) by the PD probe it is indicative of a "sleepy" or "alert" state of driver D. This can occur, for example as discussed below with reference to Figure 10.

In one or more embodiments, the reconstruction of the EEG signal to provide the two signals e_rec_drowsy and e_rec_wakeful can be performed as exemplified in Figures 6 and 7.

Figure 6 and Figure 7 are examples of diagrams of a possible network topology of an artificial neural network treatment (as exemplified by stage 424 in Figure 5) configured to correlate the PPG signal to the EEG measurements available for training the first (and second) artificial neural network (ANN) 424 stage, resulting in reconstructed EEG signals e_rec (this designation will be used indifferently for e_rec_drowsy and e_rec_wakeful) starting from the PPG signal S.

Training values, e.g., e_real (this designation will again be used indifferently for e_real_drowsy and e_real_wakeful) may include sets of EEG measurements performed on a large sample of subjects (e.g., by conventional means such as a plurality of electrodes on the surface of the head) simultaneously with the detection of a PPG S signal.

These values can be used to train both neural network circuits (eg Levenberg-Marquardt) of Figure 5.

It was found that the Levenberg-Marquardt multilayer perceptron neural network, LM-MLP NN for short, is an adequate tool for use in learning a correlation between PPG signals and EEG samples / waveforms.

The (first) stage ANN 424 will consequently facilitate the reconstruction of a subject's EEG e_rec signals for both sleepy (e_rec_drowsy) and alert (e_rec_wakeful) values, e.g. based on a Levenberg-Marquardt neural network multilayer perceptron (for short, LM MLP NN).

Figure 6 is a diagram of a perceptron multilayer, MLP for short, with an input layer 601, a hidden layer 602 and an output layer 603, having a number of parallel perceptrons n1, n2,…, nO, for each respective layer.

The perceptrons in the layers are coupled to the input node of each neuron of the downstream layer (which can be referred to as a “fully connected advance” topology) and a bias input node. For example, the input layer 601 can receive an array of input values, e.g. I1,…, Ini, and a bias input, eg. 1.

Figure 7 is an example of the topological diagram of a single perceptron, e.g. a perceptron 61 belonging to the input layer 601.

The output layer 603 of the MLP can provide the output with an array of output values O1, ..., Om, ..., Ono whose value can be described by the following equation:

The learning phase, eg. to define the values of the weights associated with the output layer, it can facilitate the minimization of a defined error function:

Levenberg-Marquardt error correction learning is expressed in the following equation:

where the weight vector w is interactively updated by the error vector and modified by the Jacobian matrix J and the scalar m.

For an LM-MLP NN (Levenberg-Marquardt MultiLayer Perceptron Neural Network), the input bias array (e.g., 1) are given the error vector values and the output bias array (e.g., 1) is set to (1-e), thereby ensuring that outliers or outliers in the inputs are scaled down in importance in the output layer.

The first artificial neural network circuit 424 can facilitate the completion of such a reconstruction of the EEG signals for both the sleepy and alert driver training sets.

In one or more embodiments, a mathematical analysis step 800 (visible in Figure 5 and further described in detail in Figure 8) can be configured to process the filtered "clean" PPG Sclean signal and the reconstructed EEG e_rec_drowsy signal for provide a set of vectors, eg, [F, L, E], to decision stage 44.

It will again be recalled that, while it is provided primarily for the sake of brevity with reference to step 42a, the same present description also applies, mutatis mutandis, to step 42b. Accordingly, in one or more embodiments, stage 42b may comprise a respective mathematical analysis stage 800 configured to process the filtered "clean" PPG Sclean signal and the reconstructed EEG e_rec_wakeful signal to provide a respective set of vectors, eg, [F ', L', E '], at decision stage 44.

For example, the mathematical analysis step 800 (both in 42a and 42b) may comprise at least one processor block (e.g., a DSP or similar processor circuit) configured, in a manner known to those skilled in the art (e.g., via software) to perform the mathematical analysis of the "clean" PPG Sclean signal from the PPG PD probe (eg as received - in digital form - by the input) to extract certain characteristics from it to support further processing in the stages of downstream artificial neural network (ANN).

