Artificial intelligence enhanced functional electrical stimulation for central nervous system disabilities

Abstract

Provided herein are systems for treating a paralysis patient in need of controlled motion to improve quality of life, and related methods. The systems comprise an implantable system formed from a hub, various electrodes, such as spinal cord (upstream) electrode, periphery nerve (downstream) electrode and an external system formed of communication pod implanted just under the skin, a computing core, and a plurality of motion sensors. A software system, providing AI compatibility are used to supply the computing core, and thereby implanted electrodes, with signal. In this manner, the specially positioned electrodes, upstream of the spinal cord injury and downstream of the spinal cord injury, can provide safe and reliable patient motion.

Description

Artificial Intelligence Enhanced Functional Electrical Stimulation for Central Nervous System Disabilities CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims benefit of priority to US Provisional App. No.63 / 635,421 5 filed April 17, 2024, which is hereby incorporated by reference to the extent not inconsistent herewith. BACKGROUND OF INVENTION

[0002] Provided are systems of implantable and wearable devices and associated software designed to restore mobility, sensations and high level of skilled movement in patients with 10 paralysis caused by spinal cord injuries. The system may use advanced neural interfaces for the spinal cord and periphery nerves. The system has applications for restoring involuntary movements, and may be used to manage cerebral palsy and spinal muscular atrophy.

[0003] 179,312 cases of Spinal Cord Injury (SCI) were reported globally in 2007. The global average was about 23 cases per million, with North America average at 40 cases per 15 million[1]. SCI usually result in loss of ability to perform voluntary and involuntary movements. These patients suffer from permanent disability and reduced life expectancy. Even with mobility devices their lives are still significantly impaired. By combining recent advances in neuroscience, bio-engineering and artificial intelligence (AI), provided herein are functional electrical systems (FES) that can restore a significant amount of mobility and 20 dexterity, without significant cost increases.

[0004] Conventional systems in the art include: US10376389 (showing a prosthesis that uses EMG activity in residual muscle to drive a prosthetic device); US20230086004 (showing an AI-Enabled prosthesis that uses residual periphery nerve signal to perform feedback control of robotic actuators); US7901368 (showing an FES system that uses control 25 signals from the motor cortex); Loeb et al. Med. Eng. Phys., vol.23, pp.9–18, 2001 [2] (showing injectable electrodes activated by a radio frequency magnetic field, called BION).

[0005] A survey of Academic literature by Popovic in 2014 concluded that [3]: “The majority of studies reported in the literature describe sophisticated FES systems that can be used for the restoration of sensory–motor function; however, to date, only simulation studies 30 or anecdotal data from healthy individuals or case studies have been reported for thesesystems.” Accordingly, there is a need in the art for systems that can facilitate restoration of motor function in individuals suffering from a paralysis. SUMMARY OF THE INVENTION

[0006] Provided herein are systems useful for treating a patient having a Central Nervous 5 System (CNS) disability. Examples including spinal cord injury or cerebral palsy. The systems are a specially-configured combination of implantable components, external components, and machine learning (e.g., AI software). The implantable portion, including electrodes, are used to transmit and provide electrical signals around various body tissue, including spinal cord nerve, peripheral nerve, and muscle. Reliable and efficient signal 10 generation is achieved by AI and appropriate training thereof.

[0007] Embodiments of the invention include, but are not limited to:

[0008] 1. An implantable functional electrical stimulator (FES) comprising: an electrode configured for implantation and to functionally integrate with a target biological tissue; and 15 a feedback control system comprising a processor in electronic communication with the electrode, wherein the processer executes processor-executable instructions provided by a machine learning algorithm, to cause the electrode to interface with the target biological tissue.

[0009] 2. The implantable FES of claim 1, wherein the processer implements an interface 20 scheme with the target biological tissue for one or more of: adaptive control; feedback control; and / or predictive control.

[0010] 3. The implantable FES of claim 2, wherein the feedback control comprises one or 25 more motion sensors connected to an external system of the FES for tracking a location of a limb or limb extremity.

[0011] 4. The implantable FES of claim 2, wherein the adaptive control is provided by an adaptive control algorithm implemented via artificial intelligence.

[0012] 5. The implantable FES of any one of claims 1-4, wherein the processor is provided in 30 a computing device that further comprises a memory device communicatively coupled to the processor.

[0013] 6. The implantable FES of claim 5, wherein the memory device includes a non-transi- tory computer readable medium storing processor-executable instructions encoded as software, which, when executed by the processor, cause the processor to implement an interface scheme with the target tissue. 5

[0014] 7. The implantable FES of claim 6, wherein the interface scheme is AI-generated to provide predictive control of the target biological tissue.

[0015] 8. The implantable FES of any one of claims 1-7, wherein the processor and electrode are configured to: monitor an electrical parameter of the target tissue; 10 monitor a target tissue motion; electrically actuate the target tissue; or monitor the electrical parameter of the target tissue and electrically actuate the target tissue.

[0016] 9. The implantable FES of any one of claims 1-8, wherein the electrode is a biohybrid15 electrode comprising a grafted or seeded neuronal cell in electrical contact with a syn- thetic portion of the electrode.

[0017] 10. The implantable FES of any one of claims 1-9, further comprising: a network of electrodes electrically connected to hermetically sealed implantable electronic components; and 20 an external processing element containing at least the processor.

[0018] 11. The implantable FES of any one of claims 1-10, further comprising a pod config- ured to be wearable by a user, wherein the pod comprises a motion sensor for position tracking.

[0019] 12. The implantable FES of any one of claims 1-11, wherein the processor is con- 25 tained in a cellphone, and is in wireless communication with the electrode.

[0020] 13. The implantable FES of any one of claims 1-12, further comprising an assistive exoskeleton in electrical communication with the processor, wherein the assistive exo- skeleton is configured to provide physical support of a biological system and a measure of a kinetic parameter. 30

[0021] 14. The implantable FES of any one of claims 1-10, wherein the electrode comprises: a spinal cord electrode configured to be positioned upstream of a spinal cord lesion and to implantably connect to living cells; a periphery nerve electrode configured to be positioned downstream of the spinal cord lesion and to implantably connect to living cells; andthe processor is configured to process, store and / or transmit electrical spike trains.

[0022] 15. The implantable FES of any one of claims 1-14, further comprising a muscle elec- trode configured to electrically stimulate a muscle based on a signal from the processor 5 and / or a cortical electrode configured to electrically stimulate a cortex layer for tactile sensation.

[0023] 16. The implantable FES of any one of claims 1-15, further comprising: an external system, wherein the external system comprises: a communication pod configured to spatially align with a hub implanted under 10 the skin, wherein the hub is in electronic communication with the electrode(s), a computing core communicatively coupled to the communication pod, a plurality of motion sensors; and a software system operably connected to the computing core, the software system comprising: 15 an adaptation onboard neural network for a subject / encoding adaptation; and a simulation onboard neural network for simulation of a local spinal circuit.

[0024] 17. The implantable FES of any one of claims 1-16, further comprising: a hub config- ured for: implantation beneath skin, a hub in electrical contact with the electrode(s); 20 inductive power charging of an inductively charged battery; and inductive or radio communication with an externally-positioned processor or a communication pod.

[0025] 18. The implantable FES of any one of claims 1-17, wherein the electrode comprises a cellular electrode array (CEA). 25

[0026] 19. The implantable FES of any one of claims 1-18, wherein at least one electrode is a conformable electrode for conformal contact with the target tissue having a shaped sur- face for positioning of biological cells to follow a peripheral nerve to the target tissue comprising a destination muscle.

[0027] 20. The implantable FES of claim 19, provided in a cuff configuration for implanta- 30 tion around a major peripheral nerve, including at a lesion to bridge a nerve gap.

[0028] 21. The implantable FES of any one of claims 1-20, wherein the electrode is a spinal cord electrode configured for implantation onto a spinal cord of the patient.

[0029] 22. The implantable FES of claim 21, wherein the spinal cord electrode is seeded with: (i) cells that are configured to attract axon collateral from upper motor neurons and / or project dendrite directly onto the upper mountain neurons; (ii) sensory neuronal cells; (iii) neural precursor cells; and / or (iv) embryonic neural cells, including provided in 5 an array pattern.

[0030] 23. The implantable FES of any one of claims 1-22, wherein the periphery nerve elec- trode is a bi-directional electrode.

[0031] 24. The implantable FES of any one of claims 1-23, wherein the periphery nerve elec- trode is a cuff electrode. 10

[0032] 25. The implantable FES of any one of claims 1-24, wherein the periphery nerve elec- trode is seeded with: (i) cells that are configured to attract axon collateral from lower motor neurons and / or project dendrite directly onto the lower motor neurons; and / or (ii) sensory neuronal cells.

[0033] 26. The implantable FES of any claims 1-25, wherein the periphery nerve electrode is 15 seeded with sensory and motor neuronal cells.

[0034] 27. The implantable FES of any one of claims 1-26, wherein the at least one muscle electrode comprises a uni-directional electrode.

[0035] 28. The implantable FES of any one claims 1-27, wherein the at least one muscle electrode is seeded with lower motor neuronal cells. 20

[0036] 29. The implantable FES of any one of claims 1-28, wherein the at least one spinal cord electrode, the at least one periphery nerve electrode, and / or the at least one muscle electrode are connected to the at least one hub via an electrically insulated ribbon cable that is biocompatible, flexible and / or stretchable, such as a polyimide ribbon cable.

[0037] 30. The implantable FES of any one of claims 1-28, wherein the hub comprises one or 25 more of: a battery pack; a pulse train generator; a microcontroller; an inductive coupling transceiver; a radio transceiver; an amplifier; an analog-to-digital converter; an inductive charger; a spike detector or threshold circuit; a memory device; and / or an integrated cir- cuit.

[0038] 31. A method of electrically interfacing a target tissue with an implantable FES, the 30 method comprising the steps of: providing the implantable FES of any one of claims 1-30; implanting the electrode into a patient; integrating the electrode with the target tissue;operably connecting the feedback control system to the implanted electrode to thereby provide electronic communication between the feedback control system and the implanted electrode; and training the feedback and control system using the machine learning 5 algorithm; thereby electrically interfacing the target tissue with the implantable FES.

[0039] 32. The method of claim 31, wherein the training comprises a modified DDPG (deep deterministic policy gradient) for FES control and / or a trust region method for DDPG.

[0040] 33. The method of any one of claims 31-32, wherein the processor processes and 10 transmits neural signals as pulse trains, thereby reducing a total required power.

[0041] 34. The method of any one of claims 31-33, for treating paralysis in a patient by indi- rect or direct stimulating of a muscle of the patient; and / or receiving a signal from a pe- ripheral nerve.

[0042] 35. The method of any one of claims 31-34, the method further comprising the step 15 of: training the patient with a defined motor skill to thereby train both neural networks and the patient in a closed-loop configuration.

[0043] 36. The method of any one of claims 31-35, the method further comprising: transmitting processed and / or synthetic sensations back to the patient from the periphery nerve electrode and the plurality of motion sensors worn by the patient. 20

[0044] 37. A method of making an implantable FES, the method comprising the steps of: establishing a bioconnection between biological cells and a conformable array of synthetic electrodes to provide a conformable biohybrid electrode array, wherein the conformable biohybrid electrode array is conformable to a target tissue; and providing a feedback control system comprising a processor in electronic communication 25 with the biohybrid electrode.

[0045] 38. The method of claim 37, wherein the target tissue is muscle, spinal cord, or a su- perficial cortex layer.

[0046] 39. The method of claim 38, wherein the conformable biohybrid electrode array im- planted in a patient is configured to: 30 align cells to follow an existing peripheral nerve pathway to a destination muscle and thereby form a specific connection to a muscle motor unit; attract a target axon to form a specific connection with a biohybrid electrode of the conformable biohybrid electrode array; orprovide cells to form a specific contact with target cells in a superficial cortex layer.

[0047] 40. The method of any one of claims 37-39, wherein each biohybrid electrode of the biohybrid electrode array is provided with a single biological cell, the method further comprising the steps of: 5 microfluidically introducing the single biological cell to well configured to operably connect to a single biohybrid electrode; and repeating the microfluidically introducing step for each biohybrid electrode of the conforma- ble biohybrid electrode array, wherein the repeating step is optionally simultaneous with all other biohybrid electrodes. 10

[0048] 41. The method of any one of claims 34-37, further comprising the step of using a ma- chine learning algorithm to train the implantable FES to generate an interface scheme provided from the processer to the biohybrid electrode to generate one or more biological outcomes.

[0049] Any of the electrodes provided herein may be a cellular electrode array (CEA) 15 device, including as described in WO 2024 / 079699, which is specifically incorporated by reference herein for the disclosed devices, electrodes and electrode arrays, methods of making the devices, and methods of using the devices, including to electrically interface with tissue or cells.

[0050] The CEA’s are novel micro-fabricated electrode arrays which uses living cells to 20 form contacts, and more specifically electrical contacts including synaptic contacts, between conducting electrodes (e.g., polymer electrodes) and target biological tissue, including neural tissue. In this manner, the electrodes may be referred to as a biohybrid electrode because there is an artificially-constructed portion (including inorganic electrode portions) and a cellular component. Through that physical contact, specific electrical interfacing is achieved. 25 The CEA devices, and more specifically, the biohybrid electrodes provided herein, comprise a flexible micro fabricated device made from a polymer, such as PDMS (polydimethylsiloxane) punctuated by a regular grid of wells. The polymer may be on the order of 1 mm thick, such as between about 100 μm and 2 mm thick, depending on the operating conditions and application of interest, including the electric field spread into the 30 well. Conductive polymer contacts are positioned within a well on a surface location, including located at the bottom of the wells. The other top end of the well is open. Neurons are introduced into the wells and cultured, including with hydro-gels, to guide sprouted axons and dendrites towards the open end of the wells. Those sprouted axons and dendrites areavailable to make physical contact with target neural tissue. The combination of the precisely located sprouted axons, their one-to-one contact and electrical interconnection with biological tissue, and the special electrical grounding of the device surface facing the biological tissue, provides the device with the benefit of high specificity in combination with high resolution 5 and attendant functional benefit, including: the ability to precisely, reliably and continuously electrically interface with biological tissue without damage or trauma, including down to a very small sub-regions, with each sub-region specific to an underlying electrical connection defined by a single well.

[0051] More generally, provided are biohybrid electrodes with grafted or seeded neuronal 10 cells configured to establish a biological electrical connection between the grafted neuronal cell and a target neuronal cell in a biological tissue in which the electrode is positioned. In this manner, substantially all electrical connection between the electrode and the biological tissue is through the biological electrical connection. The electrically conducting material of the electrode has an outer surface, wherein the outer surface is optionally formed of an outer 15 material that is different from a core material of the electrically conducting material, with at least one neuronal cell in electrical communication at one end with the electrode.

[0052] Provided herein are biohybrid electrodes or cellular electrodes, including in the form of CEAs, and related methods of making and methods of using any of the CEAs to electrically interface with biological tissue and for making any of the CEAs. For example, 20 provided is a CEA comprising a flexible polymer substrate having a top surface and an array of wells disposed in the flexible polymer substrate. Each well has an inner-facing surface that defines a well volume. The inner facing surface may comprise a bottom surface and side wall surfaces that extend from the bottom surface to the flexible polymer substrate top surface. A conductive contact, such as a conductive polymer contact, silver nanowire (AgNW) contact, 25 or an AgCl contact, at least partially covers the well inner-facing surface, such as positioned at the bottom surface and / or around at least a portion of the side wall. Preferably, the conductive contact is a conductive polymer contact or nanowire, although other electrically- conductive materials are compatible with the electrode arrays, particularly materials that are flexible and are not rigid, so as to not adversely impact the overall device conformability to 30 the biological tissue. In this manner, a well depth corresponding to a distance between the flexible polymer substrate top surface and the well bottom surface is defined. The well volume may be characterized in terms of the volume, and for a circular well cross-section thevolume may be calculated as Hπr2(equation (1)), with H the well depth and r the effective radius of the well, such as d / 2, with d the effective diameter of the well. Of course, the invention is compatible with any of a range of well cross-sections, including non-circular cross-sections such as ovals, rectangular, square and the like. Wells may have shapes with 5 non-constant cross sections, such as tapers and funnels, including so as to receive a droplet. Depending on the application of interest, the well volume can be between 1 μL and 2 mL, including between 10 μL and 1 mL. The well may be configured to contain a hydrogel in each well and to receive a neuronal cell. In this manner, the hydrogel facilitates directed growth of sprouted axons and dendrites from the neuronal cell in a direction toward the 10 flexible polymer substrate top surface. The ends of the sprouted axons and dendrites are available to physically contact biological cells, including biological cells that are electrically active, generating and / or responding to electric stimuli or field.

[0053] The conductive polymer or nanowire contact may be positioned at the well bottom surface and / or along sidewalls and portions thereof. 15

[0054] The polymer substrate that is flexible facilitates conformal contact with a biological tissue, including a biological tissue that is curved or has a time-varying surface shape. In this manner, the cellular electrode is configured to operably connect flush against a tissue surface, including a tissue that corresponds to a brain surface , periphery nerve, or muscle surface with the operably connected mediated by synaptic connections between the 20 neuronal cells in the wells (e.g., “grafted neuronal cell) and dendrites in the biological tissue, including dendrites from brain adjacent to the brain surface for biological tissue that is brain. Accordingly, “grafted” in the instant context refers to the neuronal cell that is reliably supported by wells such that there is sufficient proximity to an electrode surface that there is the ability of each of the grafted neuron and the electrode to electrically influence each 25 other’s electric field.

[0055] The method may further comprise the step of matching a mechanical property of the cellular electrode array with a mechanical property of the biological tissue. The mechanical property may be a modulus, such as a Young’s modulus, that is within 50%, 20% or 10% of each other. 30

[0056] The arrays provided herein are compatible with any of a range of biological tissue, including biological tissue where electric potential and electrical activation are of importance.The biological tissue may be selected from the group consisting of: brain, spinal cord, peripheral nerves, muscles, a three-dimensional cell culture (including a bioartificial organ), and organotypic cultures. Particularly useful is brain tissue, where the brain tissue comprises neuronal cells with brain dendrites that are capable of synaptically connecting to the grafted 5 neuronal cells of the array, thereby providing the specific and reliable electrical connection between the electrodes of the array and the brain tissue. Other useful tissues, particularly for treating lack of motions from patients suffering SCI, include spinal cord and peripheral nerve.

