Closed-loop neuromodulation / neurostimulation system for enabling a body function and / or treatment of a body condition of a mammal

Abstract

The present invention relates to a system (100), especially a closed-loop system, for planning and / or providing neurostimulation / neuromodulation for enabling a body function, especially a motor function and / or autonomic function, and / or treatment of a body condition of a mammal, especially a human patient (P).

Description

[0001] Closed-loop neuromodulation / neurostimulation system for enabling a body function and / or treatment of a body condition of a mammal

[0002] The present invention belongs to the technical field of neuromodulation / neurostimulation, in particular for enabling a body function and / or treatment of a body condition of a mammal, preferably a human patient.

[0003] Said body function can be a motor function, e.g. involving the upper and / or lower limbs, and / or an autonomic function of the patient.

[0004] Said body condition can be, e.g., chronic pain.

[0005] Neuromodulation / neurostimulation techniques use one or several patterns of electrical stimulation applied to the nervous system of a subject.

[0006] One example is the use of spinal cord stimulation to alleviate symptoms of neuromotor disorders such as, but not limited to, spinal cord injury (SCI), stroke, and / or Parkinson’s disease.

[0007] As a non-limiting example, Epidural Electrical Stimulation (EES) may be used to restore a lower-limb motor function of a patient after SCI or other neurological disorders.

[0008] An example of use of EES for the purpose of restoring a lower-limb motor function in a patient is illustrated in Figs. 1a-d.

[0009] Here, EES is delivered through a paddle lead placed over the lumbosacral segments of the spinal cord. Spinal roots innervate the lower-limb muscles in a person-specific rostro-caudal organization (Fig. 1a).

[0010] Fig. 1 b shows a library of optimized anode and cathode configurations and stimulation frequencies to modulate motor pools associated with key phases of a gait sequence.

[0011] A sequence of muscle activity underlying walking in healthy people, converted into a spatiotemporal map of motor neuron activity that highlights the timing and location of motor hotspot activation, is translated into a pre-programmed sequence of stimulation bursts aiming at reproducing this activation pattern (Fig. 1c).

[0012] The configurations of the electrodes targeting each hotspot are derived from the library and injected into this template (Fig. 1d). SCI may lead to severe locomotor deficits or even complete leg paralysis.

[0013] A clinical trial carried out on individuals who sustained a spinal cord injury more than four years before and are presenting permanent motor deficits or complete paralysis despite extensive rehabilitation, demonstrated that delivery of targeted spinal cord stimulation enabled voluntary control of walking. In particular, spatially-selective stimulation was delivered to the lumbosacral spinal cord with timing that coincided with the intended movement by using an Implantable Pulse Generator (IPG) with real-time triggering capabilities. Within one week, this spatiotemporal stimulation had re-established adaptive control of paralyzed muscles during over-ground walking. Locomotor performance improved during rehabilitation. After a few months, participants regained voluntary control over previously paralyzed muscles without stimulation and could walk or cycle in ecological settings during spatiotemporal stimulation [1],

[0014] An experimental study demonstrated that EES along with real-time processing of gait kinematics are powerful methods for the study of locomotion and the improvement of motor control after SCI and / or other neurological disorders. In particular, the study describes equipment and surgical procedures that can be used to acquire chronic electromyographic (EMG) recordings from leg muscles and implant targeted spinal cord stimulation systems that remain stable up to several months after implantation in rats and non-human primates. Also, it is explained how to exploit these implants to configure electrical spinal cord stimulation policies that allow control over the degree of extension and flexion of each leg during locomotion. This protocol uses real-time processing of gait kinematics and locomotor performance, and can be configured within a few days. Once configured, stimulation bursts are delivered over specific spinal cord locations with precise timing that reproduces the natural spatiotemporal activation of motoneurons during locomotion. These protocols can also be easily adapted for safe implantation of systems in the vicinity of the spinal cord and conduct experiments involving real-time movement feedback and closed-loop controllers [2],

[0015] Another clinical trial demonstrated effectiveness of EES targeting the dorsal roots of lumbosacral segments for restoration of walking in patients with SCI. Previously, EES was delivered using multielectrode paddle leads that were originally designed to target the dorsal column of the spinal cord instead of the lumbosacral dorsal root entry zones. In the context of this clinical trial, it was hypothesized that an arrangement of electrodes targeting the ensemble of dorsal roots involved in leg and trunk movements would result in superior efficacy, restoring more diverse motor activities after the most severe SCI. To test this hypothesis, a computational framework was established, that informed on the optimal arrangement of electrodes on a new paddle lead, wider and longer than the ones previously used, and guided its neurosurgical positioning. Also, a software means was developed, supporting the rapid configuration of activity-specific stimulation programs reproducing the natural activation of motor neurons underlying each activity. These neurotechnologies were tested in three individuals with complete sensorimotor paralysis as part of an ongoing clinical trial (www.clinicaltrials.gov identifier NCT02936453). Within a single day, activity-specific stimulation programs enabled these three individuals to stand, walk, cycle, swim and control trunk movements. Neurorehabilitation mediated sufficient improvement to restore these activities in community settings, opening a realistic path to support everyday mobility with EES in patients with SCI [3],

[0016] A further experimental study carried out on a rodent model, in particular on rats, emphasized that, although EES has shown promise for spinal cord injury therapy, finding effective stimuli on the employed multi-electrode stimulating arrays required a laborious manual search of a vast space for each patient. Here, the issue was addressed by the provision of a method based on Gaussian Process Batch Upper Confidence Bound (GP-BUCB), i.e. a Gaussian Process bandit algorithm. Epidural electrode arrays were implanted in n = 4 spinally transected rats, and the algorithm's performance in selecting bipolar stimuli to elicit specified muscle responses was examined. These responses were compared with temporally interleaved intra-animal stimulus selections by a human expert. GP-BUCB successfully controlled the spinal electrostimulation preparation in 37 testing sessions, selecting 670 stimuli. These sessions included sustained autonomous operations (ten-session duration). Delivered performance with respect to the specified metric was as good as or better than that of the human expert. Despite receiving no information as to anatomically likely locations of effective stimuli, GP-BUCB also consistently discovered such a pattern. Further, GP-BUCB was able to extrapolate from results of previous sessions to make predictions about performance in new testing sessions, while remaining sufficiently flexible to capture temporal variability [4],

[0017] Still further, a recent experimental study, carried out on rodents and monkeys, investigated novel, high-density neural interfaces enabling refined and multipronged neurostimulation interventions, in particular to alleviate paralysis and sensory deficits. This required the development of algorithmic frameworks capable of handling optimization in large parameter spaces. Here, Gaussian-process (GP)-based Bayesian optimization (BO) was used. The study revealed that GP-BO efficiently explores the neurostimulation space, outperforming other search strategies after testing only a fraction of the possible combinations. Through a series of real-time multi-dimensional neurostimulation experiments, optimization across diverse biological targets (brain, spinal cord), animal models (rats, non-human primates), in healthy subjects, and in neuroprosthetic intervention after injury, was demonstrated for both immediate and continual learning over multiple sessions. GP-BO can embed and improve prior expert / clinical knowledge to dramatically enhance its performance. These results advocate for broader establishment of learning agents as structural elements of neuroprosthetic design, enabling personalization and maximization of therapeutic effectiveness.

[0018] WO2018 / 129280A1 discloses a method including using a dueling bandits algorithm with correlation among stimulation arms to select a batch of stimulation arms for sequential application to a patient during a therapy session. Each of the stimulation arms specifies complex stimulation waveform parameter values. Feedback from applying the stimulation arms to a patient is recorded and used to update feedback reward values corresponding to at least some of the stimulation arms using a stimulation arm correlation index. A second batch of stimulation arms is selected based upon the updated feedback reward values and applied to a patient. The method is iteratively repeated over a number of therapy sessions until an optimal or near optimal batch of stimulation arms (defining complex stimulation waveforms) is determined.

[0019] The use of correlational dueling bandits in clinical treatment was also explored in Sui, Y, et al., J.W., 2017. Correlational dueling bandits with application to clinical treatment in large decision spaces. arXiv preprint arXiv: 1707.02375 [6],

[0020] In this study, sequential decision making under uncertainty was considered, the goal being to optimize over a large decision space using noisy comparative feedback. In detail, this problem was formulated as a K-armed Dueling Bandits problem where K is the total number of decisions. When K is very large, existing dueling bandits algorithms suffer huge cumulative regret before converging on the optimal arm. The study considered the dueling bandits problem with a large number of arms that exhibit a low-dimensional correlation structure. The problem was motivated by a clinical decision-making process in large decision space. An efficient algorithm CorrDuel was proposed, allowing optimizing the exploration / exploitation trade-off in this large decision space of clinical treatments. Regret bounds were derived, and performance in simulation experiments as well as on a live clinical trial of therapeutic spinal cord stimulation were evaluated.