In one or more embodiments, the calculation of the values of the (first) set of vectors [F, L, E] can take place as exemplified in Figure 8, which also applies to the calculation of the values of the (second) set of vectors [F ', L', E '].

In one or more embodiments, the mathematical analysis stage 800 may comprise a first analysis stage 421, a second analysis stage 425, a feature extraction stage 423 and an optional combining stage 427.

In one or more embodiments, the first analysis stage 421 may receive the filtered "clean" PPG Sclean signal from the first filtering stage 420 and provide a first analysis vector L comprising a plurality of values, e.g., characteristics filtered "clean" PPG Sclean signals, eg. L = [Ldia, Lsys, LpeakToPeak].

For example, these characteristics may include: - the length of the sub-curve of the PPG waveform, for the diastolic phase Ldia,

- the length of the sub-curve of the PPG waveform, for the Lsys phase, e

- the length of the sub-curve of the PPG waveform, between two consecutive SP peaks, LpeakToPeak.

The suffixes sys and dia indicate the systolic and diastolic phases of the PPG signal that can be identified, referring to the diagram in Figure 1, as the O-SP (systolic) and DN-DP (diastolic) portions, while the peakToPeak suffix indicates the phase between two consecutive SP peaks.

Similarly, the second analysis stage 425 can receive the reconstructed EEG signals e_rec (again this can be applied to e_rec_drowsy in 42a and to e_rec_wakeful in 42b) from the circuit of the first ANN 424 and provide a second analysis vector E comprising a plurality of values , e.g., statistical characteristics of EEG reconstructed e_rec signals, e.g. E = [µ (e_rec), σ (e_rec), µ (R (e_rec))].

For example, statistical characteristics may include:

- an average value of the reconstructed EEG signal, µ (e_rec),

- one standard deviation of the reconstructed EEG signal, σ (e_rec), e

- an average value of the autocorrelation function of the reconstructed EEG signal, µ (R (e_rec)).

Subsequently, the characteristic extraction stage 423 can be configured to receive as an input of the first analysis vector L (L ') and the second analysis vector E (E') and to provide as an output a characteristic vector F (F ' ), containing a number of mathematical features F1 to F6, e.g. F = [F1, F2, F3, F4, F5, F6], resulting from the received entry treatment.

In one or more embodiments, the characteristics F1 to F6 can be expressed by the following equations:

Basically, the mathematical characteristics F1 to F3 provide an indication of the "length" of the curve or path of the PPG signal signal in the systolic and diastolic phases (ie, that is, "as far as

time "each of these signals remains in each phase) and the" length "from peak to peak, while the characteristics F4 to F6 provide an indication of the statistical characteristics of the reconstructed EEG signal.

In one or more embodiments, optionally, the first analysis vector L (L ') and the characteristic vector F (F') can be combined in a combination stage 427, eg. can be chained together.

As a result, in one or more embodiments: - the first set of vectors [F, L, E] calculated in the mathematical analysis phase 800 in sub-phase 42a will depend on the sample data set e_real_drowsy, and

- the second set of vectors [F ', L', E '] calculated in the mathematical analysis phase 800 in sub-phase 42b will depend on the sample data set e_real_wakeful.

In one or more embodiments, the second artificial Neural Network (ANN) 426 circuit can receive the reconstructed EEG e_rec signal (e_rec_drowsy in 42a and e_rec_wakeful in 42b) from the first ANN 424 circuit and can be configured to process the signal e_rec of EEG reconstructed via an artificial neural network, e.g., such as a neural network with multi-layered motor map.

In one or more embodiments, the artificial neural network of the second ANN 426 circuit can be trained to provide a vector of selected weights U (for 42a) and U '(for 42b), comprising a plurality of selected weights, e.g. , U = [u1,…, u6]; U '= [u1', ..., u6 '] at decision stage 44, as discussed below in relation to Figure 9.

In one or more embodiments, the second ANN 426 circuit may comprise a storage area T for a collection of EEG's e_real waveforms (e_rec_drowsy in 42a and e_rec_wakeful in 42b), similarly to what has been described for the first circuit ANN 424.