[0057] The method may further comprise the steps of: providing neuronal cells to the wells; providing a hydrogel, including a collagen solution, to the wells; culturing the neuronal 10 cells to generate axon and dendrite growth. In this manner, the array is ready for contact with a biological tissue. Accordingly, the method may further comprise the step of contacting the cellular electrode array with a biological tissue so that the neuronal cells electrically interface with the biological tissue, including biological cells within the biological tissue. The method may be practiced on the living body, including a human or a non-human animal. Any of the 15 methods may be in vivo, in vitro or ex vivo.

[0058] Without wishing to be bound by any particular theory, there may be discussion herein of beliefs or understandings of underlying principles relating to the devices and methods disclosed herein. It is recognized that regardless of the ultimate correctness of any mechanistic explanation or hypothesis, an embodiment of the invention can nonetheless be 20 operative and useful. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] FIGs.1A-1C: Schematic illustration of a cellular electrode array (CEA), with a side view (FIG.1A), a top view (FIG.1B) and an implanted view with attendant electronics for measuring a signal output (FIG.1C). 25

[0060] FIGs.2A-2E: manufacturing steps for an electrode-neuron grafted array.

[0061] FIG.3 is a flow-chart schematic of certain relevant manufacturing steps for making the electrode array. Dark solid is SU8; light solid is PDMS; left hatching represents a conductive composite; right hatching represent wells and holes; horizontal hatching represents metallization; vertical hatching represents channels.

[0062] FIG.4 is similar to FIG.3, but provides another embodiment of selected manufacturing steps for making the electrode array.

[0063] FIG.5 is a schematic of the electrode in use to form a biological electrical connection with a target cell via a grafted neuronal cell. 5

[0064] FIG.6 is a schematic of a system illustrating the various portions, including the implantable system, the external system, and components thereof, on a patient in need of a paralysis treatment, as reflected by spinal cord lesion labeled “x”.

[0065] FIG.7 illustrates a Corticospinal Tract (CST) and the CEA operating as a sensor.

[0066] FIG.8 illustrates Spinocerebellar tract (SCT) and the CEA operating as a 10 stimulator.

[0067] FIG.9 illustrates Dorsal column–medial lemniscus (DCML).

[0068] FIG.10 is a flow diagram summarizing operation of the overall system in close loop feedback.

[0069] FIG.11 summarizes an analysis of a CEA well functioning as a microfluidic cell 15 sorter for reliable positioning of a cell in a well.

[0070] FIG.12 is a process flow summary schematically illustrating steps A-E for fabrication of the array, which also functions as a microfluidic system for cell placement.]

[0071] FIG.13 is an overview of AI architecture (modified DDPG with a dynamic model) useful with the feedback control system machine learning algorithm. 20 DETAILED DESCRIPTION OF THE INVENTION

[0072] In the following description, numerous specific details of the devices, device components and methods of the present invention are set forth in order to provide a thorough explanation of the precise nature of the invention. It will be apparent, however, to those of skill in the art that the invention can be practiced without these specific details. 25

[0073] “Interfacing” is used herein to refer to the ability of two materials or components to electrically interact with each other, so that a change in an electrical parameter (e.g., potential) of one material or component affects the state or status of another material orcomponent, without adversely impacting the functionality of each material or component. Interfacing is used broadly herein to include electrical stimulation (e.g., “actuation”) and / or electric parameter detection (e.g., “measurement” or “detection”). For example, the arrays provided herein may electrically stimulate biological cells and / or detect electrical activity in 5 or associated with biological cells. The interfacing may be via biohybrid electrode directly to a muscle unit or via an intervening biological tissue, such as biohybrid electrode to a peripheral nerve, and then form the peripheral nerve to a muscle unit. Accordingly, one aspect of the feedback control system is an interface with the biological target tissue that provides control of the biological target tissue, including functionally as a robotic actuator, 10 where the biological tissue that is a muscle unit is functionally the actuator.

[0074] Accordingly, an “interface scheme” is use broadly to include a variety of schemes. For example, the term can encompass “control schemes” for actuation of a biological tissue (e.g., provide signal to a nerve which, in turn, results in muscle actuation or directly to a muscle), “training schemes” for providing user and system training (via the AI and motion 15 sensors) so that the system can self-learn with the patient as to what impulses are associated with what function, “measurement scheme” for ensuring the FES is appropriately recording data generated from the biological tissue. In particular, interface scheme may include a feedback control mediated by motion sensors worn by a patient and, optionally, predictive capabilities. 20

[0075] A “feedback control system” is used herein to refer to that part of the FES the provides the electronic control of the electrode. Of particular relevance for the instant FES provided herein, is the incorporation of a machine learning algorithm, including an artificial neural network, where the FES can be effectively “trained” to ensure desired biological outcomes are generated for various input signals, including instructions from the patient. 25

[0076] The “interfacing” may be direct with the target tissue or indirectly via an intervening biological component. In the context of spinal cord applications, for example, the FES may directly contact the peripheral neve or may directly contact a muscle unit. In this example, the peripheral nerve may be considered the intervening component between muscle that is the target tissue (e.g., the FES uses the muscle unit as a type of robotic actuator) where 30 cells of the biohybrid electrode effectively grow into an axon(s) of the nerve. Of course, the FES provided herein can function just as well if the intervening nerve is bypassed by cells of the biohybrid electrode that grow directly into the muscle tissue. Of course, the FESprovided herein can function just as well with both nerve cell innervation in each of the native peripheral nerve and the muscle. The innervation can be accomplished by the grafted or seeded neuronal cell that grows along a path to integrate with a peripheral nerve and / or with a muscle tissue. 5

[0077] “Conformable electrode” refers to an electrode formed of materials that are sufficiently flexible that the electrode can change shape to adapt to a target tissue shape and facilitate interfacing without adversely impacting electrode functionality “Flexible” refers to the ability of a material, structure, device or device component to be deformed into a curved shape without undergoing a transformation that introduces significant force that would 10 otherwise result in failure. A flexible material, structure, or device may be deformed into a curved shape without introducing a force greater than the failure point of the material.

[0078] “Cellular electrode array” refers to any of electrodes as described in WO 2024 / 079699 (PCT / IB2023 / 060324), including the biohybrid electrodes comprising a neuronal cell grafted to the electrode in a well configured for directed growth to a target 15 tissue.

[0079] “Bioconnection” refers to a functional connection between a biological cell of the FES and the electrode portion of the FES. Such a bioconnection can be functionally utilized as part of a synaptic junction or a signaling molecule.

[0080] “Biological outcome” is used broadly, and can range from relatively simple 20 actuation of a muscle (motor) unit through to a clinical outcome, such as to functionally overcome disease conditions, including as cerebral palsy and spinal muscular degeneration.

[0081] “Conductive polymer composite” refers to a polymer that is capable of conducting an electric charge. The conductive polymer contact may be formed from a composite material, including a mixture of several polymers and a mixture of polymer with 25 distributed metallic elements.

[0082] “Operably connected” or “in operable connection” refers to a configuration of elements, wherein an action or reaction of one element affects another element, but in a manner that preserves each element’s functionality. For example, a cellular electrode array in operable connection with the biological tissue or surface refers to the ability of the electrodearray to electrically interface with the tissue, including cells, without adversely impacting the tissue or the array.

[0083] “Electrode” refers to the component of the cellular electrode array that can be electrically energized to electrically stimulate a biological material and / or to detect an 5 electrical parameter arising from the biological material. The electrode may be a conducting metal. The electrode may be a composite, with a core material that is covered or coated by an outer material that is different from the core material. Accordingly, this portion of the electrode is also referred to as a “synthetic portion” of the electrode. The outer material may correspond to coatings of active sites where electrical connection is desired. As but one 10 example, the electrode may be coated in Poly(ethylenedioxythiophene) (PEDOT). The electrode may be configured to graft to a neuronal cell, so that substantially all of the electrical connection with a biological tissue is via the grafted neuronal cell(s). In this context “substantially all” refers to at least 80%, at least 90%, or at least 95% of electrical current that travels between the biological tissue and the electrode via the grafted neuronal cell. 15 “Substantially all” also refers to the situation where electrical signals in the form of action potentials are provided between the electrode and the biological tissue without loss or substantial degradation of the action potentials, including the frequency of action potentials.

[0084] “Device footprint” refers to the surface area defined by the perimeter of the polymer substrate. 20

[0085] “Adjacent” refers to biological tissue that is sufficiently close to the grafted cells in the wells of the electrode array is able to make electrical contact with the tissue, including biological cells of the tissue. The term is intended to reflect that there is a constrained region of contact, as the grafted cells have a spatial limit as to how far from the well in which they are contained can establish direct electrical contact. At least part of the constraint is the 25 distance from the well the sprouted dendrite or axon can grow. Accordingly, adjacent includes a functional definition based on typical maximum axon length is less than 100 μm, including between about 20 μm and 60 μm. In one aspect, any tissue that is within 100 μm from a reference location, such as an edge of a well, is considered to be “adjacent” to that reference location. 30

[0086] “Bulk electrical interfacing” refers to an electrical interface with electrodes on the surface of the device, where an electrical parameter (e.g., electrical potential) or electricalsignal (e.g., current or potential) is measured or provided directly with the electrodes without the biological / synaptic connection.

[0087] “Well” is used broadly herein to refer to a volume in a flexible polymer substrate that can reliably support growth of a biological cell under culture conditions. The well may 5 have any of a range of shapes. For example, the cross-sectional shape may be uniform, such as cylindrical in nature. Alternatively, the wall may have a funnel or taper shape, with a larger cross-sectional area toward the flexible polymer substrate top surface, tapering to a minimum cross-sectional area in a direction away from the top surface. The well can be shaped depending on the application of interest. For example, to receive a droplet containing 10 very few neuronal cells (e.g., down to about 1 neuronal cell / droplet), the well may have a tapered configuration to help guide and confine the droplet to the well.

[0088] The invention can be further understood by the following non-limiting examples.

[0089] Example 1: CEA

[0090] Referring to FIGs.1A-1C, a cellular electrode array (CEA) 1 is formed from a 15 flexible polymer substrate 10 having a top surface 20 and an array of wells 30. Each well has an inner facing surface 40, including a bottom surface, and a conductive polymer contact 50 (illustrated as positioned to form the bottom surface). Together, a non-conductive PDMS layer and a conductive polymer layer can effectively form an insulated conductive layer that is functionally an electrical interconnect between the conductive polymer contact 50 and an 20 external electronic component 51. Well depth is illustrated by arrow 60. A hydrogel 70 can be disposed in each well. For simplicity, one neuronal 80 with sprouted axon and dendrite 90 is illustrated per each well, available for contact with a biological tissue 100. FIG.1A is a side cross-section and FIG.1B a top view that also illustrates conductive top surface 20. FIG.1C illustrates the cellular electrode array 1 implanted under the skull and electrically 25 interfacing with biological tissue that is brain. For simplicity, the interface is shown in the direction of detecting a signal. Of course, the signal may be generated by the device and used to actuate biological tissue. Both the detecting and generated signal may be spatially varying over the footprint area (see, e.g., FIG.1B showing the footprint of a 4x3 electrode array).

[0091] Referring to FIG.5, an electrode 500 (which can, functionally, be equivalent to 30 conductive polymer contact 50 of FIG.1A-1C, including by being formed of a plurality of PDMS layers 501) formed of an electrically conducting material 510 having an outer surface520. The electrode may be formed of a plurality of distinct materials, such as an outer material 530 that is different from a core material 540 (e.g., the inner body) of the electrode. Line 505 represents system boundary, such as the implanted portion footprint. A neuronal cell 550 is electrically connected to the electrode, as represented by the two-way electrical 5 current arrows 560. In this manner, neuronal cell 550 is characterized as a “grafted” neuronal cell with respect to the electrode 500. As further explained herein, the electrode 500 may correspond to a CEA well that has functionality as a microfluidic cell sorter for reliable positioning of a cell in a well. With respect to a target neuronal cell 570 in a target or biological tissue, the grafted neuronal cell 550 has a biological electrical connection 580. 10 This is reflected by the two-way arrows 585, indicating that the electrode may be used to stimulate target 570 or to detect an electrical parameter of target 570. Arrows with strike- through 590 corresponds to leakage of electric current between biological tissue and electrode, reflecting that substantially all electrical connection is between the biological electrical connection 580 with attendant arrows 560 rather than “non-specific” electrical 15 connection of electrical path 590. For example, at least 80% of the electrical current may be via path 560 compared to the non-specific path 590.

[0092] FIGs.2A – 2E are summaries of various steps that can be utilized to make an electrode-neuron grafted array. Unless noted otherwise, dimensions are only exemplary. FIG. 2A illustrates multiple potential devices on one wafer. For simplicity, subsequent figures 20 illustrate one device. All units are in mm. The designs and dimensions shown are a representative embodiment. Step nos. refer to the steps further explained in FIGs 2A-4.

[0093] FIG.2A illustrates an initial step with intersecting grooves cut into the surface of the wafer with a diamond saw. An 8 by 8 well design is shown as an example, actual well density can be much higher. 25

[0094] FIG.2B illustrates second step described in step 1.b)ii below, isotropic wet etching creates rounded profiles. The profile created using 1.b)i is similar, except missing rounds near the apex and the bases of the columns. Either process is acceptable.

[0095] FIG.2C illustrates casting of the Device Body described in step 3. The thin layer of conductive PDMS (10 μm in this example) serves as the ground and can be deposited 30 before the main body or attached as the final step. The main body should exceed the top of the columns by a known amount.

[0096] FIG.2D illustrates deposition of electrodes and part of the interconnects after steps 4.a), 4.b) and 4.c). Device Body should be etched to the top of the master, then conductive PDSM is patterned by selective curing (if photo-reactive) or by etching with a metallic mask layer (if not photo-reactive). Wiring is optimized by any number of techniques 5 and processes, including but not limited to manual determination and / or using electronic design automation (EDA) software for large arrays. Note some electrodes are not connected to contact pads in this layer because multiple layers are used for dense arrays. The interconnect between electrodes and contacts can be created manually, using an auto-routing function of most EDA software. 10

[0097] FIG.2E illustrates the completed electronics after step 4. Electrical interconnects 200 may not all fit in the same layer on a dense electrode. The interconnects 200 electrically connect the electrodes to the contact pads at or toward the device edges. Interconnects connecting to various regions of electrodes may be at different layers in the device. For example, the interconnects that electrically contact the electrodes positioned in the device 15 middle region may be on a layer above the other interconnects. Accordingly, the devices can be built in layers 210 by repeating steps 4.b), 4.c), 4.d) and 4.e) in FIGs 2A-2E. The pattern of external contact shown in this example is designed to be compatible with Multichannel System amplifiers for Glass MEA (alternative 3), described in 4.f)ii. Four layers are illustrated in the top inset figure, including top insulation. One layer is illustrated as having a 20 thickness of 20 μm. The portions of the insets with the dimensions are illustrative electrode contact dimensions. Multiple layers 230 are schematically illustrated, and reflect that the interconnects and / or electrode contacts may span multiple layers.

[0098] The CEA devices provided herein can be implanted under the pia mater, flush against the brain’s surface. When axons and dendrites from grafted neurons reach the opening 25 in the well, they can synapse onto nearby dendrites from the neural tissue. The axons in electrical contact with the tissue can then conduct electric activity stimulated by the electrodes. Similarly, cells near the device will project axon collateral which synapses with dendrites in the wells, conducting signals to grafted cells. Because the device itself is dielectric (e.g., the device surface is effectively grounded so that signals cannot leak between 30 cells and unwanted signals cannot leak from tissue to the device), the electric field cannot reach neighboring wells. Thus interference is limited by the number of cells in each well, anumber controllable by device dimensions and cell culture parameters, such as plating density.

[0099] Various manufacturing processes may be employed to make an electrode array suitable for containing cells configured to electrically interface with a biological tissue, 5 including as provided in PCT / IB2023 / 060324, which is specifically incorporated by reference herein for the CEA devices and related methods of using and methods of making the CEA devices.

[0100] One example of a method of making the device includes that summarized in FIGs 2A-2E 10

[0101] Another example is provided in FIG.3, which schematically illustrates parts of a manufacturing process showing, for simplicity, a single well and its immediate vicinity, as a flow chart of steps. Of course, the process can be used for an array of wells and is compatible with any of a range of known processes in the art related to substrate micrometer-sized handling, deposition, etching, and equivalents thereto. 15

[0102] 300 illustrates a SU-8 master comprising a pillar positioned at a corresponding well location. As noted, 300 is preferably an array of pillars corresponding to wells of the microarray. Optionally, the SU-8 master is functionalized, such as by silanization, to facilitate removal of the SU-8 master from another layer of material, such as a PDMS-based layer, including by peeling of the arrays from the SU-8 master. 20

[0103] In step 310, PDMS is cast over the master 300 to form a PDMS body 312, Preferably, PDMS body 312 is not conductive. In step 320, the PDMS body 312 is etched until the top surface 322 of the SU-8 pillars 300 are exposed.

[0104] The Conductive interconnects are subsequently patterned by applying any of a variety of techniques, such as surface machining process, so long as the techniques is 25 compatible with PDMS-based materials. Conductive interconnects are formed from PDMS based conductive composite. The subsequent paragraph provides two specific examples of patterning: by masking and etching or by channel filling.

[0105] Masking and Etching: Pattern by masking and etching

[0021] : prepare positive and negative photomasks of each interconnect layer; optionally perform reactive oxygentreatment of the wafer; transfer negative photomask to aluminum hard mask on the top surface. For example, in step 330 a protective mask 332 is provided on the PDMS top surface and in step 340 a conductive path layer 342 is provided, such as by spin coating a PDMS based conductive composite. These may be: PDMS, PEDOT:PSS, Triton X-100 and ethylene 5 glycol

[0015] ; PDMS and carbon black

[0032] ; PDMS and carbon nanotubes

[0037]

[0033] ; PDMS and silver powder or silver nano wires

[0034] .

[0106] In step 350 an etch mask 352 is provided to the positive photomask, such as an aluminum hard mask that is resistant to the PDSM etching of subsequent step 360, also generally referred to as an etch down step that removes PDMS-based material that is 10 positioned over mask 332. PDMS etching is by any of a range of techniques known in the art, including: Dry Reactive Ion Etching (DRIE) using, for example CF4 and O2 or SF6 and O2; Wet etch using NMP(N-methyl pyrrolidinone) and TBAF (tetra-butyl ammonium fluoride).

[0036] ; Combined wet / dry process, where dry etch by CF4 and O2 is followed by wet etch.

[0024] ; Wet etch using concentrated sulfuric acid.

[0038] . 15

[0107] After etching, the hard masks 332 and 352 are stripped revealing wire layer 342 over pillars 300. In step 380, insulation 382 is applied over conductive path layer 342, such as by an additional layer of PDMS by oxygen plasma treatment

[0039] ,

[0040] . In step 390 the arrays are detached and in step 400 bonded to a grounding layer 402.