[0021] The above-mentioned patent application WO2018 / 129280A1 provides a method allowing identifying complex waveforms for a neurostimulator device.

[0022] However, the method described therein is only adapted to be implemented for optimization of electrode configuration while doing an exhaustive search for proper frequency and amplitude, and is based on a dueling approach with patient and / or medical feedback. However, the method known from WO2018 / 129280A1 does not contemplate developing a proper objective function.

[0023] US11413459B2 discloses a system for planning and / or providing neurostimulation for a patient, the system including a pathological spinal cord map storage module for storing at least one pathological spinal cord map describing activation of a spinal cord of a patient; a healthy spinal cord map storage module for storing at least one reference map describing physiological activation of the spinal cord of at least one healthy subject; an analysis module for generating a deviation map, the deviation map describing an activation difference between the pathological spinal cord map and the reference map; and a compensation module for calculating, based on the deviation map, a neurostimulation protocol for compensating the activation difference.

[0024] US10981004B2 discloses systems for electrical neurostimulation of the spinal cord of a subject in need of nervous system function support. In one example, the system comprises a signal processing device configured to receive signals from the subject and operate signalprocessing algorithms to elaborate stimulation parameter settings; one or more multi-electrode arrays suitable to cover a portion of the spinal cord of the subject; and an IPG configured to receive the stimulation parameter settings from the signal processing device and simultaneously deliver independent current or voltage pulses to the one or more multiple electrode arrays, wherein the independent current or voltage pulses provide multipolar spatiotemporal stimulation of spinal circuits and / or dorsal roots.

[0025] That is, US11413459B2 and US10981004B2 provide systems allowing targeting specific locations of the spinal cord to optimize modulatory outcomes for various motor tasks.

[0026] However, both US11413459B2 and US10981004B2 still fail to provide a solution allowing achieving a robust and flexible (semi-)automated optimization of neurostimulation parameters for enabling a body function, e.g. a motor function.

[0027] Also, the techniques described therein are limited to spinal cord stimulation, while cannot be adapted to other neurostimulation techniques.

[0028] The concept of automated Gaussian-process-based optimization of a limited set of stimulation parameters for targeted single-joint movements is generally explained in a presentation illustrated at the Society for Neuroscience (Chicago 2020) (Fig. 2).

[0029] The same concept of implementing Gaussian processes for automatically optimizing electrical stimulation of the spinal cord is mentioned in Abranches de Carvalho, P., Gaussian Processes of Spinal Electrical Stimulation, Master Project Report, Departamento de Ciencia de Computadores, Faculdade de Ciencias da Universidade do Porto, 2019 [7],

[0030] However, this reference [7] as well as the corresponding content illustrated in the presentation of Fig. 2 are almost entirely focused on showing plausibility of the illustrated approach, while do not provide any disclosure on how to build a full clinical framework.

[0031] In particular, this known approach only includes a broad disclosure on the use of a Gaussian process in order to optimize stimulation parameters.

[0032] However, no details are provided regarding further, clinically-relevant aspects, e.g. how to define a relevant reward function, how to deal with and combine different stimulation parameters, and / or how to handle amplitude parameters in a robust and efficient way.

[0033] In the context of a clinical trial (www.clinicaltrials.gov identifier NCT02936453) carried out on a plurality of human patients with SCI, participants were implanted with an electrode array over the lumbosacral region of the spinal cord, to enable locomotor activities.

[0034] All participants regained the ability to perform multiple motor activities with stimulation turned on.

[0035] However, the clinical trial revealed that optimal parameter values are highly patient-specific and that individual parameters have complex interactions.

[0036] This generally requires the presence of highly-trained personnel, which limits the potential application of this therapy on a large scale.

[0037] Therefore, there is the need for an improved system that allows obtaining (semi-)automated personalization and optimization of neuromodulation / neurostimulation parameters, without requiring the intervention of highly-trained personnel, thereby enabling wide-spread applicability of neuromodulation therapy.

[0038] It is an object of the present invention to provide a neuromodulation / neurostimulation system, especially a closed-loop system, for enabling a body function, especially a motor function and / or autonomic function, and / or treatment of a body condition of a mammal, especially a human patient, allowing obtaining (semi-)automated personalization and optimization of neuromodulation / neurostimulation parameters, without requiring intervention of highly-trained personnel. The present invention relates to a system, especially a closed-loop system, for planning and / or providing neurostimulation / neuromodulation for enabling a body function, especially a motor function and / or autonomic function, and / or treatment of a body condition of a mammal, especially a human patient, said system comprising:

[0039] - a stimulation module, configured and adapted to deliver stimulation to said patient according to one or more stimulation parameters through one or more electrodes;

[0040] - a learning module configured and adapted to:

[0041] (i) receive a user command for a selected stimulation function;

[0042] (ii) define a stimulation parameter space for one or more stimulation parameters to be optimized within a respective predefined parameter range;

[0043] (iii) define an objective function where a plurality of points is defined based on a relationship between received input data and an effect to be induced through delivered stimulation;

[0044] (iv) encode the defined stimulation parameter space for the one or more stimulation parameters;

[0045] (v) define an internal model of stimulation effect space including a prediction of a value for the objective function for each point defined in the stimulation parameter space;

[0046] (vi) optimize an acquisition function for defining at least one set of stimulation parameters based on a current state of the internal model of stimulation effect space;

[0047] (vii) perform iteration of steps (ii) to (vi) based on input data collected during stimulation delivery;

[0048] (viii) define an iteration stopping criterion for the objective function, and

[0049] (ix) when the iteration stopping criterion is reached, output at least one set of stimulation parameters that lead to the highest values for the objective function, and

[0050] - a control module, configured and adapted to receive at least one set of stimulation parameters from the learning module and translate the received at least one set of stimulation parameters into a sequence of commands for the stimulation module. The present invention relates to a neuromodulation / neurostimulation system for planning and / or providing neurostimulation / neuromodulation for enabling a body function and / or treatment of a body condition of a mammal.

[0051] In particular, the system is a closed-loop system.

[0052] Preferably, said mammal is a human patient.

[0053] In particular, said human patient may have at least one of a motor complete or motor incomplete spinal cord injury (SCI) and / or ischemic or traumatic brain injury and / or a neurodegenerative condition affecting the brain and / or the spinal cord.

[0054] In particular, the body function may be a motor function of the patient.

[0055] Said motor function may comprise at least one among standing, stepping, reaching, grasping, voluntary changing a position of one or both legs, walking a motor pattern, and / or sitting down or laying down.

[0056] In addition or alternatively, the body function may be an autonomic function of the patient.

[0057] Said autonomic function may comprise at least one of a cardiovascular function, body temperature, metabolic processes, sexual function, vasomotor function, breathing, swallowing, and / or voiding the patient’s bladder or bowel.

[0058] Said body condition can be, for example, chronic pain.

[0059] The system comprises a stimulation module.

[0060] The stimulation module comprises one or more electrode arrays.

[0061] In particular, the stimulation module is configured and adapted to deliver stimulation to the patient according to one or more stimulation parameters through the one or more electrodes.

[0062] There is also a learning module provided.

[0063] In particular, the learning module is configured and adapted to:

[0064] (i) receive a user command for a selected stimulation function;

[0065] (ii) define a stimulation parameter space for one or more stimulation parameters to be optimized within a respective predefined parameter range; (iii) define an objective function where a plurality of points is defined based on a relationship between received input data and an effect to be induced through delivered stimulation;

[0066] (iv) encode the defined stimulation parameter space for the one or more stimulation parameters;

[0067] (v) define an internal model of stimulation effect space including a prediction of a value for the objective function for each point defined in the stimulation parameter space;

[0068] (vi) optimize an acquisition function for defining at least one set of stimulation parameters based on a current state of the internal model of stimulation effect space;

[0069] (vii) perform iteration of steps (ii) to (vi) based on input data collected during stimulation delivery;

[0070] (viii) define an iteration stopping criterion for the objective function, and

[0071] (ix) when the iteration stopping criterion is reached, output at least one set of stimulation parameters that lead to the highest values for the objective function.

[0072] The system further comprises a control module.