For example:

- the second ANN 426 circuit in 42a can store a collection of EEG waveforms of drivers in a drowsy e_real_drowsy state,

- the second ANN 426 circuit in 42b can store a collection of EEG waveforms of drivers in an alert and_real_wakeful state.

The reconstructed EEG signal e_rec (e_rec_drowsy in 42a and e_rec_wakeful in 42b) can therefore be calculated as a function of either of the alternate states of the driver depending on which collection of EEG waveforms between e_real_drowsy and e_real_wakeful is stored in the respective T memories of the ANN circuits 424, 426, resulting in either a reconstructed sleepy EEG signal e_rec_drowsy or a reconstructed awake EEG signal e_rec_wakeful.

In one or more embodiments, the first selected weight vector U in the substage 42a, e.g. U = [u1, ..., u6], can facilitate the calculation of a DLA attention level for a sleepy driver, while the second vector of selected weights U 'in sub-phase 42b, eg. U '= [u1', ..., u6 '], can facilitate the calculation of a level of attention DLA' for an awake driver.

In one or more embodiments, the first and second selected weight vectors U, U 'can be calculated with a similar procedure, as discussed below.

In one or more embodiments, the second ANN 426 circuit (in both 42a and 42b) can implement a motor map neural system, as exemplified in Figure 9.

In one or more embodiments, a winner-take-all algorithm, eg. for Self-Organizing Maps (SOM for short), it may be suitable for training the neural network of the second ANN 426 circuit.

It occurs that an extended SOM (Self-Adapting Map) as schematically exemplified in Figure 9 (where only the input layer IL and the output layer OL are visible) is suitable for use in one or more embodiments.

For example, in one or more embodiments, the second ANN circuit 426 may comprise an input neural layer (e.g. lattice-like, e.g. n * n = n2 neuronal nodes) 900, which can receive signals as input reconstructed EEG e_rec, and a corresponding output layer 990 (e.g. lattice-like, e.g. n * n = n2 neuronal nodes), which can provide at selected weights vector U.

In one or more embodiments, the input layer 910 can comprise the weights win (x, y, t) while the output layer can comprise the weights u (x, y, t).

In one or more embodiments, a random element µ (t) can be comprised in order to improve the learning process.

In fact, from an initial distribution of random weights, and over many iterations, the SOM can be trained to facilitate the supply of an input characteristic map at the exit. The characteristic map can be provided at the output in the form of a vector of selected weights U comprising a plurality of weight values, eg. six values. The plurality of values can be the result of the selection at the output layer 990 of the neural network of a plurality of neuron parameters of Best Matching Units (BMUs). BMU neurons can be those neurons that minimize a distance (based on certain metrics, eg Euclidean) between synaptic weights (eg, win) and EEG samples, as discussed below.

In order to do this, one can train the artificial neural network, eg. its weights win (x, y, t), u (x, y, t) can be determined through an iterative process based on the supply of “real” e_real EEG data sets to the neural network circuit.

For example, the determination of the BMU can involve the iteration through all the nodes (neuronal) and the calculation of the Euclidean distance dNi (i, j) between a weight vector of each node and a current input vector; the node with a weight vector win (x, y, t) closest to the input vector e_rec can be labeled as the BMU.

The Euclidean distance can be given as:

Nodes in the vicinity of BMU nodes (including BMU nodes) can have their weight vector adjusted according to the following equation:

where is it:

- t is the time step,

- α is the learning rate,

- β (x, y, t) is a parameter of the amount of influence that a node distance from the BMU has on its learning, e.g. a classic Gaussian function.

Basically, the new adjusted weight w <Ni> (t + 1) for a neural network node can be equal to the old weight w <Ni> (t), plus a fraction of the difference between the old weight w <Ni> (t ) and the input vector e_rec.

Each of the components, e.g., u1, ..., ui, ..., u6, of the selected weight vector U can vary over time according to the following equation:

A similar notation can be applied to the vector of selected weights U '.