[0108] Channel Filling: An alternative to steps 330-380 of FIG.3 is shown in FIG.4, 20 with initial steps 310 and 320 the same as in FIG.3. In step 410 a new positive master 412 is provided and in step 470 a PDMS negative 472 of the interconnect layer is cast. The PDMS negative 472 is detached in step 475. In step 480 the PDMS negative 472 is bonded to the top surface 473 of the device body, such as by oxygen plasma treatment of both adjacent surfaces. 25

[0109] A punch hole 486 is made in step 485 to provide access via injection of a PDMS based conductive composite 491 and subsequent curing as illustrated in step 490. Note that the composite containing ethylene glycol will form bubbles when heated and should not be cured in an oven.

[0110] As in masking and etching, steps 390 and 400 are the final steps where the wells 30 and the interconnects are bond to the separately created grounding layer. Conductive grounding layer and / or arrays for bulk interfacing are bond to the tissue-facing side byreactive oxygen treatment (see, e.g., step 400). Wafers are cut into devices and bonded to wires.

[0111] A particularly important aspect of the instant devices and methods is the ability to reliably add neuronal cells to the electronic device that is in the form of a microarray. 5 Neuronal cells, in other words, can effectively be grafted to the device and ready to make an electrical connection with biological tissue. The electrical connection is fundamentally improved herein, in that each individual electrical connection between device (e.g., a neuronal cell in the well) and tissue is to as few as less than 10 cells in the tissue, less than 5 tissue cells, or as few as 1 tissue cell. The tissue cell may be in a tissue such as brain, spinal 10 cord, or a peripheral nerve.

[0112] With respect to obtaining and providing to the device an immunologically compatible cell configured to graft to the device and provide an electrical contact with a biological tissue, the following includes embodiments for the methods of cell culturing:

[0113] 1) Obtain immunologically compatible cells for the graft. 15

[0114] (a) Adult cortical neurons can be difficult to work with. However, for auto- graft, neurons can be obtained from the subject’s own Dorsal Root Ganglion (DRG). They are routinely obtained as a part of thoracic vertebrectomy

[0042] . Neurons can be dissociated from the matrix by standard protocols.

[0115] (b) Neurons are immunologically privileged. Studies show that transplanted 20 allo-graft neurons are not rejected by the immune system

[0043] ,

[0044] .Even Xenografts are known to survive

[0045] . This open up the door to use well understood Induced Pluripotent Stemcells (IPSC). Somatic cells such fibroblast can be transformed into IPSC by transient expression of transcription factors (for example using Oct3 / 4, Sox2, c-Myc, and Klf4)

[0046] ,

[0047] . IPSC linage are commercially available (for example ATCC’s ACS1019). 25

[0116] (c) Subsequently, IPSC is induced to differentiate into neurons or precursors

[0048] , for example with ASCL1, BRN2A, and MYT1L

[0025] . In addition, IPSC derived cells are commercially available (see, e.g., Elixirgen Scientific’s QuickNeuron™ Series).

[0117] (d) Alternatively, for small animals, compatible neurons are obtained from an inbred rat strain, a bio-engineered twin or a cloned embryo.

[0118] 2) Creating the culture medium following well established receipts. Either of serum-free or otherwise:

[0119] (a) Neurobasal w / B-27, 0.5 mM L-Glutamine, 1% heat-inactivated FBS, 2.5 g / L D-glucose, 20 ng / mL NGF, and 20 lM FdU + 20 lM uridine.

[0049] 5

[0120] (b) Neurobasal medium (Invitrogen, Gaithersburg, MD) supplemented with 2% B-27 (Invitrogen), 1% penicillin–streptomycin (Invitrogen), and 0.4 mM of L-glutamine (Invitrogen)

[0050]

[0121] 3) Transfer the PDMS devices to petri-dishes or culture flasks (see 5a) with wells facing up, and sterilize using one of the following techniques. Method a is preferred because 10 it will also make the surface more hydrophilic. Note that gas sterilization is not recommended because gas will diffuse into PDMS and make it toxic.

[0122] (a) By Oxygen plasma treatment

[0051]

[0123] (b) By autoclaving

[0124] 4) PDMS is extremely hydrophobic, therefore action must be taken to improve 15 surface adhesion, especially if step 3a is not taken. Cover the PDMS devices with culture media for several hours, to allow protein to absorb onto the surfaces. (a) Optionally, after covering the devices with culture media, transfer the dishes to a low vacuum, until bubbles in the wells dissolve into the media.

[0125] 5) Plate cells at a density where, on average, there is one cell per well. 20

[0126] Optionally, transfer the PDMS device and cells to a flask and use a centrifuge adapter (Beckman Coulter SX4750) to force cells into the wells. Use PBS to wash away excess cells at the surface

[0127] 6) Provide a scaffold for axon / dendrite growth. Both the following reagents are available commercially. Because these reagents are only liquid at low temperature, there is 25 only a short time window to work with them. Therefore it is not advised to suspend cells in them, rather they should be applied in a separate step after cell-plating. Of course, the methods and devices provided herein are compatible with providing them to a surface before cell-plating.

[0128] (a) Collagen (e.g. Gibco™ A1048301)

[0129] (b) Becton Dickinson Matrigel™

[0130] 7) Follow the instruction of reagent for gelling, for example, exposure to ammonia. 5

[0131] 8) Introduce fresh media every 2-3 days. Allow 1-2 weeks for axon and dendrite growth and observe for their spread along the scaffold.

[0132] Resistive and capacitive leaks between wells without grafted cells may be tested without grafted cells. It is possible to test if stimulation artifacts appears in other channels without the cells. This test is especially important for dense arrays. The device surface facing 10 the tissue is a solid conductive grounding layer, and so there should be no current flow between wells. This aspect is relevant for achieving high specificity between device and the tissue.

[0133] Because PDMS is optically transparent, the device can be directly visualized, including with a confocal microscope, by supporting the device with a microscope slide. 15

[0134] Because PDMS is soft, the device can be fixed in a dish and sectioned, such as sliced with a vibrotome, and stained (for example, with Nissil stain) for imaging, including by a transmission microscope. This aspect is useful for assessing if grafted cells are viable and living (live vs. dead stain).

[0135] Iterate design by simulation: (a) Stimulators apply bi-polar pulses. (b)This pulse 20 introduces a voltage gradient between the grounding plate facing the tissue, and the electrodes at the bottom of the wells. (c) The amount of gradient required to trigger action potentials can be determined with in silico simulation, such as by NEURON software, and LFPy package in python. (d) Simulation and in-vitro results are used to inform device dimensions, such as well diameter, depth, geometry and spacing (density). (e) Likewise 25 similar simulation can be used to optimize the array’s ability to detect action potentials propagating inside the wells towards the cell body.

[0136] Targeting synaptic connections: (a) The type of cells grafted to the electrodes of the array influences the counter-part corresponding nearby cell (from the biological tissue) that forms a synaptic contact. This can be leveraged with targeted gene expression, tofacilitate connection with only certain types of neurons in the brain. (b) This is a relevant consideration in view of the fact that the brain comprises different types of neurons in different brain regions.

[0137] The devices and methods are compatible with a range of cell types, including 5 commercially-available cell types and custom-isolated cell types, so long as they can be provided at a low cell number to each well and are able to reliably secure electrical connection to a biological tissue.

[0138] For example, commercially available IPSC derived cells may be used to target specific type of neurons. Different types of neurons (for example Motor, Sensory, 10 Cholinergic, Dopaminergic, GABAergic) preferentially form synapse with different type of cells. This can be used to communicate with specific type of cells in the vicinity of the array. Examples of commercially available cell types for use with the instant CEA devices and methods include, but are not limited to, those provided in TABLE 1.

[0139] TABLE 1: Representative cell types (Cells derived from Induced Pluripotent 15 Stem Cells (IPSC))

[0140] High specificity: The main aspects of the CEA devices and methods of using the devices provided herein, and advantage over conventional existing devices, are those that provide the extremely high specificity between the electric contact and target neurons. This is 20 achieved by providing a brain-facing surface of the device that functionally acts as a ground in combination with the grafted cells in the wells, so that electrical current can only flow as action potentials to or from the grafted cells. The grafted cells, in turn, will only synapse with one or the few cells that they can physically reach. This combination of grounded tissue-facing surface, grafted cells in the wells, and physical connection between the grafted cells and biological cells of the tissue provides the high specificity.

[0141] Power Efficiency: Furthermore, there is about one cell, or at most several cells (e.g., less than 10, less than 5, or less than 3) per contact to the target biological tissue. Such a 5 few number of cells that synaptically connect at a particular location provides a number of functional benefits, including a signal from each cell can be transformed into a spike train with a threshold circuit. This results in a greatly reduced power consumption by a device, including a device that is a head stage. Conventional head stages historically have a serious problem with power consumption, particularly for implantable / integrated head stages. The 10 instant devices and methods have reduced power consumption, particularly as the devices and methods are effectively working with action potentials instead of analog systems with attendant transmission requirements.

[0142] Noise Resistance: because there is only one or a few cells in every well, every channel will pick up spike trains corresponding to these cells, which are in effect carrying 15 only timing information. Noise from local field potentials that penetrates the wells cannot disrupt the operation because they distort the spike trains vertically rather than horizontally. The cellular electrode can tolerate a large amount of noise before breaking down, making it tolerant to internal or external noises.

[0143] High Bandwidth: The first part of the fabrication process is the same for a Utah 20 array. Therefore the instant devices can achieve the same array density as the number of needles in state-of-art Utah array. With a spacing of 100 μm, a 10 X 10 mm array can contain up to 10,000 unique contacts. Because electrical interconnects are routed through multiple conductive polymer layers, this density is possible. In practice, however, the density may be limited to the channel count of the best amplifier systems. 25

[0144] The devices provided herein also have good temporal resolution. The response of grafted neurons serving as electrical connection can be predicted using the pulse waveform, cell type and well geometry.

[0145] The combination of high specificity and resolution allow high bandwidth communication, and facilitates transmission of meaningful encoded signal in time and space 30 (like spike trains) between electronic devices and neural tissue.

[0146] Mechanical and Chemical Bio-Compatibility: The device is formed from the polymer, including from PDMS. With PDMS as a dominant component of the system, the device has excellent mechanical and chemical characteristics for biomedical applications, including interfacing with soft biological tissue. PDMS is an elastic material which can have 5 a mechanical parameter, such as a Young’s modulus, matched to the to-be-interfaced tissue. Therefore, there is minimal abrasion between the electrode array and the neural tissue due to mismatch in mechanical properties. PDMS is also biologically inert, non-toxic and tends not to trigger immune responses. Accordingly, the device is suitable for long term use with biological tissue, including in vitro, ex vivo and in vivo. 10

[0147] Being a soft and flexible device, the device can also be rolled into a cuff electrode configuration, suitable for electrically interfacing with the spinal cord or peripheral nerves.

[0148] Histological Compatibility: PDMS is optically clear and so the device can be viewed directly with a transmission microscope. The device is sufficiently soft that it can be cut with a vibrotome and scanned by a confocal microscope. This makes the device an 15 excellent solution for long-term in vitro studies, such as organotypic cultures.

[0149] Modular Compatibility: The device may be designed to incorporate head stage electronics. Alternatively, the device can also be made compatible with any amplifier that accepts a micro-ribbon connector. This can be done by bonding or directly patterning Parylene ribbons on the device. 20

[0150] To facilitate device characterization, the device may be made to have only 32 or 64 channels. This allows the device to be tested using existing TBSI or Multichannel system head stage amplifiers.

[0151] Low cost: Mass produced arrays are made with reusable masters and photo- reactive PDMS. This reduces cost at scale and makes the devices attractive for studies with 25 small animals.

[0152] Computational Model for Design Iteration: The number of cells in each well is a Poisson distribution with the probability equal to plated cell density divided by the array density. This can be used to limit number of cell per well to several each. Cells processes largely develop in a vertical orientation along the collagen scaffold against oxygen and 30 nutrient concentrations.

[0153] Detailed computational analysis for the reaction of neural tissue to injected charges is not possible in general due to the complexity of the neural tissue structure. The well-based design insulates the analysis of electrode system from the complexity of biological structures and makes the problem tractable because of the confining of electrical 5 connection via the well and grounded nature of the device surface. A theoretical understanding of the system further facilitates particulars of the design in an iterative and / or simulation, or by an AI.

[0154] Segmented numerical model (with NEURON) can be created for dendrite and axon from grafted cells (following gradients) and towards grafted cells (against gradients). 10 Simulations are performed to predict spike trains triggered by time-varying potential gradient (for stimulation) or the extracellular potentials generated by moving action potentials (for recording). These models are used to guide design parameters. Cell morphology obtained from confocal microscopes are used to verify the models.

[0155] The distribution of number of cells in wells is expected to be a Poisson 15 distribution. Therefore, to limit the number of wells with multiple cells, a large number of wells may be empty. This situation can be improved by cell capture and handling techniques provided in various micro-fluid devices.

[0031] . In this manner, the number of cells per well is controlled via microfluidics. For example, cell sorters (see, e.g., fluorescent activated cell sorters FACS) plus fluidic controllers, positioners and the like can be used to provide precise 20 number of cells per well. A FACS can provide a single cell from a nozzle. The nozzle can be connected to a custom setup on a floating air table, where a micro manipulator moves a petri dish containing the array at a fixed speed. Therefore, a single cell can be placed at each well, as desired.

[0156] Because neural signal are conducted mostly passively in dendrite and actively in 25 axons, it is better for stimulation if grafted cells send axons into the adjacent tissue. Conversely it is better for recording if adjacent cells send axons to synapse onto grafted cells.

[0157] The CEA devices described herein comprise an electrode array for a neural interface which achieves high specificity using grafted cells in concert with special conductive and non-conductive polymer layers and regions. In this manner, the devices and 30 related methods of using the electrode array devices, outperform all systems currently on the market, including for chronic applications. As described herein, any of the CEA devices maybe incorporated into the systems described herein, including for treatment of a patient suffering a paralytic injury due to a spinal cord injury.

[0158] In another embodiment, a standard surface electrode array 240 (see, e.g., FIG.1B) is fabricated on the tissue-facing layer with standard techniques, to provide the desired 5 specificity and ability to obtain a bulk aggregate electrical interface with tissue. This secondary electrode array can be positioned on the top-most biological-facing tissue instead of part of the ground. This facilitates simultaneous collection of more “bulk” or “overall” electrical signals. This embodiment is particularly useful for debugging or serving as a back- up. Of course, there is an attendant impact with respect to ground being imperfect; the 10 imperfect ground can be accommodated by increasing device thickness.

[0159] Example 2: Artificial Intelligence (AI) Enhanced Functional Electrical Stimulation for CNS Disabilities

[0160] Theoretical Background

[0161] Equilibrium Point Hypothesis: 15

[0162] There are alternative theories on the computation performed by the spinal cord. The spinal cord is not a passive conductor. The various reflex arcs in the spinal cord performs feedback controls at the local level which stabilizes the postures and movements. [4]–[7] This is also called the lambda model.

[0163] Feldman AG discovered that, if a subject performing an intentional movement is 20 perturbed by external force for brief periods of time, their movement will usually reach the same destination[8]. He called this equifinality condition and it suggests the control signal in the spinal cord is a reference trajectory. This is known as the equilibrium point hypothesis in academic literature. Further study suggests that while the spinal cord circuit seem to stabilize voluntary movement around intended trajectories, the elasticity may not be linear. Force field 25 mapped around stable postures sometimes had nonzero curls, suggesting that a perturbation in one direction could result in a restorative force not aligned with that direction, suggesting nonlinear feedback. Based on this, in order, to restore motor function, a FES should learn nonlinear control. Artificial neural networks are capable of learning arbitrary functions, and are unconcerned with linearity of the control system.

[0164] This theory is controversial; Kawato et al. argues that, instead of set-point, the brain computes a sequence of force values requires to achieve a desired trajectory ahead of the time using an internal neural representation of the mechanical system[9]. In this case, the spinal reflex arcs are suppressed during voluntary movements. Based on this, the FES system 5 will still have to learn nonlinear control but only for posture stabilization. It would also have to detect the onset of fast movements and switch to open-loop or predictive control.

[0165] Both theories may be partly true and they describe parallel descending systems. It’s plausible only some lower motor neurons exist outside reflex arcs, and are involved explicitly in discreet or open loop control. For example, when a person desires to apply a 10 specific amount of force in one direction, or to offset a specific weight they are carrying. Activating these cells requires inhibition of local feedback to prevent antagonistic muscles from automatically activating.

[0166] In the review of FES systems, Popovic discusses some modern systems with “Hybrid” control schemes that is “defined by interaction of subsystems in the context of both 15 continuous dynamics and discrete events” [3]. Those system have both open loop controllers to handle discrete events, and close loop controllers which handles dynamic control, consistent with the conjuncture that both theories of human motion planning may be correct to some extent.

[0167] Artificial intelligence (AI) may be especially applicable in light of the lack of 20 theoretical clarity. Instead of engineering controllers optimized in accordance with a specific theory, an AI can be designed to accommodate both aspects, and will learn the best control scheme in accordance with the data. Theoretical interpretation may subsequently be made based on the parameters of the AI.

[0168] An incidentally interesting consequence of the equilibrium point model is that, 25 when the subject is relaxed, their neural trajectory can be moved by passive motion. This implies that at low speed, an exoskeleton can be used to perform passive training. See “Assistive exoskeleton” section below.

[0169] Neuron transplants in spinal cord and periphery:

[0170] Pfister et. al. developed a synthetic nerve by forcing DRG neurons to grow an 30 axon track between to population of cell bodies under mechanical stretch

[0010] . The cellcultures were than covered in 80 percent collagen hydro-gel / matrigel then rolled up into a tube to form a construct. The construct were implanted into models of lateral spinal cord injury

[0011] and the Sciatic Nerve

[0012] . Histology showed extensive integration: there were both projections from host neuron into the construct, and outgrowth of transplanted cells into 5 host tissues. These studies took place without immune suppression, suggesting that neurons are immune privileged and allograft neurons do not suffer from rejection.

[0171] Interestingly, some host axon were seen to transverse the entire construct. For instance, axons were observed entering from proximal end of the construct and exit through the distal end (Fig.4 in

[0012] . This suggest regenerating axons in periphery can follow existing 10 axon tract over long distances. This was also consistent with observations other conditions, such as embryonic hippocampal neurons transplanted into the spinal cord

[0013] .