[0073] The control module is configured and adapted to receive at least one set of stimulation parameters from the learning module, and translate the received at least one set of stimulation parameters into a sequence of commands for the stimulation module.

[0074] In particular, this translation can be a direct forwarding of the predicted parameter sets, or a more complex transformation based, e.g., on preferred stimulation timing (e.g., a set repetition rate), safety considerations (e.g., avoiding big stimulation intensity jumps by ensuring amplitude ramping), and / or efficacy considerations (e.g., using knowledge about previous responses to stimulation to estimate minimal amplitude needed to elicit a response).

[0075] For instance, the minimal amplitude to elicit a response can be estimated based on the calibration and then be exploited by the control module to define appropriate stimulation parameters.

[0076] The invention is based on the basic idea that, by the provision of a learning module configured to implement the above-mentioned steps (i) to (ix), it is possible to personalize and optimize stimulation parameters (e.g., frequency, location, amplitude, pulse-width, burst type, or the like) in a (semi-)automatic manner, without requiring the intervention of highly-trained personnel. Also, the system can be easily adapted to any neuromodulation strategy, without being limited to spinal cord stimulation. That is, the system according to the invention is characterized by an enhanced level of flexibility of application in different kinds of treatment.

[0077] Simple stimulation parameters, such as frequency and amplitude, can be encoded (see (iv)) as their real values.

[0078] On the other hand, as explained in more detail in the following, encoding (see (iv)) of more complex stimulation parameters, such as electrode configuration (possibly with multiple anodes and cathodes) and / or anode / cathode configuration with non-binary current distributions, requires either a mapping between a list of discrete options and a continuous N- dimensional space, or a method to evaluate similarity between discrete objects (such as optimal transport).

[0079] In particular, the definition of the internal model of stimulation effect space including a prediction of a value for the objective function for each point defined in the stimulation parameter space (see (v)) can be based on prior knowledge, e.g. including one or more among: generic knowledge or trends regarding the effects of different stimulation parameters; information specific to the participant, e.g. implant location, results from personalized computer simulations, or the like; previously acquired sensor and human feedback data, and / or type of model chosen, such as kernel in Gaussian processes.

[0080] For instance, the kernel may encode the relationship between the different electrical parameters by expressing the degree of dependence among each other.

[0081] As an example, frequency and pulses are strictly correlated, but electrode configuration is more independent, as a cathode choice in the wrong spatial location will continue giving erroneous outcomes even with a correct frequency, usually targeting the specific functional movement. The aspects of independence can be expressed with an additive kernel which is a composition of different kernels with relations determined by mathematical operations, such as multiplication and addition.

[0082] Here, prior knowledge about the specific problem influences the design and choice of the model that is used.

[0083] In particular, the integration of prior knowledge in the definition of the internal model of stimulation effect space (see (v)) allows speeding up the optimization process while remaining robust against patient peculiarities.

[0084] Otherwise stated, prior knowledge helps the learning module to build an initial internal representation of the stimulation parameter space, which highly speeds up efficient navigation through the parameter space, while acquired data for defining at least one set of stimulation parameters based on a current state of the internal model of stimulation effect space (see (vi)) can be immediately used to update the internal representation of the stimulation parameter space (e.g., in case an actual observation and the assumed prior knowledge do not match).

[0085] The stopping criterion can be a mathematical formula or logic description of a criterion for stopping the search.

[0086] At each iteration, if the stopping criterion is not reached, the learning module proposes a new set of stimulation parameters to be applied to the patient.

[0087] Then, as the stopping criterion is reached, the learning module outputs the at least one set of stimulation parameters that lead to the highest values for the objective function (see (ix)).

[0088] Advantageously, the learning module may be configured and adapted to implement a Gaussian Process Bayesian Optimisation algorithm, preferably a Gaussian Process with Upper Confidence Bound algorithm.

[0089] Gaussian processes are suitable for low dimensions, low data regimes, and to make sensible interpolation and extrapolation from data.

[0090] In addition or alternatively, the learning module, as well as the acquisition function, may be configured and adapted to implement a deep learning process.

[0091] Presently, deep learning approaches still fail to provide settings with limited available data (at the moment, too little data is available to characterize the entire population and stimulation effects). However, it can be reasonably expected that deep learning approaches will become useful as more data related to the effects of neuromodulation parameters becomes available.

[0092] Deep learning approaches can be used to create a more complex model which better explains the spinal cord, thereby enabling potentially better predictions.

[0093] Moreover, the way to make choices (also referred to as choice policy), equivalent to the acquisition function in Bayesian optimization approaches, can also be learned from data.

[0094] For example, based on currently collected data, a transformer neural process (TNP) could be trained in a supervised manner, and conditioned on online data to make predictions for different sessions and different patients.

[0095] Specifically, when more data is rendered available, an approach such as TNP can be improved, and the acquisition function can be learned using the same base architecture but trained according to a reinforcement learning approach.

[0096] Advantageously, the input data may comprise user input data.

[0097] In particular, the user input data may include a qualitative feedback (e.g., good / bad, better / worse, or the like) from the patient and / or operator.

[0098] In addition or alternatively, the user input data may include a pseudo-quantified feedback (e.g., by using qualitative scoring scales) from the patient and / or operator.

[0099] In addition or alternatively, the user input data may include a sporadic feedback (e.g., feeling of discomfort, safe / unsafe, specific assessment of stimulation effects quality, or the like) from the patient and / or operator.

[0100] The patient’s feedback can be provided, e.g., through oral and / or non-verbal communication with other individuals involved in the session.

[0101] In addition or alternatively, the patient’s feedback can be provided though a user interface, e.g. a keyboard, touch-screen, or the like.

[0102] The operator’s feedback may as well be provided through a user interface, e.g. a keyboard, touch-screen, or the like.

[0103] Advantageously, the system may further comprise a sensor means. The sensor means may be arranged either on the patient or in the vicinity of the patient, to acquire data representative of the effects of delivered stimulation.

[0104] The sensor means may include one or more cameras.

[0105] In addition or alternatively, the sensor means may include one or more electromyographic sensors.

[0106] In addition or alternatively, the sensor means may include one or more brain sensors.

[0107] In addition or alternatively, the sensor means may include one or more kinematic sensors.

[0108] In addition or alternatively, the sensor means may include one or more vesical pressure monitors.

[0109] In addition or alternatively, the sensor means may include one or more blood pressure sensors.

[0110] In addition or alternatively, the sensor means may include one or more electrocardiographic sensors.

[0111] For example, electromyographic signals from different lower-limb muscles can be obtained to quantify muscular activation.

[0112] In a further example, optionally in addition to the above, camera data can be obtained to observe a movement of the patient.

[0113] In a still further example, optionally in addition to the above, brain recording data can be obtained to detect a feedback from the patient.

[0114] In this case, optionally in addition to the above, the input data may comprise sensor data collected through said sensor means.

[0115] By performing iteration of steps (ii) to (vi) based on input data (see (viii)), e.g. user input data from the patient and / or operator and / or sensor data, collected during stimulation delivery, it is possible to obtain a more satisfactory and effective outcome.

[0116] For example, user input data from the patient and / or operator and / or sensor data may be processed by the learning module to update the stimulation parameter space, which can be either shrunk or extended according to the current needs. In a further example, optionally in addition to the above, the user input data from the patient and / or operator and / or sensor data may be processed by the learning module to update the defined objective function.

[0117] For instance, one or more stimulation parameters of the objective function can be adapted, e.g. to handle a trade-off between multiple stimulation effects.

[0118] In a still further example, optionally in addition to the above, user input data from the patient and / or operator and / or sensor data may be processed by the learning module to update the internal representation of the stimulation parameter space.

[0119] A calibration step may be required to calibrate the sensor means and / or the interpretation of the collected input data.

[0120] This calibration step may comprise assessment of the effects of a pre-defined sequence of stimulation parameters.

[0121] Advantageously, the system may comprise a storage module.

[0122] In particular, the storage module may be configured to store input data collected during stimulation delivery.

[0123] The storage module is operatively connected to the learning module to implement bi-directional communication.

[0124] The stored input data (user input data from the patient and / or operator and / or sensor data) may thus be stored, analysed and immediately used as additional prior knowledge and / or for adaption of the stimulation parameters.

[0125] Optionally, said input data may be collected during the calibration step.

[0126] The stimulation parameters may include one or more among: stimulation frequency; stimulation amplitude; pulse width; pulse shape; charge distribution; functional stimulation amplitude range; temporal evolution of all parameters during the stimulation; electrode(s) configuration; stimulation timing, and / or anode / cathode configuration with non-binary current distributions.