Consequently, the second ANN circuits 426 in the sub-stage 42a and the second ANN 426 circuits in the sub-stage 42b can provide respective selected weight vectors U and U ', each of the vectors comprising six values, e.g., U = [u1, ... , u6], U '= [u1',…, u6 '] indicative of parameters / weights of the output layer of the Self-Adapting Map (SOM) neural network.

Figure 10 is an exemplary flow chart of a possible mode of operation of the decision phase 44.

In one or more embodiments, the decision stage 44, coupled to the processing stage 42, can receive a "union" data set P (see Figures 4 and 5) comprising the first set of vectors [F, U , E, L] from the treatment sub-stage 42a and the second set of vectors [F ', U', E ', L'] from the treatment sub-stage 42b.

In one or more embodiments, the decision stage 44 may be configured to process data sets from the processing stage 42 via neural network processing.

It is again verified that the Levenberg-Marquardt perceptron multilayer neural network, LM-MLP NN for short, provides an adequate tool for that purpose.

In one or more embodiments as exemplified in Figure 10, a selective output layer 440 can utilize a "defuzzification" function, e.g. a Takagi-Sugeno centroid calculated as a weighted sum, which provides a DLA (rec) array of "reconstructed" driver attention level values DLAn1, ..., DLAk, ..., DLAno whose values can be calculated according to the following equation :

where the index k refers to the k-th element of the array and the remaining symbols in the equations are the same as previously defined.

For example, the components of the selected weight vector U can be considered as the membership functions (eg weights of a weighted sum) of the components F F1,…, F6, of the characteristic vector eg. according to a defuzzification function of the Takagi-Sugeno type, for example as described in T. Takagi and M. Sugeno: “Fuzzy Identification of Systems and Its Applications to Modeling and Control,” IEEE Transactions on Systems, Man and Cybernetics, vol. SMC-15, no. 1, pp. 116-132, 1985.

For example, the first vector of selected weights U can be indicative of a first membership function to be applied to the vector of mathematical characteristics F of the reconstructed sleepy EEG signal e_rec_drowsy.

Similarly, the second vector of selected weights U 'can be indicative of a second membership function to be applied to the vector of mathematical characteristics F' of the EEG reconstructed vigilant signal e_rec_wakeful.

At each reconstructed driver attention level value DLAk (rec) calculated from data based on the reconstructed EEG signals e_rec (i.e. e_rec_drowsy and e_rec_wakeful), a corresponding driver attention level value DLAk can be calculated according to the above formula (real), this time based on sample EEG signals e_real.

Next, the calculated DLA reconstructed driver attention level values (rec) can be compared with the corresponding DLA (real) real driver attention level values by calculating an error metric, e.g. distance of the square or quadratic:

where N indicates the number of reconstructed EEG signals used in the training set and the index i identifies each individual EEG sample.

It will be appreciated that an error E verified to be smaller than the error calculated previously can indicate that the system is learning "well" to the extent that it minimizes with quadratic dynamics the average error between the values of the attention level of the reconstructed driver DLAk (rec) and the corresponding “real” ones DLAk (real).

The learning procedure as described can continue until a desired accuracy is achieved or for a fixed amount of time, eg. 3 epochs ("epochs").

The neural nodes of the LM-MLP NN 441 can have a transfer function, e.g. a step function, to be provided as an output to an indicator (classification) DS, which, for example, can facilitate to assess whether the previously calculated level (s) of the driver's attention (or a combination thereof) is below or above a threshold value T, e.g. T = 0.5.

In one or more embodiments, the treatment with artificial neural network in the decision stage 44 can provide the DS indicator to user circuits A, 46.

In one or more embodiments, the DS indicator can have values that vary in a unit range, eg. DS ∈ [0,1].

In one or more embodiments, the DS indicator can be indicative of the driver's attention level as calculated by the neural network circuits 440, 441 of the decision stage 44.

For example, when the DS indicator has values within a first interval, eg. 0 <DS≤0.5, this may be indicative of a "sleepy" state of the driver, while a value of the DS indicator outside this range, eg. DS> 0.5, may be indicative of a driver's "alert" state.

In one or more embodiments, the decision step 44 will consequently facilitate the assessment of a subject's attention level, in a range of states, e.g. from sleepy to alert state.