[0172] Methods and Specifications:

[0173] The invention restores both voluntary and involuntary moments in SCI patients by simulating damaged spinal cord local circuits using artificial intelligence. The AI learns to 15 maintain reflex arcs and mimic central pattern generators according to patient’s intentions intercepted at spinal cord above the lesion, intercepted sensory nerve activity, and measurements from motion sensors.

[0174] To facilitate this AI, at least one electrode is implanted on the spinal cord upstream from the lesion, and at least one electrode is implanted downstream from the lesion. 20 The downstream electrodes can be any one or more of a spinal electrode, peripheral nerve electrode, muscular electrode, or a combination thereof.

[0175] Electrodes form networks with implantable hubs housing rechargeable batteries and low powered electronics, such as amplifiers, ADC, pulse generators. Each hub manages one or more electrodes. The hubs performs minimum processing and mainly serve relay 25 signals and generate stimulating pulses. The hubs communicate with external wearable pods (e.g., communication pods) positioned in proximity, referred herein as “spatially aligned” or “proximally positioned”.

[0176] The pods form a network with one or more computing cores, either by wireless or cables. A core is a high powered computing device capable of evaluating an AI andcommunicating with cloud services. It could be a cell phone, a commercial micro-controller or custom electronics.

[0177] The onboard AI contains two types of ANN: Feature Adaptation Networks for both motor and sensory signals are multi-layered auto-encoders that compresses neural 5 signals into a lower dimension space for control. Control Models learn to simulate biological systems in close-loop feedback. The two networks are trained end-to-end offline, such as on a cloud.

[0178] Note that any component of the system may be substituted with other commercially available components depending on the compatibility. 10

[0179] Implant Network:

[0180] The system comprises a network of implantable electrodes, at least one of which is a spinal cord electrode or interface. The terms “interface” and “electrode” are used interchangeably. Electrodes at or proximal to the lesion site are referred herein as upstream electrodes. Electrodes distal to the lesion site are referred herein as downstream electrodes. 15 Downstream electrodes can be on the spinal cord below the lesion, on a periphery nerve, or on muscles.

[0181] FIG.6 illustrates various locations of system components, with a lesion at Th10 of the spinal cord illustrated by the “x” mark on the spinal cord. Circle: Electrodes (internal): Square, Hubs (internal), Triangle: Pods (external), Cylinders: Core (external), Cross: Motion 20 sensors (external), Solid line: implanted insulated buses, dotted line: external wires or wireless communication. In particular, the implantable functional electrical stimulator can have a number of electrodes, such as spinal cord electrode 601, periphery nerve electrode 602, muscle electrode 603, feedback control system 610 (including a processor, computing core, etc.), hubs 620, pods 630 (including communication pods), motion sensors 640. Shown 25 on patient or user 690 is a spinal cord lesion 695. External system can include the components that may be wearable by the patient, such as pods 630 and motion sensors 640. Processor of 610 may be within a handheld electronic device, such as a smartphone or a tablet. Assistive exoskeleton 699, including in communication with motion sensor 640, can be used to assist with active motion and, therefore, can communicate with processer 610 of 30 feedback control system, as reflected by double-sided arrows. Of course, the communication may be mediated via pods 630 and hub(s) 620 (including a pod associated with anexoskeleton 699). The muscle electrode 603 can be positioned in the thigh, on a muscle end plate band where periphery nerves terminate, for example. Muscle electrode can be used to interact with a muscle, either directly and / or indirectly via an intervening nerve. Assistive exoskeleton 699 typically functionally have mounting / integrating cores, pods 5 (communication), motion sensors, motorized or hydraulic actuators (M) that are in communication with hubs 620 and wireless network 621.

[0182] Electrodes are connected to electronic hubs located under the skin via insulated bus, such as a biocompatible, flexible and stretchable cable, including polyimide ribbon cables. The electronic hubs contain short ranged transceivers (such as induction or radio), 10 batteries, and a microcontrollers. The controller is responsible for generating clock, stimulus, timing (such as ramp) signals. The controller should always contain amplifiers (ideally one of infinite impedance) or alternatively, a spike detection circuit for reduced power consumption. In implementation where the hub functions as head stage or spike detectors (e.g. electrodes have no built-in electronics), the hub should also contain analogue digital converter and / or 15 timing circuit.

[0183] Cell based periphery neural interface: A highly specific electrode array or “cellular electrode array” (CEA) can be provided by engineering onto a polymer substrate a regular array of wells with contact at the bottom and subsequently depositing individual cells in each well. The side facing the tissue is subsequently insulated, so the only electrical 20 connections are through synaptic junction between the cells in the wells and the cells in the tissue. See, e.g., FIGs 1A-5, Example 1, and corresponding PCT App. No. PCT / IB2023 / 060324 (Atty Ref.339732: 41-22 WO) filed Oct.13, 2023, which is specifically incorporated by reference herein. See, e.g., FIG.2E (illustrating a CEA device with tissue facing side at the bottom). 25

[0184] Because this interface is highly conformal, it can be rolled up into a cuff around a periphery nerve. This can be done by reversibly attach the CEA onto a pre-stressed backing that is formed from a biocompatible, flexible and stretchable material, such as polyimide, then allow it to relax into a roll around the target nerve before removing the backing.

[0185] The CEA can take advantage of periphery cells to synapse onto targets distant 30 from the implant site. We can use, for example, IPSC derived alpha or gamma motor neuron to target muscle by following periphery nerves. These cells express fetal (embryonic)phenotype and contain growth cones, which follow existing axon tracks because they are the same lower motor neurons which innervate muscles in nature.

[0186] For recording sensations the CEA can be additionally seeded with, for example IPSC derived pesdo uni-polar DRG cells, which should sprout axons to distant sensory 5 organs (such as Golgi and the skin). Unlike in the spinal cord, sensory and motor cells on the periphery electrodes can be uniformly or randomly distributed, as the nerve is relatively small so it would be difficult to align with specific tracts. Alternatively multiple CEA with different cell populations may be used, with the sensor electrode positioned rostral to the stimulating one. 10

[0187] Muscle electrodes are a last resort in the case of periphery nerve atrophy. It appears that over time, changes in excitability of periphery neurons make it difficult to stimulate muscles in those who suffered from SCI. In these cases it may be necessary to directly stimulate the muscles. Muscle electrodes tend to be uni-directional, because the sensory nerves serving the skin may still be functional. 15

[0188] The muscle electrodes can be CEA loaded with lower motor neurons only and positioned directly over the end plate bands. The location for these are not precisely known for some muscles

[0014] , requiring relatively larger electrodes with lower resolution to cover relatively larger regions of interest.

[0015]

[0189] Cell based spinal neural interface: 20

[0190] Likewise the CEA may be used to follow specific tracts in the spinal cord in order to reduce the surgical risk and improve implant performance. In order to do this, cells are organized into specific regions on the CEA before implantation for spinal electrodes, which can be implanted on the dorsal surface forming a hemispherical cuff around the cord using the method mentioned above for the periphery nerve electrode. 25

[0191] FIGs 7-9 are schematics of electrode location, including a CEA, and projection of various exemplary cells into the cord. The neural interface forms a hemispherical cuff around the cord. Sensory cells (e.g. IPSC derived DRG pseudo unipolar cells

[0016] ) are transplanted in the dorsal region of the CEA while motor (e.g. IPSC derived alpha or gamma motor neurons) are transplanted in the lateral region of the CEA.

[0192] FIG.7 illustrates a Corticospinal Tract (CST), here Lateral CST is shown in red as large area marked 1a on the left side of the image. This normally project onto Anterior Grey column (shown in thick solid curve towards the gray matter), and synapse onto lower motor neurons. Cells transplanted with the array (motor, cortical or cerebellar) may project 5 dendrite tree into the cord in close proximity of the tract (dash dot lines), causing de- efferented neurons from the tract to target them instead during remodeling. (shown as solid lines).

[0193] FIG.8, Spinocerebellar tract (SCT), Host DRG cells enter through dorsal horns (dashed line) and synapse onto the secondary cells which either decussate and ascend 10 Anterior SCT on the other side (4b), or turn around and ascend Posterior SCT (4a) on the same side (solid lines). Implanted pseudo-unipolar cell near the dorsal horn may follow the existing axon tract of host DRG cells to the same interneurons.

[0194] FIG.9, Dorsal column–medial lemniscus (DCML), Host DRG cells enter through dorsal horns (solid line) and ascend the DCML (3a, 3b) in proximity of the tract. The CEA 15 contain DRG cells directly dorsal to the DCML (dashed lines), which should follow these tracks due to their close proximity]

[0195] For obtaining command signals, any of a variety of spinal cord cells may be used, including Alpha or Gamma motor neurons. Any type of cells that develop large dendrite trees may be used. (Table 2). 20

[0196] Injured spinal cord may have higher plasticity that allow it to reorganize in the presence of CEA. CEA may have more difficulty integrating with the nervous system in old cases, where extensive degeneration has occurred.

[0197] Table 2: List of candidate cells useful for detecting volition / command signals and their origins Cell Type Origin P C A G R Ia Ib G N E

[01] eara e ewor : e weara e componens ncu es po s, one or eac u , worn just over the hub. These are responsible for communicating with and recharging the 5 hubs. The hubs are connected to one or more processing cores via cables or local area wireless.

[0199] The cores contain either a high powered micro-controller capable of onboard evaluation of an AI and high powered radio for communicating with the cell network, or USB dock for a cellphone. A possible cost saving measure is to dock the patient’s cellphone with 10 the system since many of these tasks are routinely performed by personal electronic devices. A mobile App may perform most of the function of the core, thus reducing its cost.

[0200] In either case, the core has to communicate with one or more motion sensor packages which keep track of the position and motion of the limb; these sensors can be housed within pods and positioned as desired on the body. 15

[0201] The core could be expanded to an exoskeleton with multiple sensor packages (see, e.g., Fig.2 in

[0017] ). The exoskeleton serves two purposes: For passive training, the exoskeleton can act as a demonstrator because it’s theorized that command signal follows the physical trajectory when the subject is relaxed; and for rehabilitation, because severe muscle atrophy occurs as a consequence of spinal cord injury. 20

[0202] Software systems: The AI controlling the FES system will largely employ two types of artificial neural networks (ANNs). The first one (“adaptation onboard neuralnetwork”) is responsible for identifying and extracting relevant biological signals and reduce dimensions. It is also referred herein as Feature Adaptation Network (FN) (731 / 831 and 732 / 832, 733, 734 in FIG.10 and 731 / 831 and 732 / 832, 733, 734 in FIG.10 and FIG.13). See, e.g, FIG.10, for input feature adaptation and output feature adaptation. This network 5 will adapt to the parameter unique to each patient, such as accounting for electrode location, specific mapping between channels and cells, or electrode drifts. It will output subject- independent data in a lower dimensional space used for the controller.

[0203] Because an input may affect future outputs, this network must have an architecture capable of accounting for time, such as recurrent networks, convolution networks 10 (where convolution act on time

[0018] ), transformer networks, or a combination of thereof. In addition, the FN can be an auto-encoder, which is for example able to re-encode ascending sensory data for subject-dependent stimulation in order to transmit them back to the CNS.

[0204] The second network (“simulation onboard neural network”) (730 in FIG.10 and 830, 833, 835 in FIG.13) is responsible for learning to control the FSE by mimicry of 15 biological system. This is also referred herein as Control Model (CM). The adaptive control network may have to learn different models depending on which theory is correct (see theoretical background). The simulation learning network / adaptive controller typically implements reinforcement learning. These network are known in the art to be capable of adaptive and feedback control of robotic systems, additionally reinforcement learning is 20 inherently predictive because it’s typically trained on entire episodes. A reinforcement learning algorithm with a value table can be considered a learning finite state controller.

[0205] The two stages (feature and control) are trained end-to-end, see Guo’s 2018 paper on applying convolution network to neural signals for details

[0018] . In addition, the CM network is not unique to the patient and therefore it can be pre-trained, using aggregate data 25 e.g. from previous patients.

[0206] To summarize a possible software architecture is essentially an adaptive controller (730 or 830) which operate between two auto-encoders running in opposite directions (See, e.g., 731 and 732 in FIG.10). However this should be considered as only one possible implementation. Other implementations include reinforcement based adaptive controllers. 30

[0207] FIG.10 is a system overview treating the patient and the FES as a single system. The software can be seen as a close loop adaptive control sandwiched between 2 auto-encoder architectures. The invention restores both the outer / discrete control path from brain to the muscle (710, 721, 731,732, 722, 711), and the inner / close loop control path (711, 723, 730, 732, 722, 711) from the brain to the cord to muscle. Solid line: onboard software processing, dashed line: biological signals, dotted line: remote services, triangle: Electrodes]. 5 The Inner Control Loop is from motion measurement, to the control model, to output feature adaptation, to periphery stimulator, to neural muscular system, back to motion measurements.

[0208] Because training an AI require a tremendous amount of power, data can be uploaded onto the cloud for AI training. The trained parameter tensor can then be downloaded to the core. This ensures that the system will still function if cut off from the 10 network, even if it is unable to improve.

[0209] Rehabilitation: An important aspect is training the AI and the patient as a single system. Kinetic and Kinematic measurements are collected by healthy trainers wearing the sensor package while they perform skilled functional motions, such as drinking from a mug, picking up an item from the floor, writing with a pen, and so on. The measurements serve as 15 training data for supervised learning.

[0210] For example, to train the adaptive Prosthesis, the patient can be asked to attempt to perform each type of movement in a controlled environment (could be a VR or Haptic simulation). The AI can assume the patient’s intention is the command signal, and learn the relationship between the neural signal and the correct motion previously performed by the 20 demonstrator.

[0211] The training for the first group of patients is fairly tedious as the AI must learn from scratch. After which, the control model can be pre-trained on the patient data, as the Control Models should be fairly similar between individuals. Therefore only the fairly shallow Feature Adaptation Network need to be tuned for individual patients. 25

[0212] This is obviously challenging for animal tests. We have therefore designed an alternative method of passive training using exoskeletons by driving passive movement. Because this method takes advantage of a feature of equilibrium point theory, where the equilibrium trajectory can be passively manipulated in relaxed subjects, that method should only be assumed to be valid when the movement is slow.

[0213] No system can be 100 percent safe because even able-bodied persons can be danger to themselves. The goal here is to provide a standard of living as close to able-bodied as possible, with a level of risk comparable to or below day to day living of an able bodied person. Different safety protocol are designed to protect researchers instead for animal tests. 5

[0214] Design Considerations: Popovic’s review [3] identifies the physical requirements of an effective FES system: “The electrodes must be designed in a manner that will not harm the preserved sensory–motor systems and will not deteriorate due to environmental conditions. The connectors should be safe, small and allow eventual replacement of the electrode or the stimulator. The power supply for the stimulator should be rechargeable and 10 also reprogrammable. Ultimately, this objective requires efficient energy transmission from and to the stimulator.”

[0215] The systems provided herein can accomplish this using living cells as contacts which result in minimum trauma due to difference in mechanical properties. The simulator, which is the hub for the downstream implant (see, e.g., FIG.6) is located at an easily 15 accessible location and can be replaced independently.

[0216] The review also mentioned directions for improvements for FES systems: “Neither surface nor implantable systems provide adequate feedback to the user. The sensors need to be integrated into a net- work that provides data similar to a natural sensing network. In this domain, it is even more important that neuroscience research and clinical testing 20 provide solid evidence regarding the optimal means of feedback. The feedback must allow the use of the system at the subconscious level and support its integration with the preserved biological control mechanism. Sensor development is equally important for the use of FES and the feedback provided to the user.”

[0217] The systems provided herein can overcome this by directly transmitting encoded 25 sensory information through the spinocerebellar tract and DCML via transplanted neurons (although the system can use a cortical electrode), and by learning to correctly encode sensory information by training the patient and the AI in close loop. The system may also query the patient on whether Conscious Proprioception has been correctly received. The system can avoid the Spinothalamic Tract as any error encoding pain or temperature would 30 cause suffering. Note this information is only omitted to the patients, the AI can still mediate any reflex arcs triggered by pain or temperature. For example, The AI will reflexivelywithdraw a patient’s hands from a hot stove, but they will only be notified their hand has changed position and is no longer in contact with a surface.

[0218] Example 3: Skilled movement restoration

[0219] A primary objective of the systems and methods provided herein are to restore 5 skilled movements in cases of SCI, and other CNS conditions by using an AI to perform adaptive control, substituting injured biological circuit. To date, no FES system has this kind of capability. See, e.g., Popovic

[0010] ^ (2014).

[0220] Conventional FES systems typically employs electrodes implanted directly in muscles. A fundamental disadvantage of those systems is that they almost always recruit 10 nearby muscles (depending on the level of activation). The instant biohybrid electrode arrays (including CEA systems) can be used to form more specific connections with individual muscle units because the transplanted cells that are part of the biohybrid electrode can follow existing tracks to their target tissue. This enables a FES system which allows precise control of muscles as if they are robotic actuators. 15

[0221] While not practical (nor useful) to explicitly directly map electrode channels to muscle units, an AI is able to learn activation of which units will lead to improved performance. Additionally, it is theoretically contentious if the neuromuscular system can be better approximated as open or closed loop control at high speed. However, Reinforcement Learning algorithms are able to learn both feedback and predictive controls. In other words, it 20 is able to simulate damaged local spinal cord circuitry.

[0222] To reduce the risk of complications and dissipate power away from the nervous system, the FES can be implemented using a network of biohybrid electrodes connected to hermetically sealed, implanted electronic hubs, which communicate with external wearable pods in physical proximity. The pods can, in turn, communicate with a high powered device, 25 such as a mobile phone, which communicate with networked services and optionally perform local AI inferences.

[0223] The use of CEA greatly reduces the power consumption because each electrode contact normally produces a spike train, which removes the need for expensive local processing in order to manipulate or transmit rapidly fluctuating local field potentials.

[0224] The system can also integrate with an optional assistive exoskeleton to facilitate training and rehabilitation. In addition, the system can optionally provide an alternative path for proprioception, for example by directly stimulating the somatosensory cortex.

[0225] Cell Generation: CEA is characterized as a biohybrid device because they 5 contain cells grafted into wells that facilitate the interfacing with target tissue. Clinical applications of the systems provide herein require functioning neurons, including from non- primary sources. While neurons are immune-prevailed, ideally they should be generated from the patient’s own somatic cells. This is generally done in three steps. First, Induced Pluripotent Stem Cells (iPSC) is derived fromfibroblasts, then, neural stem cells (NSC) and 10 neural precursors (NPC) are generated from the iPSC for the upstream spinal electrodes, lastly, the NSC and the precursors are committed to specific sensory, motor and cortical linage for downstream and cortical electrodes. In some cases,fibroblasts can be directly reprogrammed into neurons. This is not necessary for animal tests if primary cells are readily available. 15

[0226] Somatic cell to iPSC: iPSC can be generated from somatic cells as an intermediate step for producing neurons. I. Velasco et al.