[0127] In particular, the functional stimulation amplitude range denotes the range of amplitudes that the patient reacts to, which are comfortable and efficient for the purpose of achieving a desired therapeutic outcome.

[0128] Also, by defining a proper temporal evolution for all parameters during the stimulation burst, it is possible to obtain smooth and efficient restoration of a desired motor and / or autonomic function.

[0129] Advantageously, the learning module may be configured and adapted to encode the defined stimulation parameter space for the one or more stimulation parameters by implementing a mapping operation between a list of discrete options and an N-dimensional (often lowerdimensional) space, or by defining a specific metric between discrete objects (such as optimal transport).

[0130] The latter aspect might be advantageous since the original input can be used (without the need for implementing any extra steps for transformation to continuous spaces, which are subject to loss of information), thereby more reliably modelling the problem. Here, the acquisition function is optimized by using a list of pre-defined stimulation parameter combinations or by heuristically searching and increasing the size of the search space.

[0131] On the other hand, a well-built or learnt invertible map to an N-dimensional continuous space allows for efficient optimization of the acquisition function (avoiding high dimensions) to obtain the next stimulation protocol to test, without the need for a pre-defined search space, as well as the use of more standard metrics between continuous objects (such as Euclidean distance).

[0132] Moreover, such continuous space can also be a byproduct of deep learning approaches which, with the use of data, may create a specific data-driven embedding space specific for the problem at hand. Advantageously, the learning module may be further configured and adapted to estimate a level of uncertainty for one or more of the predicted values in the internal model of stimulation effect space.

[0133] For instance, estimation of the level of uncertainty can be based on prior knowledge as described above.

[0134] Estimating the level of uncertainty for one or more of the predicted values is helpful in order to enhance the overall level of accuracy.

[0135] Preferably, the learning module may be configured to estimate the level of uncertainty for one or more of the predicted values in the internal model of stimulation effect space by implementing a Gaussian Process, e.g. according to the formula explained in the following: tW = fc(x) ■ (Kt+ ■ / t)“1■ yt

[0136] In particular, by placing a zero-mean Gaussian Process as a prior distribution over f, after t iterations with input x and noisy observations y, it is possible to describe the posterior distributions.

[0137] Here, k(x) is the corresponding kernel matrix and t is the number of tested points.

[0138] Advantageously, the learning module may be configured and adapted to perform the optimization of the acquisition function (see (vi)) by implementing a designed or learned heuristic that aims to deal with the exploration-exploitation dilemma.

[0139] Accordingly, the learning module chooses the next stimulation parameters to be tried based on the current state of the internal model of stimulation effect space.

[0140] Advantageously, the learning module may be configured and adapted to define the objective function based on a multidimensional, mathematical relationship between the received input data (e.g., user input from the patient and / or operator and / or sensor data) and an effect to be induced through delivered stimulation. Accordingly, delivered stimulation can be automatically and effectively optimized to obtain a desired therapeutic effect, based on received input data from the user and / or operator and / or sensor means.

[0141] Advantageously, the iteration stopping criterion may include an adaptable threshold.

[0142] This allows defining the criterion for stopping iteration with enhanced accuracy and reliability.

[0143] Advantageously, the adaptable threshold may be based on a predefined objective function.

[0144] In addition or alternatively, the adaptable threshold may be based on a predefined limit on the number of iterations.

[0145] In addition or alternatively, the adaptable threshold may be based on a binary user feedback.

[0146] Advantageously, the user command for a selected stimulation function may include an indication of a target function.

[0147] Also, the user command may include one or more among: a predefined set of stimulation parameters; patient-related information; previously-acquired input data, and / or one or more optimization algorithm parameters.

[0148] Preferably, said previously-acquired input data may include user input data from the patient and / or operator and / or sensor data.

[0149] Advantageously, the stimulation module may be configured and adapted to deliver epidural electrical stimulation (EES).

[0150] In addition or alternatively, the stimulation module may be configured and adapted to deliver transcutaneous electrical stimulation (TENS).

[0151] In addition or alternatively, the stimulation module may be configured and adapted to deliver dorsal root ganglion (DRG) electrical stimulation.

[0152] In addition or alternatively, the stimulation module may be configured and adapted to deliver functional electrical stimulation (FES). In addition or alternatively, the stimulation module may be configured and adapted to deliver brain stimulation.

[0153] The system according to invention is therefore suitable for use in a wide variety of either invasive or non-invasive therapeutic applications.

[0154] The system of the invention is also suitable for use in different applications, e.g. for neuro- or muscular stimulation for sports and / or cognitive augmentation.

[0155] The system described above may be used for implementing a method for planning and / or providing neurostimulation / neuromodulation for enabling a body function, especially a motor function and / or autonomic function, and / or treatment of a body condition of a mammal, especially a human patient, said method comprising the steps of:

[0156] - providing a stimulation module, said stimulation module being configured and adapted to deliver stimulation to said patient according to one or more stimulation parameters through one or more electrodes;

[0157] - providing a learning module, said learning module being configured and adapted to implement the following steps:

[0158] (i) receiving a user command for a selected stimulation function;

[0159] (ii) defining a stimulation parameter space for one or more stimulation parameters to be optimized within a respective predefined parameter range;

[0160] (iii) defining an objective function where a plurality of points is defined based on a relationship between received input data and an effect to be induced through delivered stimulation;

[0161] (iv) encoding the defined stimulation parameter space for the one or more stimulation parameters;

[0162] (v) defining an internal model of stimulation effect space including a prediction of a value for the objective function for each point defined in the stimulation parameter space;

[0163] (vi) optimizing an acquisition function for defining at least one set of stimulation parameters based on a current state of the internal model of stimulation effect space; (vii) performing iteration of steps (ii) to (vi) based on input data collected during stimulation delivery;

[0164] (viii) defining an iteration stopping criterion for the objective function, and

[0165] (ix) when the iteration stopping criterion is reached, outputting at least one set of stimulation parameters that lead to the highest values for the objective function, and

[0166] - providing a control module, said control module being configured and adapted to implement the following steps: receiving at least one set of stimulation parameters from the learning module; translating the received at least one set of stimulation parameters into a sequence of commands for the stimulation module, and forwarding the obtained sequence of commands to the stimulation module.

[0167] In particular, the body function may be a motor function of the patient, e.g. comprising at least one of standing, stepping, reaching, grasping, voluntary changing a position of one or both legs, walking a motor pattern, and / or sitting down or laying down.

[0168] In addition or alternatively, the body function may be an autonomic function of the patient, e.g. comprising at least one of a cardiovascular function, body temperature, metabolic processes, sexual function, vasomotor function, breathing, swallowing, and / or voiding the patient’s bladder or bowel.

[0169] The patient may have at least one of a motor complete or motor incomplete spinal cord injury (SCI) and / or ischemic or traumatic brain injury and / or a neurodegenerative condition affecting the brain and / or the spinal cord.

[0170] Advantageously, the method may further include implementing a Gaussian Process Bayesian Optimization algorithm, preferably a Gaussian Process with Upper Confidence Bound algorithm.

[0171] In addition or alternatively, the method may include implementing a deep learning process.

[0172] Advantageously, the input data may include user input data from the patient and / or operator.

[0173] Preferably, said user input data include one or more among: a qualitative feedback from the patient and / or the operator; a pseudo-quantified feedback from the patient and / or the operator, and / or a sporadic feedback from the patient and / or the operator.

[0174] In addition or alternatively, the input data may include sensor data obtained through sensor means of the system.

[0175] Preferably, said sensor means of the system may include one or more among: one or more cameras; one or more electromyographic sensors; one or more brain sensors; one or more kinematic sensors; one or more vesical pressure monitors; one or more blood pressure sensors, and / or one or more electrocardiographic sensors.

[0176] Advantageously, the method may further include storing input data collected during stimulation delivery.

[0177] This step is implemented by using storage means of the system.

[0178] Advantageously, the one or more stimulation parameters include at least one among: stimulation frequency; stimulation amplitude; pulse width; pulse shape; charge distribution; functional stimulation amplitude range; temporal evolution of all parameters during the stimulation burst; stimulation amplitude profile; electrode(s) configuration; stimulation timing, and / or anode / cathode configuration with non-binary current distributions.

[0179] Advantageously, the method may further include the step of implementing a mapping operation between a list of discrete or continuous options and an N-dimensional space, for encoding the defined stimulation parameter space for the one or more stimulation parameters.