In one or more embodiments, the DS indicator can be supplied to further treatment units and can be used to trigger an alarm on an interface (see eg. A in Figure 4) such as a display, eg. on the dashboard of the vehicle V.

In one or more embodiments, the indicator DS can be provided to the error monitoring stage 46, which can trigger the activation of the feedback loop path 48 towards the decision stage 44 and / or the processing stage 42 for restart the learning phase of the respective circuits of the neural network 426, 429, 44. For example, the neural network circuits can be retrained to take into account the changes in the dynamics of the received PPG, Sclean S signals.

In one or more embodiments, the feedback loop path 48 can be operated to be activated:

- periodically, according to a safety check schedule planned in the error monitoring stage 46, e.g. Once a month;

- due to a significant deviation (above a certain tolerance threshold) of the measured PPG signals S acquired by the sensor (i) DP from the PPG signals used for the training phase of the neural network, eg. due to a change in the driver;

- due to the wear of / as a consequence of the safety checks carried out on the electronic circuits.

In one or more embodiments, the error monitoring stage 46 can trigger the activation of the branches of the feedback loop path 48 in a certain sequence. For example, said sequence may include:

- trigger the start of a training phase for decision stage 44, eg. retraining of the neural network,

- wait (for example, for its convergence) a desired time interval, for example. a fixed number of eras,

- if the training phase does not converge within the desired time interval, trigger the start of a training phase for the treatment stage 42, in particular for the second ANN 426 circuit.

In one or more embodiments, optionally, it may be possible to iterate the procedure until the training phase for the decision stage 44 converges within the desired time interval.

In one or more embodiments, a method for processing electrophysiological signals may comprise:

- collect a photoplethysmography signal (for example, S), PPG for short, via a PPG sensor (for example, PD),

- treat (e.g., 42, 44; 42a, 42b) the PPG signal, wherein the processing may comprise generating through the treatment of the artificial neural network (e.g., 424, 426, 44) the reconstructed PPG signal a ElectroEncephalogram signal, EEG for short, (e.g., e_rec), the artificial neural network processing of the PPG signal comprising training at least one artificial neural network circuit on a training set of signals (e.g., e_real) produced during the sampling of a sample set of EEG signals, e

- provide the reconstructed EEG signal to a user circuit (for example, 800, 44, A).

In one or more embodiments, the PPG signal processing may comprise at least one of filtering (e.g., 420) and normalization (e.g., 422) of the PPG signal collected by a PPG sensor prior to neural network processing. artificial (e.g., 424, 426) of the PPG signal.

One or more embodiments may include: - detecting (for example, 46, A) a reconstruction error signal (for example, ER) indicative of the accuracy of the reconstructed EEG signals, and

- activating (e.g., 48), as a result of detecting said error signal, the training of said at least one artificial neural network circuit (e.g., 424, 426, 44) on an updated training set (e.g. example, e_real) of signals produced during the sampling of a sample set of EEG signals.

One or more embodiments may comprise triggering the training of said at least one artificial neural network circuit on an updated training set as a result of receiving a retraining trigger signal from at least one of:

- an error monitoring phase (e.g. 46), - a periodic internal trigger generator,

- an alarm interface (for example, A).

In one or more embodiments, the PPG signaling artificial neural network processing may include:

- first treatment of the artificial neural network (eg 424) to map reconstructed EEG signals on PPG signals;

- second treatment of the artificial neural network (for example, 426) of the reconstructed EEG signals (e_rec) mapped on the PPG signals to produce a selected set of output weights (for example, U, U ').

In one or more embodiments:

- the first artificial neural network treatment (e.g., 424) may include a Levenberg-Marquardt multilayer perceptron treatment, and / or

- the second treatment of the artificial neural network (for example, 426) may comprise a treatment with a Self-Adapting Map motor map, in short SOM.

One or more embodiments may comprise fuzzy inference processing, preferably by a Takagi-Sugeno centroid-like fuzzy operator, of said selected set of output weights.