[0015] ^ )(p.2811) describes this is accomplished using 4 transcription factors (“(OCT4, SOX2,KLF4, and c-MYC) was achieved for thefirst time with mouse cells in 2006 and with humanfibroblasts in 2007 using retro-viral vectors. Many researchers have produced induced pluripotent stem cells (iPSCs) 20 by expressing the Yamanaka factors through integrative or nonintegrative methods.”). Those factors are further confirmed by Ymamanaka

[0016] (p.663) (“Oct3 / 4, Sox2, c-Myc, and Klf4, under ES cell culture conditions.”).

[0227] Neural induction of iPSC: The neural linage appears to be the default fate of stem cells in the absence of other signals. This is known as the “default model” of neural induction 25

[0017] ^. There has been a number of study of deriving NSC from iPSC with various degrees of success and consistency. For a overview of differentiation method methods, see A. A. Galiakberova and E. B. Dashinimaev

[0018] ^.

[0228] There has been a fairly detailed protocol for an induction process that does not use any viruses by S. Bell et al

[0019] ^. Precise media formulation are provided on page 14 and 15. 30 A method for generating NSC and NPC at scale using an adherent culture system and no embryoid bodies (EB) was shown in a well illustrated publication by L. D’Aiuto et al.

[0020] .^Yet another method using botulinum hemagglutinin was published by M. H. Kim, N. Thanuthanakhun, and M. Kino-oka

[0021] ^.

[0229] As further alternative, review

[0015] ^ by Velasco describes (p.2812; Fig.1) various protocols for neuronal differentiation of human PSCs (embroyoid bodies and treatment with 5 retinoic acid; neurectoderm inducer; pharmacological inhibition of transforming growth factor-b and bone morphogenic protein pathways (dual SMAD inhibition)).

[0230] Motor cell generation: Spinal motor neurons can be reliably produced from neuroepithelial cells (NE) derived from iPSC. See, e.g., B. Y. Hu et al

[0022] ^ on page 4337- 4338. 10

[0231] Sensor cell generation: DRG-like population sensory neurons were generated at scale in a study by T. Deng et al.

[0023] ^. Pseudounipolar neurons can be produced from Pluripotent stem cells (PSC) using a method published by T. Deng et al.

[0023] ^. The exact formulation for NOCI induction mediums can be found in Hu at Fig.1a. CEPT is a molecule cocktail containing Chroman 1, Emricasan, Polyamines,and Trans-ISRIB

[0024] ^ It’s frequently 15 used in human stem cultures. Note that iPSC derived sensory neurons are also available commercially, for instance Reprocell Inc. RCDN004N.

[0232] Cortical Cell Generation: A review paper by

[0025] ^ included a summary of protocol for generating cortical Telencephalic Glutamatergic Neurons in Table 1 on page 1628, using the various factors. GABAergic cortical interneurons can be derived from a stem cell line. 20 See, e.g., p.559 of A. M. Maroof et al.

[0026] ^.

[0233] Direct Reprogramming of somatic cells and PSC: It is known in the art that a combination of 3 expression factors can directly reprogram somatic cells into neurons, allowing allogenic grafts. See, e.g,. Velasco

[0015] ^ at p.2814 and Fig.1 at p.2812. Fibroblasts can be reprogrammed directly into spinal motor neurons using a combination of 7 25 transcription factors. See, e.g., p.206 and Fig.1 of E. Y. Son et al

[0027] ^.

[0234] Application of CEA Electrodes: Medical application of the CEA Electrodes require culturing the electrodes with appropriate cells in vitro and implanting them in the patients. At the minimum it requires one electrode array rostral to the lesion to receive command signals, and one array caudal to the lesion to actuate muscles. The rest of this 30 section will discuss various candidate locations and associated protocols. The CEA array ismechanically similar to known conformal electrode arrays and can adopt known surgical protocols. Exemplary methods of constructing an CEA Arrays are provided herein.

[0235] The CEA electrodes can receive various formulations of cells and scaffold both on its surface and in the wells. In addition to their flexibility, this allows them to function as 5 Nervous Tissue Constructs (Tissue Engineered Nerve Constructs TENG are examples of Nervous Tissue Constructs) with additional electronic components. Results for Nerve Tissue Constructs reflect the physiological properties of the CEA electrodes. Upstream Synthnerve Electrode: Rolled Electrode in Lesion: A study by B. Q. Lai et al.

[0028] ^ exemplifies a construct containing (transgenic) Neural Stem cell and Schwann cells in a collagen sponge 10 scaffold. That study showed extensive growth of host neural tracts into the construct and functional improvements that suggest synaptic integration. Lai at p.5.

[0236] CEA electrodes can be configured to be implanted into a spinal lesion. A CEA electrode can be seeded with neural progenitors. An additional coating of collagen sponge can be applied to its surface, followed by neural progenitors and Schwann cell. The electrode 15 can then be rolled into a spiral. In this configuration, an electrode surface forms a lumen enclosing the collagen sponge, and is designed to attract axons from host neurons to facilitate the interfacing.

[0237] While this configuration can be effective in complete transection models, clinical application is expected to be limited due to the relatively small number of cases. As complete 20 tetraplegia is estimated to be less than 20% of the cases and complete paraplegia is estimated to be less than 30%. Cavitation only forms if necrosis is substantial. It can, however, be extremely useful in animal models. Accordingly, any of the FES and related methods provided herein may be practiced on a non-human animal.

[0238] Spinal Cuff: A CEA of the FES provided herein can also be configured as a 25 conformal array on the dorsal side of the spinal cord. It can be implanted using known technique for epidural simulators via dorsal laminectomy and sutured to the dura. Exemplary implantation techniques are described by Gad et al.

[0029] ^ (p.6, Fig.1). However, the instant CEA is smaller and can fit under one vertebrate, there is no need to pull it into place with a suture. 30

[0239] Additionally, CEA is generally unable to penetrate the Dura and may be too thick to fit under the dura, therefore the dura under the area with wells is preferably opened prior toimplantation. This can be accomplished surgically by making small incisions, by biodegradable guides, or using retractable needles. See, e.g., D. McCreery et al.

[0030] ^ at p.197, Fig.4 (implanting arrays of micro-needles into the spinal cord using a tool which applies the array at a controlled velocity). Accordingly, any needle array, including an array of 5 microneedles, can be used as a guide to create opening on the dura and subsequently retracted.

[0240] Cortical electrodes: A CEA of the instant FES can operate in Electrocorticography (ECOG) configuration using a standard cranial window procedure for flexible (film) electrode arrays, including the method described in Yeager et al

[0031] ^. 10

[0241] The method can be adjusted because the CEA is not able to penetrate the dura mater. The array can be implanted on the cortical surface or on the pia mater. For this reason a larger cranial window is used.. Adhere the array to the cortex by surface tension over the motor or auditory and somatosensory areas. Close the dura over the array, leaving only a small incision where the ribbon cable exits.” In addition, the control pods can be secured to 15 the bone screws using dental acrylic in animal tests.

[0242] Modifications are made for human patients as any open wound on scalp and skull is an unacceptable risk of infection. In medical applications screw electrodes are avoided as the CEA has built in grounds.

[0243] Additionally, the insulated ribbon should be routed between the skull and the dura 20 to the base of the skull, then follow the spine to the control pod under the armpit (see the section on Subcutaneous Transmitters.) This can be accomplished by guiding the ribbon cable with a lubricated Neodymium magnet. See, e.g., Figure 1 of U. J. Jeong et al. et al

[0032] .^ Alternatively, the control hub of the cortical electrode can be specifically constructed from a high strength material e.g. titanium or ceramics to replace a section of the skull (for example 25 by using CT scans), and attached to adjacent skull sections by cementing (such as with polymethyl methacrylate PMMA or calcium-phosphate)

[0033] ^.

[0244] Downstream CEA Electrode: D. B. Popović

[0010] ^ identified certain physical requirements of an effective FES system (avoid harming preserved sensory-motor systems; avoid deterioration due to environmental conditions; safe; small; replaceable electrodes or 30 stimulator; rechargeable supply; reprogrammable; efficient energy transmission).

[0245] The FES provided herein accomplishes this using living cells as contacts which result in minimum trauma to neural tissue due to difference in mechanical properties. The simulator, which is the hub for the downstream implant (see FIG.1A-1C) is located at an easily accessible location and can be replaced independently. 5

[0246] Periphery Cuff Stimulator: Integration with the host periphery nervous system was previously observed in study involving transplantation of nervous tissue constructs. In a study by

[0013] ,^ constructs containing GFP positive neurons replaced a section of the sciatic nerve in the rat. The cells in the construct were observed to follow the host nerves for a significant distance with a suggestion that the neuritic bundles may be myelinated. These results suggest 10 that transplanted cells can follow host periphery nerves over long distance. Once again, because CEA electrodes may be considered a neural tissue construct with additional electronic component, those results are applicable for the instant CEA configuration. As additional evidence, figure 4 on page 8 on a paper by K. S. Katiyar et al.

[0034] ^ on TENG showed transplanted Neuron penetrate a significant distance into the distal nerve stump. Only 15 5mm of the distal nerve stump was imaged in that study, but the images show the transplanted cell extending to the edge of this zone. Periphery sensory electrodes are similar, except they are preferably seeded with sensory population of cells similar to a DRG

[0247] After the initial shock phase, some reflex arcs will recover in Chronic Spinal Injury patients, making them spastic. This means they are stiffer than able-bodied persons 20 and resist fast movements. This imposes a drawback on stimulation electrodes in the periphery: Directly stimulating muscles will activate muscle spindles and tendon organs, triggers reflex arcs and automatically activate antagonistic muscles. This co-activation impedes fast movement and precise control. For these electrodes, Selective dorsal Rhizotomy can manage the sparsity to achieve skilled movements. 25

[0248] Selective dorsal Rhizotomy: A procedure is known to manage cerebral palsy. See, e.g., K. Aquilina, D. Graham, and N. Wimalasundera

[0035] (p.2 and Figs 1-2). See also, T. S. Park and J. M. Johnston

[0036] ^.

[0249] Muscular Stimulators: As an alternative to spiral cuff electrodes, stimulating CEA electrodes can be implanted on muscles, positioned directly over the end-plate bands. 30 Compared to the spiral cuff simulators on periphery nerves, this approach uses multiple electrodes, but the density of wells on each can be lower. However, if positioned directly overmuscles, the CEA axons does not have to be myelinated. This may be relevant in specific medical conditions cases, for example multiple sclerosis (MS), where the host immune system attacks the myelin sheath.

[0250] Spinal Stimulators (Caudal): Another attractive alternative is a stimulating cuff 5 electrode in or around the spinal cord distal to the lesion. It is identical to the spinal recording electrodes, with the exception that it contains upper motor neurons. A spinal stimulating electrode is capable of modulating the local reflex arcs, thus eliminating the need of a Rhizotomy. A. Iwata

[0012] at p.108^ showed that DRG cells are able to penetrate into host spinal cord, when implanted in the lesion. 10

[0251] Optional proprioceptive and tactile Feedback: D. B. Popović’s review

[0010] ^ notes that implantable systems do not “provide adequate feedback to the user.” mentions directions for improvements for FES systems related to adequate feedback. The FES systems presented herein address this by directly transmitting encoded sensory information from a downstream recording electrode directly to the somatosensory cortex. Somatosensory reorganization is a 15 well-documented phenomenon. See, e.g., T. R. Makin and H. Flor

[0037] ^;

[0038] ; and

[0039] ^. So long as somatosensory information is directed back to areas that previously served areas affected by the spinal cord injury, the stimulation does not have to be accurate on a microscopic level. So long as a patient receives adequate training, they are able to learn to interpret the data via Somatosensory reorganization. For this reason, the devices and methods 20 preferably accommodate an interface scheme that is a training scheme.

[0252] Science Corporation conducted a study where a rigid array of microwells fabricated from SU8 carrying optogenetics cells were transplanted into mice [2]^ (transplanted embryonic neurons penetrated into the host cortex and formed functional connections).

[0253] Control Hubs and Power Reduction by CEA: CEA are connected to low powered 25 electronic subsystem that serves to amplify, digitize and transmit spike trains as well as generate pulse trains for stimulation, herein referred to as Control Hubs or, “hub”. To carry out their functions, Hubs preferably contain power means (e.g., batteries), microprocessors, memory blocks, ADC converters, DAC converters, Spike detectors, amplifiers, pulse generators, transmitters, receivers, inductive chargers and / or integrated circuit performing the 30 functions of one or more of the above. These subsystems serve similar function to head stages in electrophysiology, but with two major differences.

[0254] The first difference that hubs are hermetically sealed and fully implantable; the only openings on the casing are ports for insulated ribbon cables from the CEA. Prior to implantation, the CEA is attached, and the ports are sealed with polymers (e.g. PDMS). This allows for clinical application because open wounds are avoided for medical applications. 5

[0255] The second difference is that hubs do not necessarily contain amplifiers or buffers. The CEA can obtain spike trains from one or a few neurons. This can greatly reduce the power required for processing and transmission because the hubs need only to encode and transmit spike timing or spike count. Specific designs of low-powered implantable electronics are provided herein. 10

[0256] In human patients the hubs are preferably implanted in subcutaneous device pockets, similar to Subcutaneous Implantable Cardioverter Defibrillator (S-ICD) or other Cardiac Devices, including by adopting methods provided in Ferrari et al

[0040] (including Fig. 1 at p.224)^.

[0257] Wearable Network: Hubs are in turn supported by a network of wearable systems, 15 including pods. They manage communication, power and perform high powered computations necessary for AI inferences. These wearable components communicate with each other via low area wireless such as Bluetooth and / or cables. AI inference takes place locally for safety reasons, however AI training can take place remotely on a cloud or by high powered server. 20

[0258] Conformal Communication Stations (pods): The wearable components includes “pods” or “pod” (also referred herein as “Stations”) (“pod”), one for each hub, worn just over the hub. The pods are responsible for communicating with and recharging the hubs, as well as taking constant motion measurements. To perform their functions, a pod may contain batteries, microprocessors, memory blocks, ADC converters, DAC converters, amplifiers, 25 transmitters, receivers, kinetic sensors, position sensors, inductive chargers, USB ports, GPUs, TPUs, or IC performing the functions of one or more of the above.

[0259] The pod is much larger than the hubs and can have various shapes, designed to conform to their location on the body. For instance, the communication pod for a cortical hub may also be an earphone; the communication pod for the arm may be shaped into shoulder 30 pads. the communication pod over a hand can contain a glove. They are typically attached with elastic belts and may contain memory foam padding to improve comfort as they areintended to be worn over long periods. Conductive gels are preferably avoided. The exact form of communication stations / pods is recognized by those skilled in the art.

[0260] The pod(s) should be worn over hub(s) mostly for recharging efficiency. The communication stations can carry a significant amount of battery allowing them to recharge 5 the hubs by inductive coupling. While low frequency radio can pass through a significant amount of tissue, charging efficiency is greatly reduced and the eddy current can heat the intervening tissue.

[0261] Wearable inertial sensors are known in the art of rehabilitation. See, e.g., T. Liu et al

[0041] at p.979 and Figs 2-3) (gyroscopes; accelerometers). See also E. Rocon et al

[0042] ^. 10

[0262] Additionally, it is desirable for pods having Communication Stations functionality to have both wireless and wired communication capabilities. They should already have USB ports for charging, allowing them to optionally communicate via cables add a layer of redundancy.

[0263] Processing Cores: AI inference (but not necessarily training) should be conducted 15 locally for safety, including to avoid system collapse if the patient is cut off from the wireless network. AI inference typically requires the use of Graphical Processing Units (GPUs) or Tensor Processing Units (TPUs), which are high powered devices. The Cores should also be responsible for communication with a cellar network, for upload of data to remote AI training services. There are three options to implement Processing cores. The combinations of any 20 two or three approaches are viable.

[0264] Distributed Processing: The systems herein facilitate a direct integration of GPUs or TPUs into the pod (Communication Stations), distributing the inference over them. This makes the pods more expensive individually, increases their heat production, and rapidly drains their batteries. The benefit is the redundancy. 25

[0265] Dedicated Processing Electronic: The systems herein also facilitate engineered dedicated electronic module to perform the AI inference, or modify a desktop grade graphics card. Such a device would for example, be worn as a backpack or integrated into a wheelchair. Dedicated systems can include liquid or forced air cooling, making them potentially much more powerful. The drawback is weight, power consumption, and expense.

[0266] Patient handheld electronic: A possible cost saving measure is to dock the patient’s handheld electronic (e.g., cellphone, tablet, etc.) with a communication station since AI inference is routinely performed by modern personal electronic devices. A mobile App can perform the functions of the core, thus reducing the cost. A docking port can be added to 5 one of the hubs (communication stations) for this purpose. The obvious drawback is the challenge with using their cellphone and FES at the same time; as handhelds become more powerful, this drawback is reduced; alternatively, a dedicated smartphone can be utilized.

[0267] Assistive Exoskeleton: The FES system can be supported by an optional exoskeleton with multiple sensor packages (see, e.g., Fig.2 in

[0043] ^). The exoskeleton serves 10 three purposes: For kinetic measurement, as its joints are able to measure the moment or torque applied on them, measurement of force and torque is otherwise difficult without a rigid structure; For passive training, the exoskeleton can act as a demonstrator because it’s theorized that command signal follows the physical trajectory when the subject is relaxed (as described herein); and for rehabilitation, because severe muscle atrophy occurs in the 15 aftermath of spinal cord injury

[0044] ^.

[0268] Hydraulically actuated exoskeletons system are known in the art. See, e g., R. Kobetic et al.

[0043] ^ (p.451 and Fig.2).

[0269] Machine learning and AI Architecture: Feature layers / Autoencoders: The inputs and control variable of this system exists in high dimensions, for instance CEA can have 20 thousands of even tens of thousands of channels. The state of a limb can be described in a very small number of variables in comparison: For instance, the human arm have 6 degrees of freedom, plus 2 per finger. The system state can be fully described with these angles, plus their derivatives (angular velocity).

[0270] There must be a Feature Adaptation Network F which reduces correlated high 25 dimensional neural volitional signal ^^ே(^^)to uncorrelated low dimensional command signals ^^^(^^). Unsupervised learning algorithm are designed for this purpose, and any of them could be used to initialize the feature layers. This includes Principle Component Analysis (PCA), Generalized Adversarial Network (GAN) and Restricted Boltzmann Machine (RBM). Alternatively, a Visual Transformer can detect relevant neural signals. Once initialized, back 30 propagation of reinforcement errors can adjust the weights of the feature layers to select features most relevant to control.