[0180] Advantageously, the method may further include the step of estimating, through the learning module, a level of uncertainty for one or more of the predicted values in the internal model of stimulation effect space.

[0181] Preferably, the step of estimating the level of uncertainty for one or more of the predicted values in the internal model of stimulation effect space is carried out by implementing a Gaussian Process.

[0182] Preferably, the step of performing an acquisition function for defining at least one set of stimulation parameters based on a current state of the internal model of stimulation effect space (see (vi)) is carried out by implementing a designed heuristic or learned acquisition function that aims to deal with the exploration-exploitation dilemma.

[0183] Advantageously, the method may include defining the objective function based on a multidimensional, mathematical relationship between the received input data (e.g., user input data from the patient and / or operator and / or sensor data) and an effect to be induced through delivered stimulation.

[0184] Advantageously, the method may include setting the stopping criterion to include an adaptable threshold.

[0185] Preferably, the adaptable threshold may be based on one or more among: a predefined objective function; a predefined limit on the number of iterations, and / or a binary user feedback.

[0186] Advantageously, the user command for a selected stimulation function may include an indication of a target function.

[0187] Also, optionally in addition to the above, said user command may include one or more among: a predefined set of stimulation parameters; patient-related information; previously-acquired input data, and / or one or more optimization algorithm parameters.

[0188] Preferably, said previously-acquired input data include user input data (e.g., from the patient and / or operator) and / or sensor data.

[0189] The method is adapted to carry out invasive and / or non-invasive therapeutic treatments, and in particular to deliver one or more among: epidural electrical stimulation (EES); transcutaneous electrical stimulation (TENS); dorsal root ganglion (DRG) electrical stimulation; functional electrical stimulation (FES), and / or brain stimulation.

[0190] Further details and advantages of the present invention will now be disclosed in connection with the drawings, where:

[0191] Figs. 1a-e are diagrams showing an exemplary use of EES for restoring a motor function of the lower-limbs of a patient. In detail: Fig. 1a. Stimulation is delivered through a paddle lead placed over the lumbosacral segments of the spinal cord. Spinal roots innervate the lower-limb muscles in a person-specific rostro-caudal organization. Fig. 1 b. Library of optimized anode and cathode configurations and stimulation frequencies to modulate motor pools associated with the key phases of gait. Fig. 1c. Sequence of muscle activity underlying walking in healthy people, converted into a spatiotemporal map of motor neuron activity that highlights the timing and location of motor hotspot activation, translated into a pre-programmed sequence of stimulation bursts (template) that aims to reproduce this activation pattern. The configurations of electrodes targeting each hotspot are derived from the library and injected into this template. Fig. 1d. Software enabling live adjustments of stimulation patterns and parameters based on real-time feedback from muscle activity and kinematic sensors that are synchronized with stimulation sequences. Fig. 1e. Walking on a treadmill with stimulation after less than 1 hour of configuration and independent walking between parallel bars less than 1 week after the onset of the therapy, including sequence of stimulation and underlying muscle activity.

[0192] Fig. 2 is a diagram showing the concept of using automated Gaussian-process based optimization of a limited set of stimulation parameters for targeted single joint movements in a patient, according to the prior art.

[0193] Fig. 3 is a schematic view of a system for planning and / or providing neurostimulation / neuromodulation for enabling a body function, especially a motor function and / or autonomic function of a mammal, in particular a human patient, according to an embodiment of the present invention. Here, components that are optional and / or alternative are depicted in dashed lines.

[0194] Fig. 4 is a diagram showing the operative interactions between components of the system of Fig. 3.

[0195] Figs. 5a-b are diagrams showing EES system setups used with different participants in a clinical trial carried out on human patients (Patients 1-5). In detail: Fig. 5a. System used for Patients 1-4: here, EES commands from a custom-built stimulation software are transmitted to commercially-available IPG Medtronic Activa® RC via Bluetooth (1) to a module that converts them into infrared signals (2), which are then transferred to the stimulation programmer device (2’). The stimulation programmer transmits EES commands into the IPG via induction telemetry (4), using an antenna (3) taped to the skin of the patient and aligned to the IPG. EES is delivered through the paddle array (5). The first participants were implanted with commercially-available Medtronic Specify® 5-6-5 paddle lead, whereas the later participants were implanted with an investigational new paddle lead. This communication chain allows the control of up to 4 concomitant stimulation waveforms in real-time, with a response latency of approximately 120 ms. Fig. 5b. System used for Patient 5: here, EES commands from a custom-built stimulation software are transmitted to a different IPG (Onward ARCIM®) via Bluetooth to the main controller (1). The commands are sent to the IPG via induction telemetry (2). EES is finally delivered through the lead (3).

[0196] Fig. 6 is a diagram showing an example of automated optimization process and preliminary results of the above-mentioned clinical trial. In detail: (1) Flow diagram. First, the target functional movement is selected. Then, the learning algorithm determines the stimulation to explore, the control module checks the stimulation parameters before delivering the electrical stimuli, then feedback (from sensors, physiotherapist, and / or patient) is collected and the learning algorithm’s knowledge is updated before making the next choice. (2) Main properties of the learning algorithm: a. Parameter search space, b. Objective function, i.e. reward measure. This measure is highly interpretable and the importance of each group (ipsilateral undesirable, antagonist, or contralateral) can be adjusted based on physiotherapist feedback, c. Encoding of the electrical stimulation, d. Internal model: Gaussian Process and Acquisition Function. (3) A control module ensures stimulations are always safe and comfortable for the participant. (4) Stimulation system to deliver EES. (5) Trains of EES elicit lower-limb movements while participant is in supine position. (6) The effects of the stimulation are observed and quantified and lead to an update of the internal model. Preliminary results on three different patients. The time to convergence corresponds to the amount of time needed for the learning algorithm to explore high reward stimulations.

[0197] Figs. 7a-d are diagrams showing preliminary test results on three different patients in the above-referred clinical trial. In detail: Fig. 7a. Increased search space used to optimize stimulation parameters. Here, a subset is sampled for online optimization in a dynamic fashion. Fig. 7b. Time needed for the learning algorithm to find 6 predefined target movements (here, results shown for the three patients). Fig. 7c. Results of an online run along with algorithm predictions for the most performant cathode locations. The best configurations found are shown alongside polar plots that represent selectivity of movement. Fig. 7d. Representation of another online run, in particular on Patient 4. Here, model predictions by iteration for the cathodes, frequency and pulse are shown.

[0198] Figs. 8a-c are diagrams showing optimization of left knee extension in Patient 1 , in the above-referred clinical trial. In detail: Fig. 8a. Predictions for cathode locations for a fixed frequency and pulse, using two Gaussian processes with two distinct kernels priors. Darkest areas denote the highest values. The area predicted by the upper Gaussian process is closer to the desired target for left knee extension on Patient 1 . The lower predictions are more spread out. Fig. 8b. Comparison with and without transfer knowledge: when the learning algorithm contains more explored points, it converges faster to better stimulation parameters, depicted as a higher reward. Fig. 8c. Comparison between predictions error using the two Gaussian process used in (a). With an incorrect kernel, with respect to the problem at hand, the algorithm makes worse predictions, nevertheless given enough data, the difference in the error diminishes. The results shown are based on 10 runs of the algorithm.

[0199] Figs. 9a-b are diagrams showing further results concerning the above-referred clinical trial. In detail: Fig. 9a. Reward values of an offline run using a TNP trained on one session data and fitted in 30 points of the actual session it was tested on. Fig. 9b Showcasing of 3 examples with respect to the predictions of the value of quantified EMGs for different situations. Predictions to other patients can be improved the more data collected. Fig. 10 provides diagrams showing further results concerning the abovereferred clinical trial, in particular illustrating optimization of functional movements including: hip flexion, knee extension, and whole leg flexion. In detail: 1. Definition of an automated algorithm to optimize EES. 2. Multimodal characterisation of functional movements. 3. An interpretable reward defines movement quality. 4. Rapid personalization of EES parameters. 5. Results are functional and comparable to expert- driven optimization.

[0200] Fig. 3 provides a schematic overview of a neuromodulation / neurostimulation system 100 according to an embodiment of the invention.

[0201] Operative interactions between components of the system 100 are illustrated in Fig. 4.

[0202] The system 100 is a closed-loop system that is configured and adapted for planning and / or providing neurostimulation / neuromodulation for enabling a body function in a mammal.

[0203] In the present embodiment, said mammal is a human patient P (Fig. 3).