One or more embodiments may comprise generation by artificial neural network processing of the PPG signal:

- a first reconstructed EEG signal (e.g., e_rec_drowsy) as a function of a first signal training set (e.g., e_real_drowsy) produced when sampling a sample set of EEG signals in a first state of mental attention, and - a second reconstructed EEG signal (e.g., e_rec_wakeful) as a function of a second signal training set (e.g., e_real_wakeful) produced when sampling a sample set of EEG signals of a sample set of EEG in a second state of mental attention.

One or more embodiments may include the acts of:

- compute (for example, 440) from said first reconstructed EEG signal (for example, e_rec_drowsy) a first reconstructed attention level indicator (for example, DLA),

- calculate (for example, 440) from said second reconstructed EEG signal (for example, e_rec_wakeful) a second reconstructed level of attention indicator (for example, DLA '),

- calculate (e.g. 440, 441) a resulting attention indicator (e.g., DS) as a combination of the first reconstructed attention level indicator and the second reconstructed attention level indicator,

- comparing (for example, 441) said resulting attention indicator with a threshold value, e

- producing an attention indicator signal (e.g., A) as a function of the result of said comparing act.

One or more embodiments may comprise generating a reconstructed EEG signal via artificial neural network circuits, ANN for short (e.g., 424, 426, 44; 42, 42a, 42b) as a function of the PPG signals (e.g. , S, Sclean) which may comprise training artificial neural network circuits with data sets of EEG signals (e.g., e_real) stored in a memory space (e.g., T).

One or more embodiments may comprise: - collecting said PPG signal from the driver (for example, D) of a vehicle (for example, V) via a PPG sensor on board the vehicle,

- supplying said reconstructed EEG signal to a user circuit (for example, 800, 44, A) on board the vehicle, in which the reconstructed EEG signal can be indicative of a driver's attention level (for example, D ).

A system (for example, 40) according to one or more embodiments can comprise:

- a PPG sensor (for example, PD), configured to collect a photoplethysmography signal, PPG for short, - processing circuitry (for example, 42, 44; 42a, 42b) coupled to the PPG sensor to receive said PPG signal therefrom, the processing circuitry comprising processing circuits of the artificial neural network (for example, 424, 426, 44) and configured to provide a reconstructed EEG signal (for example, e_rec) to a user circuit (for example, 800, 44, A) according to embodiments. One or more embodiments may comprise a vehicle (e.g., V) which can be equipped with a system (e.g., 40) according to embodiments in combination with at least one driver assistance device (e.g., A), the driving assistance device (A) configured to function as a function of said reconstructed EEG signal (e.g., e_rec).

One or more embodiments may comprise a computer product that can be loaded into the memory of at least one processing circuit and comprising portions of software code for carrying out the steps of the method according to embodiments when the product is performed on at least one processing circuit.

It will otherwise be understood that various individual implementation options exemplified for all figures accompanying this description are not necessarily intended to be adopted in the same combinations exemplified in the figures. One or more embodiments can consequently adopt these options individually (otherwise not mandatory) and / or in different combinations with respect to the combination exemplified in the attached figures.

Without prejudice to the underlying principles, the details and embodiments can vary even considerably, with respect to what has been described by way of example only, without departing from the scope of protection. The scope of protection is determined by the attached claims.

Claims (14)