[0271] Likewise there is an Actuation Network A which expands the low dimensional manipulated variable ^^(^^)back to high dimensional signal ^^ே(^^)for the stimulating electrode. The Actuation Network can be initialized using “Impulse Tests”, where a small number of motor units are stimulated. This generates microscopic movement that could be 5 measured by haptic systems with enough precision. The relation between the electrodes activated and the various DOF measured by the comm stations can be estimated as a matrix, and the Actuation Network can be initiated with the (pseudo) inverse of that matrix.

[0272] Reinforcement learning for adaptive control: Because the system has a biological plant, it faces an additional challenge. It is obviously not possible to back-propagate error 10 gradient through a biological system. This can be addressed by using reinforcement learning algorithms. Reinforce is actually an acronym for "REward Increment = Nonnegative Factor x Offset Reinforcement x Characteristic Eligibility"

[0045] .^ This encompasses an entire category of algorithms including Actor-Critic(A2C), Deep deterministic Policy Gradient (DDPG), Trust Region Policy Optimization (TRPO), Proximal Policy Optimization (PPO), Soft Actor- 15 Critic (SAC), Dynamic Sampling Policy Optimisation (DAPO) and so on Reinforcement learning have been used extensively in robotics to implement adaptive control

[0046] .^ Likewise any variation or combination of known reinforcement learning algorithms could be used to control this FES system.

[0273] To formulate the control problem as a reinforcement learning problem, we define 20 the following. During each time window or step of 1-100ms, The reinforcement system observes neural states ^^ே(^^) from the upstream sensors, which encodes subject volition, and the kinematic state ^^^(^^) from wearable sensors. It must take an action ^^ே(^^), which is the signal sent to stimulation electrodes. After which the system the system receives a reward ^^[^^^(^^)] and transition to the next state / time step, observing by ^^ே(^^ + 1) and ^^^(^^ + 1).25 the reward ^^[^^^(^^)] is typically some measure of performance as a function of the physical motion, and can be measured by additional external auxiliary sensor (required only for training and not for operation of the system), for instance, body weight supported during gait or forces on a grasped object. Typically ^^ே(^^)is known as driving input in the Control Systems Engineering literature, but is considered a state for reinforced learning. Adoption of 30 existing reinforcement algorithm following the above formalization is known in the art.

[0274] In systems with an assistive exoskeleton, the control variable ^^(^^) is used to drive the exoskeleton directly. Systems with downstream recording electrode will have anadditional feature adaptation network ^^^which converts the tactile and proproceptive signal ^^ே^(^^) to the decorrelated low dimensional vector ^^^^(^^) (input feature adaptation,733 of FIG.10).

[0275] Care must be taken to ensure gradient flow from the reward J to the parameters of 5 actuation network ^^^so the reinforcement algorithm learns better contact mapping to improve actuation. This could for example be accomplished by adding a second Q network in a DDPG algorithm. For specific implementation, see embodiment F.

[0276] Note that reinforcement algorithms are predictive by default because of the way they are trained. Reinforcement learning is trained to account for future outcomes by 10 evaluating rewards over an episode, using discounted future rewards, using time differences, or by experience replay. Due to their ability to learn both feedback and predictive control, Reinforcement algorithms are particularly useful machine learning algorithms that support the feedback control system.

[0277] Training and Rehabilitation: Integrated Training: As the patient and the AI are a 15 part of the same feedback control loop, they should be trained as a single system with a single criterion. Training should take place in multiple phases establishing increasing level of mobility. This is because neither subject nor the AI is able to make meaningful improvement if nothing they do makes a meaningful difference because the task is too difficult. This is generally referred herein as a “training scheme” of the feedback control system interface 20 scheme.

[0278] The first stage is to initialize the actuation mapping A by impulse test. This can be done by stimulating individual or a small number channels in sequence or at random. Sending a burst of spikes to a motor unit will generate a microscopic amount of force. This will typically not lead to a deflection as the force is not sufficient to overcome gravity or static 25 friction. However, if the weight is supported (for example by a haptic system), this force is measurable with an instrument with sufficient precision. A single motor unit in humans typically produces a force in the range of 0.1 to 10 Newtons. Partially stimulating it with a single channel produces forces at the lower end of the range of this. The name “impulse test” originates in control system literature where a single spike is used to test system response. 30

[0279] The second stage aims to achieve stability of posture in the absence of neural input. That means standing upright in the absence of neural commands and returning to anupright posture when perturbed. This phase establishes the safety of the system in the absence of volitional input.

[0280] The next phase involves practicing limited volitional movement under weight supported conditions, such is walking on a treadmill while wearing a harness. The patients 5 can be assumed to make the best effort to optimize the volitional command signal, based on their memory and the available feedback, such as the observation of the training session. However, we must assume they are unable to accommodate for unpredictable changes in system dynamics that arise from reinforcement learning. The validity of these assumptions can be seen with an intuitive analogy. Trying to learn to use a control system with non- 10 stationary dynamics would be akin to learning to play baseball while forced to switch between completely different bats after each swing.

[0281] Therefore, to allow the patient to familiarize with the system dynamics, they must be trained in long sessions during this phase, and the AI should only update its parameter between sessions. Alternating training between the AI and the subject can benefit from the 15 use of Trust Region Methods and / or Dynamic Models. For an overview of how this can be accomplished specifically for DDPG, see below.

[0282] After the patient has become a skilled user and the parameter of the AI parameters are near local optimal values, the patient is trained to perform skilled movement (optionally with assistance from the exoskeleton) in the final phase. They will also be given the option to 20 switch the AI to online learning, where it will make adjustment after each trial, or even in real time. The gradient can be capped to avoid collapse caused by leaving the trust region).

[0283] Safety Protocols: Mobility assist system are designed to prevent injury to the patient. However, a distinction must be made between intentional and unintentional self- harm. For instance, no practical safety margin can prevent a patient from intentionally 25 stabbing themselves. For any hard-coded margin, the patient can use a longer knife.

[0284] However, the system is capable of preventing most accidental injuries using a physician prescribed velocity limit, and a voice-activated cutoff function. The speed limit provides the patient ample time to react and issue a voice command, and even quadriplegics typically retain their ability to speak. Since the system is designed to be stable in absence of 30 neural input, upon receiving a cutoff command, it will make the best attempt to come to a stop and remain in the upright position.

[0285] Treatment of other conditions: Method and specifications takes place in context of SCI by default. However there are two other conditions where this invention is particularly applicable.

[0286] Spinal muscular atrophy: In spinal muscular atrophy, lower motor neurons in the 5 ventral horn of the spinal cord are lost. The ascending DRG sensory neurons are relatively unaffected. The system can be applied with minor changes. The grafted motor neurons should have no problem following existing periphery nerves. However, spinal muscular atrophy is a genetic condition, meaning that motor neurons derived from the patient’s own cells would inherit that condition. Genetic editing (e.g. by virus vectors) must be used to over-express 10 survival motor neuron (SMN) protein to ensure that their survival.

[0287] Spastic Cerebral Palsy: Spastic cerebral palsy, the most prevalent subtype of CP, accounting for approximately 70% to 80% of all cases, is frequently associated with damage to the motor cortex and / or the pyramidal tracts. While pyramidal cells can be lost in some cases, the more typical pathological finding especially in premature infant, is damage to the 15 corticospinal axons in the developing white matter. Since periventricular leukomalacia (PVL), a main cause of Spastic Cerebral Palsy, causes lesions to the corticospinal tract inside the brain, the main purpose of the invention is to infer the intent from partial motor signals. This require no additional changes because the amount of redundancy in neural signals (e.g. the dimensionality of neural signals is far larger than degrees of freedom, as described 20 herein).

[0288] This can be explained intuitively, such as by hypothetically assuming that a cerebral palsy patient can only send command signal to their biceps femoris muscle (which is of course, unrealistic). They would be unable to perform any action other than flexing their knee joint. However, this invention can reinterpret the commands to the various motor units 25 of the biceps femoris muscle as commands to other muscles, allowing the patient to perform other actions. The patient must adjust the encoding of their command signals to accommodate the signal, and they would be able to achieve this by training. Selective Dorsal Rhizotomy is used for managing Spastic Cerebral Palsy. Optionally a cortical recording electrode on the motor cortex can be used. 30

[0289] Example 4: Exemplary Embodiments

[0290] An embodiment herein illustrates how the instant FES systems can be implemented as a test in rats, recognizing one skilled in the art can make adaptations relevant for human patients. Electrodes used in this system may include the CEA described in PCT Pub. No. WO2024 / 079699, which is specifically incorporated by reference herein. 5

[0291] Electrode Construction: The specificity CEA, and its ability to record spike trains instead of LFP depends on an array of microwells containing a single cell each. WO2024 / 079699 describes how this can be achieved by a combination FACS and micro manipulators. Described herein is an alternative microfluidics based method with improved scaling characteristics. 10

[0292] Microfluidics chamber design: An array of wells with funnel-like cross-sections open to both sides can function as cell traps in a microfluidics chamber, ensuring a single cell per well (FIG.11). The chamber contains 2 vessels with the Array in-between with the larger opening facing down. Both vessels can be connected to perfusion pumps maintaining a continuous flow. Media is cycled through upper vessel A, while a cell suspension is cycled 15 through the lower vessel B. The pumps have slightly different output pressures. Because the cells are larger than the smaller opening towards vessel A, they are aspirated into the conical wells by the pressure difference, and obstruct further cells from entering the same well.

[0293] To achieve this, the chamber must satisfy two conditions. First, it must have nearly the same pressure everywhere; in the other words, both vessels must have nearly the20 same pressure at input and output. Mathematically, ^^ ≈ 0. The second condition is that thepressure near the entrance of the wells must be approximately the same as the pressure in the respective vessel.

[0294] To satisfy the design constraints, spacious vessel should be used as buffer volumes, (and the volume is controllable via padding). Additionally, all the pressure drop 25 should occur at the exit of the vessels, therefore both vessels require control mechanisms (e.g. needle valves, pressure regulators, or Mass Flow Controllers (MFCs). The flow chamber may optionally contain inlet or outlet manifolds, or manifolds.

[0295] Consider a well that does not contain a cell. It is possible to apply the flow resist equation for a conical segment under the assumption of incompressible Newtonian Fluid, and 30 laminar flow conditions (Poiseuille’s law):^^^ − ^^^ =128^^^^^^ 1 1 ^^^ −related to inverse of the fourth power of the diameter, so we can assume the flow rate in micrometer scale wells to be small compared to the flow in the vessels. 5

[0297] Also consider a well that is obstructed by a cell. The difference of pressure exerted on the two sides of the cell approximately counterbalances the mass of the cell, minus its buoyancy, plus the force exerted by the walls: (^^^ − ^^^)^^^ ≈ ^^(^^^ − 4⁄ 3 ^^ଷ^^ுଶை) + ^^௪because the cytoplasm have approximately the same density as media, we can approximate 10 the force exerted on the cell by the walls with the difference in vessel pressures times the cross section area of the cell. This provides the estimate of the upper limit of the pressure difference without destroying the cells. Approximate mechanical property of cells are known in the art (see, e.g., R.M. Hochmuth

[0047] ^).

[0298] Microfluidics cell sorting: 15

[0299] The fabrication of the array to be used with the chamber can be accomplished in a number of standard steps, as summarized below and in FIG.12, (illustrating steps A – E as explained below).

[0300] A. Construct a Photoresist master for the conical columns using one of the following methods: (1) Direct-Write Grey scale Lithography (um scale 3D printing); (2) 20 Dithered dot matrix mask with negative photoresist, with the transparency gradually increasing towards edges. (3) A set of circular masks that can be aligned with precision on the level of 1um (e.g. chrome on lemon glass).

[0301] B. Coat PDMS until it’s level with the master: (1) Spin-coat PDMS while monitoring the coating rating; (2) Wet etch TBAF to the top of the photoresist columns 25

[0302] C. Pattern Silver Nanowire conductive pathways: (1) 1. Pattern a negative stainless steel stencil Mask (a.k.a shadow mask) of the interconnects; (2) Apply O2 or air plasma process so that only areas not protected by the stencil becomes hydrophilic; (3) Remove the stencil mask and drop cast AgNW dispersion and agitate, AgNW will onlyadhere to PDMS surface treated with plasma; (4) Back at 200C for 20 min and set aside for 24 hours.

[0303] D. Attaching a bulk PDMS backing: (1) Use Computer numerical control (CNC) machine to create a plastic (or other mold material known in the art) bulk mold with 5 macroscopic windows over areas with the wells; (2) 3D print a circular wafer holder with raised edges that matches the mold for the mold in D1; (3) Place the wafer from completed step A-C in the holder and close with the mold from D1; (4) Cast bulk PDMS backing.

[0304] E. Detach the PDMS device from the bulk mold from D1 and the photoresist master in step A. Dice using a paper slicer (or method for dicing PDMS known in the art). 10

[0305] Pretensioned cuff electrode: The engineering of spiral cuff nerve electrodes known in the art. See, e.g., Naples et al

[0048] (p.907).

[0306] Spiral cuff electrodes can be characterized as snap rings where ^^^^ = 2.25 ∗Where, ^^^^ is the change in Gap Distance, ^^^^ is the change in Internal Pressure, ^^ is the 15 Mean snap Ring Diameter, ℎ is the snap ring thickness and ^^ is Young’s Modulus.

[0307] The method may use spring loaded or magnetic mechanisms to hold the electrodes in tension prior to implantation, such as during culturing.

[0308] Culture Protocols: Clinical applications require iPSC derived cells expression motor and sensory neuron phenotype, because primary cells are not available. However, a 20 rodent study would not require them as primary cells are readily isolated from rats and mice. However, protocols can be followed to test human cells in animals. For such tests, immune suppressants are administered for the entire duration.

[0309] Surgical Protocols: Animal model:

[0310] Spinal cord injury model in rats has been adapted from A. Iwata

[0012] ^. However a 25 full spinal transection model instead lateral hemisection model can be used because natural plasticity makes it difficult to detect functional improvements in hemisection models.

[0311] Spinal, rolled: The primary sensor in this embodiment is a CEA electrode rolled around a hydrogel core seeded with DRG and / or motor neurons. To accomplish this, a self- curling variant of CEA is grown under tension (see electrode construction). Prior to implantation, the device can be coated with hydrogel, and one end can be released, causing it 5 to relax into a roll. For formulation see, e.g., A. Iwata et al.

[0012] ^. Following which, the rolled device should take place of the hydrogel placed in the lesion site following the injury^.

[0312] Periphery, Actuator Cuff: The principle actuator in this embodiment is another pretensioned self-curling CEA electrode implanted as a cuff around the sciatic nerve. The surgical protocol for implanting cuff electrodes in this fashion has been previously 10 established in F. J. Rodríguez et al.

[0049] (p.111 and Fig.3)^ and G. G. Naples et al

[0048] ^. ^. Essentially, when released, the electrode previously under tension should automatically form a spiral cuff around the periphery nerve when released. As explained, Rhizotomy should be performed on the roots associated with the limb to remove the spastic state associated with SCI. The procedure in rat has been established (A. I. Basbaum

[0050] ^ (p.491)). 15

[0313] Subcutaneous Transmitters: The electrodes should be connected to a hermetically sealed casing using insulated ribbons. This casing contains battery, ADC, pulse generator, transmitter and micro-controllers as an integrated control pod. Since rodents have a flexible skeleton, there is freedom to the place the control hubs. In rodents, device pockets can be created via inserting gelatinous sheets into subcutaneous sites, and inflated by injecting a 20 viscous mixture. This method is detailed in R. Kuwabara and H. Iwata

[0051] . After the pocket is created and inflated, the control hub is inserted into the pocket displacing the Hyaluronate solution.

[0314] Periphery, Sensory Cuff: A second optional CEA Cuff Electrode can be used to collect ascending neural signals from the periphery nerve for feedback control. The protocol 25 for implanting this electrode is identical to the cuff electrode actuation. However, this device should be seeded with sensory / spinal neurons or neural precursors. This electrode should be implanted during the same surgery as the actuator cuff electrode, immediately rostral to the former, therefore signal generated by stimulation does not interfere with it. Alternatively, the two can be implemented as a single CEA cuff electrode seeded with both motor and sensory 30 neurons.

[0315] Spinal, Conformal: An optional CEA electrode can wrap around the dorsal spinal cord and function as a backup, or an alternative to the rolled spiral electrode in the spinal lesion. (in situations where, for instance, there is no cavity or there is tissue infiltration into the lesion). For animal tests, this requires a modification of the SCI model. The open-door 5 expansive laminoplasty

[0012] ^ should only be performed on Th10-Th11 spinal vertebrae, leaving the spinal cord under Th9 intact.

[0316] An open-door expansive laminoplasty should also be performed on Th9 during the operation to implant the rolled electrode. A small incision should be made to open the Dura where the wells are positioned, and an array should be attached to the dura and sutured to it. 10

[0317] Electronic Design: An important design considering for implantable systems is to minimize power dissipation into the nervous system because it creates biological stress as the subjects must dissipate the same heat. Overheating of the brain can be lethal. Therefore, it should be best practice to move power dissipating components away from the nervous system and reduce their power consumption. 15

[0318] The CEA electrodes aid greatly in this as it outputs a spike train for each channel instead of rapidly fluctuating field potential. This can be easily transformed into a binary value by a spike detector. Using output of the spike detector to gate the output of a timer will further compress the signal into spike timing. Using the binary signal from the spike detector as clock to a counter will compress the signal into spike counts. 20

[0319] A spike detector can be created on the CEA itself using just 2 components for each channel: A single MOSFET with its Gate connected to the well, with Source connected to a microcontroller (e.g. via DAC), and the drain connected through a Resistor to a DC rail. At near-threshold conditions: ^^^25 Where, ^^^is a device‐dependent current scale (the current at threshold); ^^ is the subthreshold slope factor; ^^௧is a temperature related constant

[0320] For a simplicity of analysis, further assume that: ^^^ = 1^^^^, ^^ = 1, ^^௧ = 25^^^^, ^^ = 1^^^^

[0321] Assuming that the microcontroller is calibrated to generate a bias exactly opposite of the threshold voltage, so that ^^ ௌ − ^^^^ℎ is the membrane potential. This predictsapproximately 0.06V across the resistor during rest (-70mV membrane potential). As well as approximately 3.32V across the resistor during activation (+30mV membrane potential). This 5 can be easily detected by a ADC or op Amp (voltage comparator). The exponential behavior near the threshold is obviously nonlinear therefore this circuit would distort signal that are not spike trains.