[0204] In particular, the patient P may have one or more of a motor complete or motor incomplete spinal cord injury (SCI) and / or ischemic or traumatic brain injury and / or a neurodegenerative condition affecting the brain and / or the spinal cord. In the present embodiment, said body function may be a motor function of the patient P, e.g. standing, stepping, reaching, grasping, voluntary changing a position of one or both legs, walking a motor pattern, and / or sitting down or laying down.

[0205] Also, said body function may be an autonomic function of the patient P, e.g. a cardiovascular function, body temperature, metabolic processes, sexual function, vasomotor function, breathing, swallowing, and / or voiding the patient’s bladder or bowel.

[0206] The system 100 may be configured and adapted to deliver invasive and / or non-invasive stimulation.

[0207] In particular, in the present embodiment, the system 100 may be configured and adapted to deliver at least one among: epidural electrical stimulation (EES); transcutaneous electrical stimulation (TENS); dorsal root ganglion (DRG) electrical stimulation; functional electrical stimulation (FES); and / or brain stimulation.

[0208] That is, the system 100 can be easily adapted to a wide variety of either invasive or non- invasive stimulation strategies, thereby showing an enhanced level of flexibility.

[0209] The one or more stimulation parameters include at least one among: stimulation frequency; stimulation amplitude; pulse width; pulse shape; charge distribution; functional stimulation amplitude range; temporal evolution of all parameters during the stimulation burst; stimulation amplitude profile; electrode(s) configuration; stimulation timing, and / or anode / cathode configuration with non-binary current distributions.

[0210] The system 100 includes a stimulation module 10 (Figs. 3-4).

[0211] The stimulation module 10 includes one or more electrodes 12 (Fig. 3).

[0212] In particular, the stimulation module 10 is configured and adapted to deliver stimulation to the patient P according to one or more stimulation parameters, through said one or more electrodes 12.

[0213] The system further includes a learning module 14 (Figs. 3-4).

[0214] In particular, the learning module 14 is configured and adapted to:

[0215] (i) receive a user command for a selected stimulation function;

[0216] (ii) define a stimulation parameter space for one or more stimulation parameters to be optimized within a respective predefined parameter range;

[0217] (iii) define an objective function where a plurality of points is defined based on a relationship between received input data and an effect to be induced through delivered stimulation;

[0218] (iv) encode the defined stimulation parameter space for the one or more stimulation parameters;

[0219] (v) define an internal model of stimulation effect space including a prediction of a value for the objective function for each point defined in the stimulation parameter space;

[0220] (vi) perform an acquisition function for defining at least one set of stimulation parameters based on a current state of the internal model of stimulation effect space;

[0221] (vii) perform iteration of steps (ii) to (vi) based on input data collected during stimulation delivery;

[0222] (viii) define an iteration stopping criterion for the objective function, and (ix) when the iteration stopping criterion is reached, output at least one set of stimulation parameters that lead to the highest values for the objective function.

[0223] There is also a control module 16 provided (Figs. 3-4).

[0224] In particular, the control module 16 is configured and adapted to receive at least one set of stimulation parameters from the learning module 14, and translate the received at least one set of stimulation parameters into a sequence of commands for the stimulation module 10 (Fig. 4).

[0225] As mentioned, by implementing the above-mentioned steps (i) to (ix) through the learning module 14, it is possible to personalize and optimize stimulation parameters (e.g., frequency, location, amplitude, pulse-width, burst type, or the like) in a (semi-)automatic manner, without requiring intervention of highly-trained personnel.

[0226] In the present embodiment, the learning module 14 is configured and adapted to implement a Gaussian Process Bayesian Optimisation algorithm, preferably a Gaussian Process with Upper Confidence Bound algorithm.

[0227] In addition or alternatively to the above, the learning module 14 may be configured and adapted to implement a deep learning process.

[0228] In the present embodiment, the input data include user input data, preferably comprising one or more among: a qualitative feedback from the patient P and / or the operator; a pseudo-quantified feedback from the patient P and / or the operator, and / or a sporadic feedback from the patient P and / or the operator.

[0229] In the present embodiment, the system 100 further comprises a sensor means 18 (Figs. 3-4).

[0230] Therefore, in addition or alternatively to the above, the input data may include sensor data collected through the sensor means 18 (Fig. 4).

[0231] In particular, the sensor means 18 may include one or more among: one or more cameras; one or more electromyographic sensors; one or more brain sensors; one or more kinematic sensors; one or more vesical pressure monitors; one or more blood pressure sensors, and / or one or more electrocardiographic sensors.

[0232] In the present embodiment, the system 100 further comprises a storage module 20 (Figs. 3- 4).

[0233] The storage module 20 is configured and adapted to store input data collected during stimulation delivery.

[0234] Also, the storage module 20 is operatively connected to the learning module 14 for implementing bi-directional communication (Fig. 4).

[0235] In the present embodiment, the learning module 14 is configured and adapted to encode the defined stimulation parameter space for the one or more stimulation parameters by implementing a mapping operation between a list of discrete options and an N-dimensional space, or by defining a specific metric between discrete objects.

[0236] Also, the learning module 14 is configured and adapted to estimate a level of uncertainty for one or more of the predicted values in the internal model of stimulation effect space.

[0237] Preferably, the level of uncertainty for one or more of the predicted values in the internal model of stimulation effect space is estimated by implementing a Gaussian Process.

[0238] In the present embodiment, the learning module 14 is configured and adapted to perform said acquisition function by implementing a designed heuristic or learned acquisition function that aims to deal with the exploration-exploitation dilemma.

[0239] Also, the learning module 14 is configured and adapted to define the objective function based on a multidimensional, mathematical relationship between the received input data (e.g., user input data from the patient P and / or operator and / or sensor data) and an effect to be induced through delivered stimulation.

[0240] In the present embodiment, the iteration stopping criterion includes an adaptable threshold. Preferably, said adaptable threshold is based on one or more among: a predefined objective function; a predefined limit on the number of iterations, and / or a binary user feedback.

[0241] In the present embodiment, the user command for a selected stimulation function includes an indication of a target function.

[0242] In addition or alternatively to the above, said user command includes one or more among: a predefined set of stimulation parameters; patient-related information; previously-acquired input data, and / or one or more optimization algorithm parameters.

[0243] Preferably, said previously-acquired input data include user input data (e.g. from the patient P and / or operator) and / or sensor data.

[0244] The system 100 may be used for implementing a method for planning and / or providing neurostimulation / neuromodulation for enabling a body function, especially a motor function and / or autonomic function, and / or for treatment of a body condition of a mammal, especially a human patient, such as the method described in detail in the foregoing.

[0245] It should be noted that the example control and estimation routines included herein can be used with various system configurations. The control methods and routines disclosed herein may be stored as executable instructions in non-transitory memory and may be carried out by a control unit such as a microcontroller (or a computer) in combination with the described components of the system 100. The specific routines described herein may represent one or more of any number of processing strategies such as event-driven, interrupt-driven, multitasking, multi-threading, and the like. As such, various actions, operations, and / or functions illustrated may be performed in the sequence illustrated, in parallel, or in some cases omitted. Likewise, the order of processing is not necessarily required to achieve the features and advantages of the example embodiments described herein but is provided for ease of illustration and description. One or more of the illustrated actions, operations and / or functions may be repeatedly performed depending on the particular strategy being used. Further, the described actions, operations and / or functions may graphically represent a code to be programmed into non-transitory memory of a computer readable storage medium in a control unit (e.g. a microcontroller) of the system 100, where the described actions are carried out by executing the instructions in a system including the various hardware components in combination with an electronic control unit.

[0246] Clinical trial on human participants with chronic traumatic spinal cord injury (SCI)

[0247] A clinical trial has been carried out with respect to a plurality of human patients (Patients 1-5), involved in experimental trials directed at restoring a motor function of the lower-limbs after SCI (www.clinicaltrials.gov identifier NCT02936453 and NCT04632290, respectively).

[0248] All involved participants presented chronic traumatic SCI with very limited or no residual control over the lower limbs.

[0249] Exemplary stimulation setups used in the context of these clinical trials are shown in Figs. 5a- b.

[0250] More specifically, Fig. 5a shows of the stimulation setup used for Patients 1-4, while Fig. 5b shows the stimulation setup used for Patient 5.

[0251] The participants were treated through delivery of Epidural Electrical Stimulation (EES).

[0252] An exemplary implementation of targeted EES for restoration of a lower-limb motor function is illustrated in Figs. 1a-e, described in detail in the foregoing.

[0253] The participants were implanted with a 16-electrode paddle lead over the lumbosacral region of the spinal cord.