CLAIMS 1. A method for treating electrophysiological signals, the method comprising: - collect a photoplethysmography signal (S), PPG for short, via a PPG sensor (PD), - treat (42, 44; 42a, 42b) the PPG signal (S), in which to treat (42, 44; 42a, 42b) includes generating the PPG signal by treatment with an artificial neural network (424, 426, 44) (S) a reconstructed ElectroEncephalogram (e_rec) signal, EEG for short, the artificial neural network treatment (424, 426, 44) of the PPG signal (S) including training at least one artificial neural network circuit (424, 426, 44) on a set of training signals (e_real) produced during the sampling of a sample set of EEG signals, and - supplying the reconstructed EEG signal (e_rec) to a user circuit (800, 44, A). 2. Process according to claim 1, wherein treating (42, 44; 42a, 42b) the PPG signal (S) comprises at least one of filtering (420) and normalizing (422) the PPG signal (S) collected by means of a PPG sensor (PD) before treatment with artificial neural network (424, 426) of the PPG signal (S). Process according to claim 1 or claim 2, comprising: - detect (46, A) a reconstruction error signal (ER) indicative of the accuracy of the reconstructed EEG signals (e_rec), and - activating (48), as a result of the detection of said error signal (ER), the training of said at least one artificial neural network circuit (424, 426, 44) on an updated set of training signals (e_real) produced during the sampling of a sample set of EEG signals. The method according to claim 3, comprising triggering the training of said at least one artificial neural network circuit (424, 426, 44) on an updated training set (e_real) as a result of receiving a retraining trigger signal by at least one of: - an error monitoring stage (46), - a periodic internal trigger generator, - an alarm interface (A). Method according to any of the preceding claims, wherein the treatment with artificial neural network (424, 426, 44) of the PPG (S) signal comprises: - a first treatment with artificial neural network (424) to map reconstructed (e_rec) EEG signals on PPG (S) signals; - a second treatment with artificial neural network (426) of the reconstructed EEG signals (e_rec) mapped on PPG signals (S) to produce a selected set of output weights (U, U '). The method according to claim 5, wherein: - the first artificial neural network treatment (424) comprises a multi-layered Levenberg-Marquardt perceptron treatment, and / or - the second treatment with artificial neural network (426) includes a treatment with a Self-Adapting Map motor map, in short SOM. 7. Process according to claim 5 or claim 6, comprising a treatment with fuzzy inference, preferably by means of a fuzzy operator like Takagi-Sugeno center, of said selected set (U, U ') of output weights. Method according to any of the preceding claims, comprising generating the PPG (S) signal by treatment with an artificial neural network (424, 426, 44): - a first reconstructed EEG signal (e_rec_drowsy) as a function of a first set of training signals (e_real_drowsy) produced during the sampling of a sample set of EEG signals in a first state of mental attention, and - a second (e_rec_wakeful) reconstructed EEG signal as a function of a second set of training signals (e_real_wakeful) produced during the sampling of a sample set of EEG signals in a second state of mental attention. 9. Process according to claim 8, comprising the acts of: - calculate (440) from said first reconstructed EEG signal (e_rec_drowsy) a first reconstructed attention level indicator (DLA), - calculate (440) from said second reconstructed EEG signal (e_rec_wakeful) a reconstructed second level of attention indicator (DLA '), - calculate (440, 441) a resulting attention indicator (DS) as a combination between the first reconstructed attention level indicator (DLA) and the second reconstructed attention level indicator (DLA '), - compare (441) said resulting attention indicator (DS) with a threshold value, e - producing an attention indicator signal (A) as a function of the result of said comparing act. Method according to any of the preceding claims, wherein generating a reconstructed EEG signal (e_rec) via artificial neural network circuits, in short ANN (424, 426, 44; 42, 42a, 42b) as a function of the signals (S , Sclean) of PPG comprises training artificial neural network circuits with data sets of EEG signals (e_real) stored in a memory space (T). 11. Process according to any of the preceding claims, comprising: - collecting said PPG signal (S) from the driver (D) of a vehicle (V) via a PPG sensor (PD) on board the vehicle, - supplying said reconstructed EEG signal (e_rec) to a user circuit (800, 44, A) on board the vehicle, in which the reconstructed EEG signal (e_rec) is indicative of a driver's attention level (D). 12. System (40), comprising: - a PPG sensor (PD), configured to collect a photoplethysmography (S) signal, PPG for short, - processing circuitry (42, 44; 42a, 42b) coupled to the PPG sensor (PD) to receive said PPG signal (S) therefrom, the processing circuitry comprising processing circuits with artificial neural network (424, 426, 44 ) and configured to provide a reconstructed EEG signal (e_rec) to a user circuit (800, 44, A) with the method according to any of claims 1 to 11. Vehicle (V) equipped with a system (40) according to claim 12 in combination with at least one driving assistance device (A), the driving assistance device (A) configured to function as a function of said reconstructed signal of EEG (e_rec). Computer product that can be loaded into the memory of at least one processing circuit and comprising portions of software code for carrying out the steps of the method according to any of claims 1 to 11 when the product is performed on at least one processing circuit.