[0322] A circuit of this type relies on precise calibrations of bias. Even temperature fluctuations would impact its operations. However, the electrode is expected to be implanted 10 in a temperature regulated mammalian CNS, and operating on spike trains produced by the CEA. It has some major advantages, for example it is extremely compact, and the power dissipation at the transistor is extremely small, with P_rest≈0.36μW and P_spike≈8.9μW (estimated using ^^ = ^^^ ⋅ ^^^ௌ). Note that action potentials only last a few ms, and at theresting potential even 1000 channels dissipates less than 1mW. Since resistor and most of the 15 processing elements are located inside the control hub, power from those components will dissipate away from the biological system. This addresses the long standing power dissipation issue associated with implantable devices.

[0323] Radio-frequency identification (RFID) technology is recommended method of communications between the implantable and wearable electronics, for example 20 ISO^11784 / 11785, aka Animal Tags, and ISO 14443. RFID devices have extremely low power consumption and do not drain the battery. They are known to be able to transmit from inside animals

[0052] . The bit rate for low frequency 11784 / 11785 in FDX (full duplex) is 4194 bits / s and in HDX (half duplex) it is between 7762 and 8387

[0053] ^. Bit rate in this range is sufficient to support hundreds of spike trains. Low frequency RFID have an important 25 advantage in their ability to penetrate conductive material. As an approximation of far field attenuation with good conductors is known in the art (G. Benelli and A. Pozzebon

[0054] ). Estimated penetration depth in sea water is 68mm for 13.56MHz and 710mm for 125kHz. Using similar calculations, we estimate the penetration depth for saline (with electrical properties similar to animal tissue) is about 3 times as large as sea water , at approximately 30 1.94 meters for 125kHz. This suggest there should be no more than 20cm of tissue between implantable hubs and the wearable pods for 13.56MHz RFID system. For 134^kHz RFIDsystems their relative positioning is largely irrelevant for the purpose of communication. However, charging efficiency, not communication, is a primary consideration for placements.

[0324] Of course, these estimates are not always accurate because RFID devices are based on magnetic coupling in the near field. But as a general rule of thumb, low Frequency 5 RFID is better for robust system, with smaller channel and high frequency RFID better for systems with large number of channels.

[0325] AI Architecture (machine learning algorithms): The reinforcement learning algorithm Deep Deterministic Policy Gradients (DDPG) is commonly used in robotics for adaptive control in continuous action spaces. It accomplishes this using two network 10 connected end to end, a actor or policy network, called μ, which learns the mapping from the state s to the optimal action a, and a critic network, called Q, which evaluate the quality of taking the action a from the state s. See, e.g., T. P. Lillicrap et al

[0055] (p.3). Also provided is the performance of the algorithm on a set of typical robotic tasks, including pendulum swing- up. The algorithm has been used in real world robotics tasks, such as robotic grasping (H. 15 Zhang et al.

[0056] ).^ Also note that DDPG is implicitly predictive, because like all reinforcement algorithm, it is trained using entire episodes of data, or alternatively discounted future rewards and Temporal Difference.

[0326] A feature network must decode the volitional command signal from electrodes in the spinal cord. This network is comprised of a series of feed-forward feature layers. Such an 20 algorithm is known in the art (see, e.g.,

[0057] ).^ While similar, the decoders for CEA must be different from decoders for traditional electrode arrays that collect field potentials. The CEA produces a spike train or spike count instead of a rapidly fluctuating field potential per channel.

[0327] Therefore, instead of power of the signals in particular bands, spectrograms or 25 short time Fourier Transforms of the Local Field Potentials, the independent raw features should be spike counts. In addition, time averages in windows of different lengths, moving average with different time constant, and time-delayed spike counts can be used as additional features.

[0328] In principle, DDPG is capable of generating a gradient that propagates back to the 30 decoder. However the layers should be pretrained with an unsupervised algorithm to produce uncorrelated features in lower dimensions as a part of initialization. For instance, byinitializing the feature layers as an auto-encoder

[0058] .^ This is recommended because the online training of a reinforcement system is expensive. Conversely, neural signal will always be available as long as the implant is switched on, even if the subject is performing unrelated tasks. And these background neural activities can be used train the feature layers. 5

[0329] FIG.13 is an overview of the AI architecture. The ANN receives high dimensional neural recording ^^ே(811)from recording electrodes (601 / 721 / 821) in the spinal cord, and Kinematic State, ^^^(814) from motion sensors (640). Feature extraction network ^^ (731831) Converts neural signals ^^ேto an uncorrelated low-dimensional control state ^^^. (813). Policy network ^^ (830) learns to generate the optimal low-dimensional control action 10 ^^ (810) from the states ^^^and ^^^.

[0330] Additionally, if the system has downstream sensory electrodes, a low dimensional decorrelated version of the tactile and proprioceptive signal ^^^^(^^)(output of 724, FIG.10) should be added as an additional input to the policy ^^, and optionally to critics Q for the same reasons connecting ^^^to Q below. 15

[0331] At the same time, the Criterion Function computes reward ^^ (836) from ^^^and measurement of auxiliary sensors. Note that the Criterion is not a neural network. In this case it defines the goodness of the gait and, therefore, does not require neural signals. For example, it could simply be distances walked on a treadmill, or the percentage of body weight supported. 20

[0332] Following the standard DDPG implementation, a critic or Q network (833) learns to predict the reward ^^ from the kinematic state ^^^and control action ^^. Note that the control state ^^^is omitted from the Q network because it encodes the subject’s intention and the gradient to F already flow through ^^. If this is included, it creates a bypass connection from the input of Q to output of F. This reduce the effective depth of the network and may allows F 25 to train faster. This connection is optional.

[0333] A separate actuation network ^^ (732832) converts the control action ^^ back to high-dimensional Neural action ^^ேfor the stimulating electrode array (602722822). In order ensure that gradient flows through ^^, we introduce a second critic network ^^^to predict reward ^^ from the kinematic state ^^^and neural action ^^ே. In addition, a stop gradientoperation is placed before the actuation network ^^ to prevent the gradients from flowing from ^^ back to ^^.

[0334] Optimization could be carried out for product of the output of the two critics ^^ and ^^^. Since the gradient relative to a product is scaled by the other quantity in the product, optimizing the product encourages the smaller of the two terms to increase faster. Since intuitively ^^ and ^^^should predict the same reward ^^, if they differ greatly, the learning should focus on the path which predicts the smaller reward.

[0335] Furthermore, Either Q or ^^^may contain convolution in time and may be recurrent networks, because it may not be possible to learn certain criterion from instantaneous statistics, e.g. penalties for rapid accelerations and deceleration.

[0336] Lastly, to facilitate offline training, we introduce a forward model ^^(^^) to the DDPG network. This model is trained to estimate the Jacobian of the stiffness and viscosity of the system as a vector function of the state. It can be a linear approximation, such that^^^^ = ^^(^^)[^^|^^]′ (The later term [^^|^^]′ is not a conditional probability but rather theconcatenation of state and action vectors.) Alternatively, in the more general case ^^௧ା^=^^(^^௧, ^^௧). This internal model is constantly updated by supervised learning, for example byminimizing the vector distance distance between state predicted from the previous state and the actual action and the new observed states e.g. ^^^^^^ಾ^^^^^^[^^(^^௧ି^, ^^௧ି^) − ^^௧]ଶ. ఏ

[0337] To improve stability, the forward model can take into account multiple previousstates, e.g. ^^(^^௧, ^^௧, ^^௧ − 1, ^^௧ − 1, ... ). It could be convolution in time and denselyconnected between states, like previously provided in Guo et al

[0059] ^. It could also be implemented as a recurrent network e.g. with LSTM. These optimizations are known to those skilled in the art. This network is used solely to increase data efficiency during training, including in the embodiment below.

[0338] Integrative Training: Training on the system can only be accomplished through the collaboration of the subject and the AI. Therefore, both should be motivated towards the same goal. In the other word, the criterion should be some objective performance, and the subject should benefit from improving the criterion. In human patients this can be automatic because it is in their best interest. Rodent should receive rewards such as sugar pellets or juice proportional to their performance.

[0339] Weight supported treadmill training for spinal injury model of rats has been previously established See, e.g., J. Cha et al ^

[0060] (Fig.1A atp.1002)^. Alternatively, simple Treadmill Harnesses, for instance those offered by MazeEngineers can also be used for this purpose. 5

[0340] Phase one (impulse tests) can be skipped in rodents. Haptic system small enough for rats are uncommon and attaching haptic to rats would result in significant opposition from the subject. Actuation network should be initialized randomly instead. Phase two (stability) can proceed as normal via standard reinforcement technique using the measurement of body weight supported as criterion. 10

[0341] Weight supported walking (Phase Three) should alternate between behavioral training and offline optimization. Rats with the AIFES perform one hour treadmill sessions, containing 30 one minute trials and one minute of rest in between. O each trial the treadmill is set to a random speed profile of walk, run and stops.

[0342] During the training sessions ^^ and ^^ are fixed, while the dynamic model ^^ and 15 critics ^^ and ^^^are trained to predict the criterion function in supervised learning, using a fixed (the previous) policy. Moreover, both neural and motion data is logged. After the session, successful trials are selected by criterion score to perform gradient optimization.

[0343] During optimization, ^^ and ^^^are locked. Inversely ^^, ^^and ^^ are unlocked, and updated using the the logged neural and motion data. Optimization is carried out using a 20 Model Based Trust Region Method. During each iteration, the policy being optimized ^^ is rolled out using the recorded neural data ^^ேand the dynamic model ^^ trained on the physical data. This produce a trajectory of actions ^^(^^)and estimated states ^^^̂(^^). The trajectory of neural action ^^ே(^^)is subsequently calculated from actions ^^(^^). During offline optimization, the true reward ^^ is not available. However, Critic network ^^ and ^^^trained on 25 the physical data will produce estimated reward trajectories ^^(^^)as usual. Policy is optimized to maximize these reward over time as per standard reinforcement learning updates.

[0344] The optimization ends if the predicted trajectory of kinematic states under the new policy exits some boundary around the physical trajectory. Define hyper-parameter ^^ such that:30 ∀௧[^^^̂(^^) − ^^^(^^)]ଶ < ^^ଶ

[0345] If the unrolled trajectory exits the trust region at any point, no subsequent optimization can be trusted. This is because Q functions are always defined on some policy. In this case, the old policy used to collect the physical data.

[0346] This is different from TRPO in several ways. In a TRPO:Surrogate loss: = ^^ ^గ(^|^)^^గ^^^ ^constraint: < ^^where A is advantage (typically estimated by Generalized Advantage Estimation). It differs from Q by a constant baseline ^^(^^) defined for each state. It should be apparent that, because DDPG computes a deterministic policy, it does not enumerate actions or produce explicit probabilities for each action ^^(^^|^^). This makes it difficult to compute the surrogate loss or the KL divergence for DDPG.

[0347] The motivation of TRPO is to increase data efficiency perform multiple gradient update on a single unroll. However, because this control problem has a very low dimensionality, unroll using a model is computationally inexpensive. The goal is therefore, increase efficiency for the physical data, which is expensive to collect. This method is also similar to experience replay. The difference is that it imagines the outcome of alternative policies, if these outcomes are similar to the previous policy.

[0348] Training should proceed to the final phase when the rat is largely able to walk on its own. Mathematically the optimization is similar to Phase Three. However, ^^ is smaller and model unroll and optimization occurs after individual trials instead of entire sessions. This can of course be asynchronous, where the animal can perform the next trial while the network analyzes the previous trial. STATEMENTS REGARDING INCORPORATION BY REFERENCE AND VARIATIONS

[0349] All references throughout this application, for example patent documents including issued or granted patents or equivalents; patent application publications; and non- patent literature documents or other source material; are hereby incorporated by reference herein in their entireties, as though individually incorporated by reference, to the extent eachreference is at least partially not inconsistent with the disclosure in this application (for example, a reference that is partially inconsistent is incorporated by reference except for the partially inconsistent portion of the reference).

[0350] The terms and expressions which have been employed herein are used as terms of 5 description and not of limitation, and there is no intention in the use of such terms and expressions of excluding any equivalents of the features shown and described or portions thereof, but it is recognized that various modifications are possible within the scope of the invention claimed. Thus, it should be understood that although the present invention has been specifically disclosed by preferred embodiments, exemplary embodiments and optional 10 features, modification and variation of the concepts herein disclosed may be resorted to by those skilled in the art, and that such modifications and variations are considered to be within the scope of this invention as defined by the appended claims. The specific embodiments provided herein are examples of useful embodiments of the present invention and it will be apparent to one skilled in the art that the present invention may be carried out using a large 15 number of variations of the devices, device components, methods steps set forth in the present description. As will be obvious to one of skill in the art, methods and devices useful for the present methods can include a large number of optional composition and processing elements and steps.

[0351] As used herein and in the appended claims, the singular forms "a", "an", and "the" 20 include plural reference unless the context clearly dictates otherwise. Thus, for example, reference to "a cell" includes a plurality of such cells and equivalents thereof known to those skilled in the art. As well, the terms "a" (or "an"), "one or more" and "at least one" can be used interchangeably herein. It is also to be noted that the terms "comprising", "including", and "having" can be used interchangeably. The expression “of any of claims XX-YY” 25 (wherein XX and YY refer to claim numbers) is intended to provide a multiple dependent claim in the alternative form, and in some embodiments is interchangeable with the expression “as in any one of claims XX-YY.”

[0352] When a group of substituents is disclosed herein, it is understood that all individual members of that group and all subgroups are disclosed separately. When a 30 Markush group or other grouping is used herein, all individual members of the group and all combinations and subcombinations possible of the group are intended to be individually included in the disclosure.

[0353] Every device, system, formulation, combination of components, or method described or exemplified herein can be used to practice the invention, unless otherwise stated.

[0354] Whenever a range is given in the specification, for example, a size range, a density range, a temperature range, a time range, or a composition or concentration range, all 5 intermediate ranges and subranges, as well as all individual values included in the ranges given are intended to be included in the disclosure. It will be understood that any subranges or individual values in a range or subrange that are included in the description herein can be excluded from the claims herein.

[0355] All patents and publications mentioned in the specification are indicative of the 10 levels of skill of those skilled in the art to which the invention pertains. References cited herein are incorporated by reference herein in their entirety to indicate the state of the art as of their publication or filing date and it is intended that this information can be employed herein, if needed, to exclude specific embodiments that are in the prior art. For example, when composition of matter are claimed, it should be understood that compounds known and 15 available in the art prior to Applicant's invention, including compounds for which an enabling disclosure is provided in the references cited herein, are not intended to be included in the composition of matter claims herein.

[0356] As used herein, “comprising” is synonymous with "including," "containing," or "characterized by," and is inclusive or open-ended and does not exclude additional, unrecited 20 elements or method steps. As used herein, "consisting of" excludes any element, step, or ingredient not specified in the claim element. As used herein, "consisting essentially of" does not exclude materials or steps that do not materially affect the basic and novel characteristics of the claim. In each instance herein any of the terms "comprising", "consisting essentially of" and "consisting of" may be replaced with either of the other two terms. The invention 25 illustratively described herein suitably may be practiced in the absence of any element or elements, limitation or limitations which is not specifically disclosed herein.

[0357] One of ordinary skill in the art will appreciate that starting materials, biological materials, reagents, synthetic methods, purification methods, analytical methods, assay methods, and biological methods other than those specifically exemplified can be employed 30 in the practice of the invention without resort to undue experimentation. All art-known functional equivalents, of any such materials and methods are intended to be included in thisinvention. The terms and expressions which have been employed are used as terms of description and not of limitation, and there is no intention that in the use of such terms and expressions of excluding any equivalents of the features shown and described or portions thereof, but it is recognized that various modifications are possible within the scope of the 5 invention claimed. Thus, it should be understood that although the present invention has been specifically disclosed by preferred embodiments and optional features, modification and variation of the concepts herein disclosed may be resorted to by those skilled in the art, and that such modifications and variations are considered to be within the scope of this invention as defined by the appended claims. 10 REFERENCES (Example 1) [1] J. Park, S. Kim, J. Li, and A. Han, “Axon length quantification microfluidic culture platform for growth and regeneration study,” Methods in Molecular Biology, vol.1162. pp. 85–95, 2014. doi:10.1007 / 978-1-4939-0777-9_7. [2] E. Malishev et al., “Microfluidic device for unidirectional axon growth,” J. Phys. Conf. 15 Ser., vol.643, no.1, p.012025, Oct.2015, doi: 10.1088 / 1742-6596 / 643 / 1 / 012025. [3] A. S. Asan, S. G. Id, and M. Sahin, “Electrical fields induced inside the rat brain with skin, skull, and dural placements of the current injection electrode,” pp.1–15, 2019. [4] A. M. Turner and W. T. Greenough, “Differential rearing effects on rat visual cortex synapses. I. Synaptic and neuronal density and synapses per neuron,” Brain Res., vol.329, 20 no.1–2, pp.195–203, 1985, doi: 10.1016 / 0006-8993(85)90525-6. [5] M. Chernov and A. W. Roe, “Infrared neural stimulation: a new stimulation tool for central nervous system applications,” Neurophotonics, vol.1, no.1, p.011011, 2014, doi: 10.1117 / 1.nph.1.1.011011. [6] C. Seok, F. Y. Yamaner, M. Sahin, and O. Oralkan, “A Sub-Millimeter Lateral Resolution 25 Ultrasonic Beamforming System for Brain Stimulation in Behaving Animals,” Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. EMBS, pp.6462–6465, Jul.2019, doi: 10.1109 / EMBC.2019.8857627. [7] H. Nollet, L. Van Ham, P. Deprez, and G. Vanderstraeten, “Transcranial magnetic stimulation: review of the technique, basic principles and applications,” Vet. J., vol.166, no. 30 1, pp.28–42, Jul.2003, doi: 10.1016 / S1090-0233(03)00025-X. [8] L. Huber et al., “Ultra-high resolution blood volume fMRI and BOLD fMRI in humans at 9.4 T:Capabilities and Challenges,” Neuroimage, vol.178, p.769, Sep.2018, doi:10.1016 / J.NEUROIMAGE.2018.06.025. [9] J. H. Duyn, “The future of ultra-high field MRI and fMRI for study of the human brain,” Neuroimage, vol.62, no.2, p.1241, Aug.2012, doi:10.1016 / J.NEUROIMAGE.2011.10.065.

[0010] J. Wu et al., “Kilohertz two-photon fluorescence microscopy imaging of neural activity 5 in vivo,” Nat. Methods, vol.17, no.3, p.287, Mar.2020, doi: 10.1038 / S41592-020-0762-7.