[0254] In particular, the first participants were implanted with a commercially-available Medtronic Specify® 5-6-5 paddle lead used off-label, whereas the later participants were implanted with a more targeted, investigational paddle lead.

[0255] For Patients 1-4, the paddle lead was connected to commercially-available IPG Activa® RC from Medtronic, endowed with real-time triggering capacities through research firmware (Fig. 5a).

[0256] A different IPG (Onward ARCIM®) was used for Patient 5.

[0257] This allowed custom-built software platforms to communicate with the IPG and trigger the correct stimulation at the correct location and time.

[0258] For Patients 1-4, the EES commands from a custom-built stimulation software were transmitted to the IPG (Activa® RC, Medtronic) via Bluetooth (1) to a module that converted them into infrared signals (2), which were then transferred to the stimulation programmer device (2’). The stimulation programmer transmitted EES commands into the IPG via induction telemetry (4), using an antenna (3) taped to the skin of the patient and aligned to the IPG. EES was delivered through the paddle array (5)

[0259] This communication chain allowed the control of up to 4 concomitant stimulation waveforms in real-time, with a response latency of approximately 120 ms.

[0260] For Patient 5, EES commands were transmitted to the used IPG (Onward ARCIM®) via Bluetooth to the main controller (1). The commands were sent via induction telemetry (2) and EES was finally delivered through the paddle array (3).

[0261] This communication chain allowed the control of up to 6 concomitant and independent waveforms real time.

[0262] An example of automated optimization process and preliminary results in the context of the present clinical trial is illustrated in Figs. 6-7.

[0263] First, a target functional movement was selected. Then, a learning algorithm determined stimulation to explore, and a control module checked stimulation parameters before delivering the electrical stimuli. A feedback, e.g., a user input from the patient and / or operator (such as a physiotherapist), and / or sensor data, were collected.

[0264] Due to the highly interpretable nature of the reward measure, the user, e.g., a physiotherapist. Is enabled to alter the importance given to the predefined muscle groups in the optimisation process.

[0265] In this case, the measure can be scaled according on the feedback of the physiotherapist to take into account the specific responses of each patient.

[0266] The knowledge of the learning algorithm was updated before making the next choice (see (1)).

[0267] As defined in (2), here the main properties of the learning algorithm are: a. parameter search space; b. objective function, e.g. reward measure; c. encoding of the electrical stimulation, and d. internal model (Gaussian Process and Acquisition Function). A control module ensured that delivered stimulation was always safe and comfortable for the participant (3).

[0268] The system setup used to deliver EES to each participant is illustrated under (4).

[0269] Trains of EES were delivered to elicit lower-limb movements of each participant, while the participant was lying in supine position (5).

[0270] The effects of stimulation were observed and quantified, and lead to an update of the internal model.

[0271] Results on three different participants are shown under (6).

[0272] Here, the time to convergence corresponds to the amount of time needed for the learning algorithm to explore high-reward stimulations.

[0273] The clinical trial revealed that the use of relevant prior knowledge for the choice of algorithm parameters and internal model initialization highly speeds up the optimization process, as illustrated in Figs. 8a-c.

[0274] In particular, Figs. 8 refers to left knee extension optimization in Patient 1.

[0275] The clinical trial also revealed that the use of incorrect prior knowledge, e.g. due to assumptions that are not valid for a given participant due to specific anatomy or innervation, does slow down the optimization process, but overall does not prevent it from finally converging to appreciable results.

[0276] This ensures the robustness of the optimization process.

[0277] Fig. 8a shows predictions for cathode locations, for a fixed frequency and pulse, using two Gaussian processes with two distinct kernels priors.

[0278] Here, darkest areas denote the highest values.

[0279] The area predicted by the upper Gaussian process is closer to the desired target for left knee extension in the patient.

[0280] The lower predictions are more spread out.

[0281] Fig. 8b provides a comparison with and without transfer knowledge. It may be noted that, when the learning algorithm contains more explored points, it converges faster to better stimulation parameters depicted as a higher reward.

[0282] Fig. 8c provides a comparison between predictions error using the two Gaussian process used in (a).

[0283] With an incorrect kernel, the algorithm makes worse predictions with respect to the problem at hand.

[0284] Nevertheless, given enough data, the difference in the error diminishes.

[0285] The results shown are based on 10 runs of the algorithm.

[0286] The clinical trial also revealed that integration of multimodal feedback implies multiple clear advantages.

[0287] Measurements of muscular activity cannot give information about participant comfort and / or potential pain. By taking direct patient feedback into account, uncomfortable and / or painful stimulations can rapidly be excluded from the search space.

[0288] Moreover, in combination with a smart ramping strategy, post-session analysis showed that the initial stimulation at each step is always below the patient’s discomfort level.

[0289] The search space given to the algorithm was built dynamically, being a subset from a significantly bigger one.

[0290] From the calibration step, or possibly from previous personalized modelling, a probability mass function (PMF) regarding the performance of cathode locations with respect to the desired target was built.

[0291] Then, at each iteration, a cathode location was sampled using such PMF.

[0292] Subsequently, a base electrode configuration was chosen, which has as cathode the previously sampled one.

[0293] This electrode configuration, along with the kernel of the Gaussian process, was used to sample a subset of points from a larger search space.

[0294] This last sampling was biased by the similarity between the points, given by the kernel, where electrode configurations that are more similar to the base electrode configuration have higher chances of being put in the space. Nonetheless, in order to elicit exploration, even with a lower chance, other configurations can also be incorporated in the search space given to the Gaussian process.

[0295] Fig. 9a. shows reward values of an offline run using a TNP trained on one session data and fitted in 30 points of the actual session it was tested.

[0296] Fig. 9b provides a showcasing of 3 examples with respect to the predictions of the value of the quantified EMGs for different situations.

[0297] Predictions for other patients can be improved the more data collected.

[0298] Fig. 10 shows further experimental results obtained in the context of above-referred clinical trial, in particular regarding optimization of three functional movements, namely hip flexion, knee extension, and whole leg flexion.

[0299] A semi-automated algorithm was introduced, allowing for rapid personalization of targeted EES (1.).

[0300] Three different functional movements, i.e., hip flexion, knee extension, and whole leg flexion, were selected (2.).

[0301] These movements are crucial as they represent the fundamental building blocks for locomotion.

[0302] A reward measure was established by characterizing motor output through wireless electromyography and kinematics sensors (3.).

[0303] Compared to human-driven approaches, the present solution allowed for a significant reduction of the required time to achieve personalized and efficient stimulation (4.).

[0304] In particular, the results revealed that the present approach allows obtaining a therapeutic outcome that is comparable to that obtained through highly-trained experts, but with a significant reduction in time required to achieve personalized and efficient stimulation (approx, a 20-fold reduction in time) (5.).

[0305] These results establish a clear path toward facilitated and wide-spread clinical integration of targeted EES. Reference list

[0306] 100 Neuromodulation / neurostimulation system

[0307] 10 Stimulation module

[0308] 12 Electrode(s) (stimulation module)

[0309] 14 Learning module

[0310] 16 Control module

[0311] 18 Sensor means

[0312] 20 Storage module

[0313] P Patient

[0314] Bibliography

[0315] [1] Wagner, F.B., Mignardot, J.-B., Le Goff-Mignardot, C.G., Demesmaeker, R., Komi, S., Capogrosso, M., Rowaid, A., Seanez, I., Caban, M., Pirondini, E., Vat, M., McCracken, L.A., Heimgartner, R., Fodor, I., Watrin, A., Seguin, P., Paoles, E., Van Den Keybus, K., Eberle, G., Schurch, B., Pralong, E., Becce, F., Prior, J., Buse, N., Buschman, R., Neufeld, E., Kuster, N., Carda, S., von Zitzewitz, J., Delattre, V., Denison, T., Lambert, H., Minassian, K., Bloch, J., Courtine, G., 2018. Targeted neurotechnology restores walking in humans with spinal cord injury. Nature 563, 65-71 , https: / / doi.org / 10.1038 / s41586-018-0649-2

[0316] [2] Capogrosso, M., Wagner, F.B., Gandar, J., Moraud, E.M., Wenger, N., Milekovic, T., Shkorbatova, P., Pavlova, N., Musienko, P., Bezard, E., Bloch, J., Courtine, G., 2018. Configuration of electrical spinal cord stimulation through real-time processing of gait kinematics. Nat Protoc 13, 2031-2061 , https: / / doi.org / 10.1038 / s41596-018-0030-9

[0317] [3] Rowaid, A., Komi, S., Demesmaeker, R., Baaklini, E., Hernandez-Charpak, S. D., Paoles, E., Montanaro, H., Cassara, A., Becce, F., Lloyd, B., Newton, T., Ravier, J., Kinany, N., D’Ercole, M., Paley, A., Hankov, N., Varescon, C., McCracken, L., Vat, M., Caban, M., Watrin, A., Jacquet, C., Bole-Feysot, L., Harte, C., Lorach, H., Galvez, A., Tschopp, M., Herrmann, N., Wacker, M., Geernaert, L., Fodor, I., Radevich, V., Van Den Keybus, K., Eberle, G., Pralong, E., Roulet, M., Ledoux, J.-B., Fornari, E., Mandija, S., Mattera, L., Martuzzi, R., Nazarian, B., Benkler, S., Callegari, S., Greiner, N., Fuhrer, B., Froeling, M., Buse, N., Denison, T., Buschman, R., Wende, C., Ganty, D., Bakker, J., Delattre, V., Lambert, H., Minassian, K., van den Berg, C.A.T., Kavounoudias, A., Micera, S., Van De Ville, D., Barraud, Q., Kurt, E., Kuster, N., Neufeld, E., Capogrosso, M., Asboth, L., Wagner, F.B., Bloch, J., Courtine, G., 2022. Activity-dependent spinal cord neuromodulation rapidly restores trunk and leg motor functions after complete paralysis. Nature Medicine, 28(2), 260-271 , https: / / doi.Org / 10.1038 / S41591 -021 -01663-5

[0318] [4] Desautels, T.A., Choe, J., Gad, P., Nandra, M.S., Roy, R.R., Zhong, H., Tai, Y.C., Edgerton, V.R. and Burdick, J.W., 2015. An active learning algorithm for control of epidural electrostimulation. IEEE Transactions on Biomedical Engineering, 62(10), pp.2443- 2455.

[0319] [5] Marco Bonizzato, Rose Guay Hottin, Sandrine L. Cote, Elena Massai, Leo Choiniere, Uzay Macar, Samuel Laferriere, Parikshat Sirpal, Stephan Quessy, Guillaume Lajoie, Marina Martinez, Numa Dancause, Autonomous optimization of neuroprosthetic stimulation parameters that drive the motor cortex and spinal cord outputs in rats and monkeys. Cell Reports Medicine, Volume 4, Issue 4, 2023, 101008, ISSN 2666-3791 , https: / / doi.Org / 10.1016 / j.xcrm.2O23.101008

[0320] [6] Sui, Y., Yue, Y. and Burdick, J.W., 2017. Correlational dueling bandits with application to clinical treatment in large decision spaces. arXiv preprint arXiv: 1707.02375.

[0321] [7] Abranches de Carvalho, P., Gaussian Processes of Spinal Electrical Stimulation, Master Project Report, Departamento de Ciencia de Computadores Faculdade de Ciencias da Universidade do Porto 2019.

Claims

CLAIMS1. A system (100), especially a closed-loop system, for planning and / or providing neurostimulation / neuromodulation for enabling a body function, especially a motor function and / or autonomic function, and / or treatment of a body condition of a mammal, especially a human patient (P), said system (100) comprising:- a stimulation module (10), configured and adapted to deliver stimulation to said patient (P) according to one or more stimulation parameters through one or more electrodes (12);- a learning module (14) configured and adapted to:(i) receive a user command for a selected stimulation function;(ii) define a stimulation parameter space for one or more stimulation parameters to be optimized within a respective predefined parameter range;(iii) define an objective function where a plurality of points is defined based on a relationship between received input data and an effect to be induced through delivered stimulation;(iv) encode the defined stimulation parameter space for the one or more stimulation parameters;(v) define an internal model of stimulation effect space including a prediction of a value for the objective function for each point defined in the stimulation parameter space;(vi) optimize an acquisition function for defining a set of stimulation parameters based on a current state of the internal model of stimulation effect space;(vii) perform iteration of steps (ii) to (vi) based on input data collected during stimulation delivery;(viii) define an iteration stopping criterion for the objective function, and(ix) when the iteration stopping criterion is reached, output at least one set of stimulation parameters that lead to the highest values for the objective function, and- a control module (16), configured and adapted to receive a set of stimulation parameters from the learning module (14) and translate the received set of stimulation parameters into a sequence of commands for the stimulation module (10).

2. The system (100) according to claim 1 , characterized in that the learning module (14) is configured and adapted to implement a Gaussian Process Bayesian Optimisation algorithm, preferably a Gaussian Process with Upper Confidence Bound algorithm, and / or implement a deep learning process.

3. The system (100) according to claim 1 or 2, characterized in that said input data include user input data, preferably wherein said user input data include one or more among: a qualitative feedback from the patient (P) and / or the operator; a pseudo-quantified feedback from the patient (P) and / or the operator; a sporadic feedback from the patient (P) and / or the operator.

4. The system (100) according to any one of the preceding claims, characterized in that said system (100) further comprises a sensor means (18), wherein said input data include sensor data collected through said sensor means (18), preferably wherein said sensor means (18) include one or more among: one or more cameras; one or more electromyographic sensors; one or more brain sensors;one or more kinematic sensors; one or more vesical pressure monitors, one or more blood pressure sensors, and / or one or more electrocardiographic sensors.

5. The system (100) according to any one of the preceding claims, characterized in that the system further comprises a storage module (20) configured and adapted to store input data collected during stimulation delivery, said storage module (20) being operatively connected to the learning module (14) for implementing bi-directional communication.

6. The system (100) according to any one of the preceding claims, characterized in that said one or more stimulation parameters include at least one among stimulation frequency, stimulation amplitude, pulse width, pulse shape, charge distribution, functional stimulation amplitude range, temporal evolution of all parameters during the stimulation burst, electrode(s) configuration, stimulation timing, and / or anode / cathode configuration with non-binary current distributions.

7. The system (100) according to any one of the preceding claims, characterized in that the learning module (14) is configured and adapted to encode the defined stimulation parameter space for the one or more stimulation parameters by implementing a mapping operation between a list of discrete options and an N-dimensional space, or by defining a specific metric between discrete objects.

8. The system (100) according to any one of the preceding claims, characterized in that the learning module (14) is further configured and adapted to estimate a level of uncertainty for one or more of the predicted values in the internal model of stimulation effect space, preferably wherein the learning module (14) is configured and adapted to estimate the level of uncertainty for one or more of the predicted values in the internal model of stimulation effect space by implementing a Gaussian Process.

9. The system (100) according to any one of the preceding claims, characterized in that the learning module (14) is configured and adapted to perform said acquisition function by implementing a designed or learned heuristic that aims to deal with the explorationexploitation dilemma.

10. The system (100) according to any one of the preceding claims, characterized in that the learning module (14) is configured and adapted to define the objective function based on a multidimensional, mathematical relationship between the received input data and an effect to be induced through delivered stimulation.11 . The system (100) according to any one of the preceding claims, characterized in that said iteration stopping criterion includes an adaptable threshold, preferably wherein said adaptable threshold is based on one or more among: a predefined objective function;a predefined limit on the number of iterations, and / or a binary user feedback.

12. The system (100) according to any one of the preceding claims, characterized in that said user command for a selected stimulation function includes: an indication of a target function, and one or more among: a predefined set of stimulation parameters; patient-related information; previously-acquired input data, preferably wherein said previously-acquired input data include user input data and / or sensor data, and / or one or more optimization algorithm parameters.

13. The system (100) according to any one of the preceding claims, characterized in that said stimulation module (10) is configured and adapted to deliver at least one among: epidural electrical stimulation (EES); transcutaneous electrical stimulation (TENS); dorsal root ganglion (DRG) electrical stimulation; functional electrical stimulation (FES), and / or brain stimulation.

14. The system (100) according to any one of the preceding claims, characterized in that said patient (P) has at least one of a motor complete or motor incomplete spinal cord injury and / or ischemic or traumatic brain injury and / or a neurodegenerative condition affecting the brain and / or the spinal cord.

15. The system (100) according to any one of the preceding claims, characterized in that said function is a motor function, which comprises at least one of standing, stepping, reaching, grasping, voluntary changing a position of one or both legs, walking a motor pattern, sitting down or laying down, and / or said function is an autonomic function, which comprises at least one of a cardiovascular function, body temperature, metabolic processes, sexual function, vasomotor function, breathing, swallowing, and / or voiding the patient’s bladder or bowel, and or / said body condition is chronic pain.

Images