[0011] S. M. Rajguru, A. I. Matic, and C. P. Richter, “Optical Stimulation of Neurons,” Laser Imaging Manip. Cell Biol., pp.99–112, 2010, doi:10.1002 / 9783527632053.ch6.

[0012] E. Musk, “An integrated brain-machine interface platform with thousands of channels”. 10

[0013] K. E. Jones, P. K. Campbell, and R. A. Normann, “A Glass / Silicon Composite Intracortical Electrode Array,” pp.423–437, 1992.

[0014] R. Bhandari, S. Negi, and F. Solzbacher, “Wafer-scale fabrication of penetrating neural microelectrode arrays,” pp.797–807, 2010, doi: 10.1007 / s10544-010-9434-1.

[0015] R. Luo, H. Li, B. Du, S. Zhou, and Y. Zhu, “A simple strategy for high stretchable, 15 flexible and conductive polymer films based on PEDOT:PSS-PDMS blends,” Org. Electron., vol.76, no. April 2019, p.105451, 2020, doi: 10.1016 / j.orgel.2019.105451.

[0016] R. K. Pal, S. Pradhan, L. Narayanan, and V. K. Yadavalli, “Micropatterned conductive polymer biosensors on flexible PDMS film,” vol.259, pp.498–504, Apr.2018, Accessed: Mar.15, 2022.[Online]. Available:http: / / dx.doi.org / 10.1016 / j.snb.2017.12.082 20

[0017] B. H. Cong and T. Pan, “Photopatternable Conductive PDMS Materials for Microfabrication,” pp.1912–1921, 2008, doi: 10.1002 / adfm.200701437.

[0018] B. X. Niu, S. Peng, L. Liu, W. Wen, and P. Sheng, “Characterizing and Patterning of PDMS-Based Conducting Composites **,” pp.2682–2686, 2007, doi:10.1002 / adma.200602515. 25

[0019] C. Lang and D. Boning, “Spin Coating Modeling and Planarization Using Fill Patterns for Advanced Packaging Technologies,” pp.192–197, 2017.

[0020] L. M. Peurrung and D. B. Graves, “Spin Coating Over Topography,” vol.6, no. 9205040, 1993.

[0021] J. Garra, G. Advanced, and J. Currie, “Dry etching of polydimethylsiloxane for 30 microfluidic systems,” pp.975–982, 2002, doi: 10.1116 / 1.1460896.

[0022] M. Zhang and P. Watson, “ScholarlyCommonsReactive Ion Etching Selectivity of Si / SiO2: Comparing of two fluorocarbon gases CHF3 and CF4 Reactive Ion Etching Selectivity of Si / SiO2: Comparing of two fluorocarbon gases,” pp.4–9, 2019.

[0023] S. A. Gupta and R. K. Gupta, “A Parametric Study of Spin Coating over Topography †,” vol.5885, no.97, pp.2223–2227, 1998.

[0024] E. Smela, “Patterning PDMS using a combination of wet and dry etching Patterning PDMS using a combination of wet and dry etching”, doi:10.1088 / 0960-317 / 19 / 4 / 047002. 5

[0025] N. R. A. M. Ejia, “PLURIPOTENT STEM CELLS Concise Review: Generation of Neurons From Somatic Cells of Healthy Individuals and Neurological Patients Through Induced Pluripotency or Direct Conversion,” pp.2811–2817, 2014.

[0026] H. Kanno, “Regenerative therapy for neuronal diseases with transplantation of somatic stem cells,” vol.5, no.4, pp.163–171, 2013, doi:10.4252 / wjsc.v5.i4.163. 10

[0027] D. Bodas and C. Khan-malek, “Formation of more stable hydrophilic surfaces of PDMS by plasma and chemical treatments,” vol.83, pp.1277–1279, 2006, doi: 10.1016 / j.mee.2006.01.195.

[0028] E. H. Chang, S. A. Frattini, S. Robbiati, and P. T. Huerta, “Construction of Microdrive Arrays for Chronic Neural Recordings in Awake Behaving 15 Mice,” JoVE (Journal Vis. Exp., no.77, p. e50470, Jul.2013, doi: 10.3791 / 50470.

[0029] “The cerebral cortex of the rat. - PsycNET.” https: / / psycnet.apa.org / record / 1990-98262- 000 (accessed May 20, 2022).

[0030] S. A. Deadwyler, T. Bunn, and R. E. Hampson, “Hippocampal ensemble activity during spatial delayed-nonmatch-to-sample performance in rats,” J. Neurosci., vol.16, no.1, pp. 20 354–372, Jan.1996, doi:10.1523 / JNEUROSCI.16-01-00354.1996.

[0031] P. Kim, K. W. Kwon, M. C. Park, S. H. Lee, and S. M. Kim, “Soft Lithography for Microfluidics: a Review,” vol.2, no.1, pp.1–11, 2008.

[0032] X. Cheng, C. Bao, X. Wang, F. Zhang, and W. Dong, “Soft surface electrode based on PDMS - CB conductive polymer for electrocardiogram recordings,” Appl. Phys. A, pp.1–7, 25 2019, doi: 10.1007 / s00339-019-3124-5.

[0033] C. Liu, “Microfabrication of conductive polymer nanocomposite for sensor applications,” 2012.

[0034] R. Karimi-Chaleshtori, A. H. Nassajpour-Esfahani, M. R. Saeri, P. Rezai, and A. Doostmohammadi, “Silver nanowire-embedded PDMS with high electrical conductivity: 30 nanowires synthesis, composite processing and electrical analysis,” Mater. Today Chem., vol. 21, p.100496, Aug.2021, doi: 10.1016 / J.MTCHEM.2021.100496.

[0035] W. Chen, R. H. W. Lam, and J. Fu, “Photolithographic surface micromachining of polydimethylsiloxane (PDMS),” Lab Chip, vol.12, no.2, pp.391–395, Dec.2011, doi: 10.1039 / C1LC20721K.

[0036] M. Redecker et al., “Topographical Micropatterning of Poly(dimethylsiloxane) Using Laminar Flows of Liquids in Capillaries**,” Electrochem. Methods & Fundamentals Appl., vol.76, 2000.

[0037] J. H. Kim et al., “Simple and cost-effective method of highly conductive and elastic 5 carbon nanotube / polydimethylsiloxane composite for wearable electronics,” Sci. Reports 201881, vol.8, no.1, pp.1–11, Jan.2018, doi: 10.1038 / s41598-017-18209-w.

[0038] S. Z. Szilasi and L. Juhasz, “Selective etching of PDMS : etching as a negative tone resist,” pp.1–16.

[0039] B. Gale, B. Gale, M. A. Eddings, M. A. Johnson, and B. K. Gale, “Determining the 10 optimal PDMS – PDMS bonding technique for microfluidic devices Determining the optimal PDMS – PDMS bonding technique for microfluidic devices”, doi: 10.1088 / 0960- 1317 / 18 / 6 / 067001.

[0040] C. Chen and K. Wharton, “RSC Advances Characterization and failure mode analyses of air plasma oxidized PDMS – PDMS bonding by peel testing †,” pp.1286–1289, 2017, doi: 15 10.1039 / c6ra25947b.

[0041] D. C. Duffy, J. C. Mcdonald, O. J. A. Schueller, and G. M. Whitesides, “Rapid Prototyping of Microfluidic Systems in Poly ( dimethylsiloxane ),” vol.70, no.23, pp.4974– 4984, 1998.

[0042] P. R. Ray et al., “RNA profiling of human dorsal root ganglia reveals sex differences in 20 mechanisms promoting neuropathic pain,” Brain, vol.146, no.2, pp.749–766, Feb.2023, doi: 10.1093 / BRAIN / AWAC266.

[0043] A. Björklund et al., “Neural transplantation for the treatment of Parkinson’s disease,” Lancet Neurol., vol.2, no.7, pp.437–445, Jul.2003, doi: 10.1016 / S1474-4422(03)00442-3.

[0044] D. G. Southwell, R. C. Froemke, A. Alvarez-Buylla, M. P. Stryker, and S. P. Gandhi, 25 “Cortical Plasticity Induced by Inhibitory Neuron Transplantation,” Science, vol.327, no. 5969, p.1145, Feb.2010, doi: 10.1126 / SCIENCE.1183962.

[0045] X. Yin, J. C. Xu, G. S. Cho, C. Kwon, T. M. Dawson, and V. L. Dawson, “Neurons Derived from Human Induced Pluripotent Stem Cells Integrate into Rat Brain Circuits and Maintain Both Excitatory and Inhibitory Synaptic Activities,” eNeuro, vol.6, no.4, Jul. 30 2019, doi: 10.1523 / ENEURO.0148-19.2019.

[0046] K. Takahashi et al., “Induction of Pluripotent Stem Cells from Adult Human Fibroblasts by Defined Factors,” Cell, vol.131, no.5, pp.861–872, Nov.2007, doi: 10.1016 / j.cell.2007.11.019.

[0047] N. Malik and M. S. Rao, “A Review of the Methods for Human iPSC Derivation”, doi: 10.1007 / 978-1-62703-348-0_3.

[0048] M. Frega et al., “Rapid Neuronal Differentiation of Induced Pluripotent Stem Cells for Measuring Network Activity on Micro-electrode Arrays,” JoVE (Journal Vis. Exp., vol. 5 2017, no.119, p. e54900, Jan.2017, doi: 10.3791 / 54900.

[0049] J. R. Loverde, V. C. Ozoka, R. Aquino, L. Lin, and B. J. Pfister, “Live Imaging of Axon Stretch Growth in Embryonic and Adult Neurons,” https: / / home.liebertpub.com / neu, vol.28, no.11, pp.2389–2403, Nov.2011, doi: 10.1089 / NEU.2010.1598.

[0050] G. C. Magou, B. J. Pfister, and J. R. Berlin, “Effect of acute stretch injury on action 10 potential and network activity of rat neocortical neurons in culture,” Brain Res., vol.1624, pp.525–535, Oct.2015, doi: 10.1016 / J.BRAINRES.2015.07.056.

[0051] S. Bhattacharya, A. Datta, J. M. Berg, and S. Gangopadhyay, “Studies on surface wettability of poly(dimethyl) siloxane (PDMS) and glass under oxygen-plasma treatment and correlation with bond strength,” J. Microelectromechanical Syst., vol.14, no.3, pp.590–597, 15 Jun.2005, doi: 10.1109 / JMEMS.2005.844746.

[0052] S. Gok and M. Sahin, “Rat forelimb movement components segregated by corticospinal tract activity,” in International IEEE / EMBS Conference on Neural Engineering, NER, 2017, pp.312–315. doi: 10.1109 / NER.2017.8008353.

[0053] C.R. Vistas et al. “Silanization of glass chips—A factorial approach for optimization.”20 Applied Surface Science 286: 314-315 (Dec.2013). C. Moraes et al. “Solving the shrinkage- induced PDMS alignment registration issue in multilayer soft lithography.” J. Micromech. Microeng.19 (2009) (6 pp.). REFERENCES (Examples 3 and 4) [1] B. B. Lee, R. A. Cripps, M. Fitzharris, and P. C. Wing, “The global map for traumatic 25 spinal cord injury epidemiology: Update 2011, global incidence rate,” Spinal Cord, vol.52, no.2, pp.110–116, Feb.2014, doi: 10.1038 / sc.2012.158. [2] J. Brown et al., “Optogenetic stimulation of a cortical biohybrid implant guides goal directed behavior,” Nov.2024, doi: 10.1101 / 2024.11.22.624907. [3] G. E. Loeb, R. A. Peck, W. H. Moore, and K. Hood, “BIONTMsystem for distributed 30 neural prosthetic interfaces,” Med. Eng. Phys., vol.23, no.1, pp.9–18, Jan.2001, doi: 10.1016 / S1350-4533(01)00011-X. [4] A. G. Feldman and M. F. Levin, “The equilibrium-point hypothesis - Past, present and future,” Adv. Exp. Med. Biol., vol.629, pp.699–726, 2009, doi: 10.1007 / 978-0-387- 77064-2_38.[5] A. G. Feldman, S. V. Adamovich, and M. F. Levin, “The relationship between control, kinematic and electromyographic variables in fast single-joint movements in humans,” Exp. Brain Res., vol.103, no.3, pp.440–450, Nov.1995, doi: 10.1007 / BF00241503 / METRICS. 5 [6] A. G. Feldman and M. L. Latash, “Testing hypotheses and the advancement of science: Recent attempts to falsify the equilibrium point hypothesis,” Exp. Brain Res., vol.161, no.1, pp.91–103, Feb.2005, doi: 10.1007 / S00221-004-2049-0 / METRICS. [7] F. A. Mussa-Ivaldi, N. Hogan, and E. Bizzi, “Neural, mechanical, and geometric factors subserving arm posture in humans,” J. Neurosci., vol.5, no.10, pp.2732– 10 2743, Oct.1985, doi: 10.1523 / JNEUROSCI.05-10-02732.1985. [8] A. G. Feldman, “Once More on the Equilibrium-Point Hypothesis (λ Model) for Motor Control,” J. Mot. Behav., vol.18, no.1, pp.17–54, 1986, doi: 10.1080 / 00222895.1986.10735369. [9] M. Kawato, “Internal models for motor control and trajectory planning,” Curr. Opin. 15 Neurobiol., vol.9, no.6, pp.718–727, Dec.1999, doi: 10.1016 / S0959-4388(99)00028- 8.

[0010] D. B. Popović, “Advances in functional electrical stimulation (FES),” Dec.01, 2014, Elsevier. doi: 10.1016 / j.jelekin.2014.09.008.

[0011] B. J. Pfister, J. H. Huang, N. Kameswaran, E. L. Zager, and D. H. Smith, “Neural 20 engineering to produce in vitro nerve constructs and neurointerface,” Neurosurgery, vol.60, no.1, pp.137–142, Jan.2007, doi: 10.1227 / 01.NEU.0000249197.61280.1D.

[0012] A. Iwata, K. D. Browne, B. J. Pfister, J. A. Gruner, and D. H. Smith, “Long-term survival and outgrowth of mechanically engineered nervous tissue constructs implanted into spinal cord lesions,” Tissue Eng., vol.12, no.1, pp.101–110, 2006, 25 doi: 10.1089 / ten.2006.12.101.

[0013] J. H. Huang et al., “Long-Term Survival and Integration of Transplanted Engineered Nervous Tissue Constructs Promotes Peripheral Nerve Regeneration,” https: / / home.liebertpub.com / tea, vol.15, no.7, pp.1677–1685, Feb.2009, doi: 10.1089 / TEN.TEA.2008.0294. 30

[0014] Y. Li and G. Raisman, “Long axon growth from embryonic neurons transplanted into myelinated tracts of the adult rat spinal cord,” Brain Res., vol.629, no.1, pp.115–127, Nov.1993, doi: 10.1016 / 0006-8993(93)90489-A.

[0015] I. Velasco et al., “Concise review: Generation of neurons from somatic cells of healthy individuals and neurological patients through induced pluripotency or direct 35 conversion,” Stem Cells, vol.32, no.11, pp.2811–2817, 2014, doi: 10.1002 / stem.1782.

[0016] K. Takahashi and S. Yamanaka, “Induction of Pluripotent Stem Cells from Mouse Embryonic and Adult Fibroblast Cultures by Defined Factors,” Cell, vol.126, no.4, pp.663–676, Aug.2006, doi: 10.1016 / J.CELL.2006.07.024 / ATTACHMENT / A7BA2E0F-99EF-4A86-88AA- 5 418202149347 / MMC1.PDF.

[0017] I. Muñoz-Sanjuán and A. H. Brivanlou, “Neural induction, the default model and embryonic stem cells,” Nat. Rev. Neurosci., vol.3, no.4, pp.271–280, 2002, doi: 10.1038 / nrn786.

[0018] A. A. Galiakberova and E. B. Dashinimaev, “Neural Stem Cells and Methods for Their 10 Generation From Induced Pluripotent Stem Cells in vitro,” Front. Cell Dev. Biol., vol. 8, p.559464, Oct.2020, doi: 10.3389 / FCELL.2020.00815 / PDF.

[0019] S. Bell et al., “Differentiation of Human Induced Pluripotent Stem Cells (iPSCs) into an Effective Model of Forebrain Neural Progenitor Cells and Mature Neurons,” Bio- protocol, vol.9, no.5, p. e3188, Mar.2019, doi: 10.21769 / BIOPROTOC.3188. 15

[0020] L. D’Aiuto et al., “Large-scale generation of human ipsc-derived neural stem cells / early neural progenitor cells and their neuronal differentiation,” Organogenesis, vol.10, no.4, pp.365–377, 2014, doi: 10.1080 / 15476278.2015.1011921.

[0021] M. H. Kim, N. Thanuthanakhun, and M. Kino-oka, “Stable and efficient generation of functional iPSC-derived neural progenitor cell rosettes through regulation of collective 20 cell-cell behavior,” Front. Bioeng. Biotechnol., vol.11, p.1269108, Jan.2023, doi: 10.3389 / FBIOE.2023.1269108 / BIBTEX.

[0022] B. Y. Hu et al., “Neural differentiation of human induced pluripotent stem cells follows developmental principles but with variable potency,” Proc. Natl. Acad. Sci. U. S. A., vol.107, no.9, pp.4335–4340, Mar.2010, doi: 25 10.1073 / PNAS.0910012107 / SUPPL_FILE / PNAS.200910012SI.PDF.

[0023] T. Deng et al., “Scalable Generation of Pseudo-Unipolar Sensory Neurons from Human Pluripotent Stem Cells,” bioRxiv, p.2022.03.24.485622, Mar.2022, doi: 10.1101 / 2022.03.24.485622.

[0024] Y. Chen et al., “A versatile polypharmacology platform promotes cytoprotection and 30 viability of human pluripotent and differentiated cells,” Nat. Methods 2021185, vol. 18, no.5, pp.528–541, May 2021, doi: 10.1038 / s41592-021-01126-2.

[0025] C. Compagnucci, M. Nizzardo, S. Corti, G. Zanni, and E. Bertini, “In vitro neurogenesis: development and functional implications of iPSC technology,” Cell. Mol. Life Sci. C., vol.71, no.9, p.1623, 2013, doi: 10.1007 / S00018-013-1511-1. 35

[0026] A. M. Maroof et al., “Directed differentiation and functional maturation of cortical interneurons from human embryonic stem cells,” Cell Stem Cell, vol.12, no.5, pp. 559–572, May 2013, doi: 10.1016 / J.STEM.2013.04.008 / ATTACHMENT / 453177EA- B01A-4494-8F7D-824C7DAAE70D / MMC1.PDF.

[0027] E. Y. Son et al., “Conversion of mouse and human fibroblasts into functional spinal motor neurons,” Cell Stem Cell, vol.9, no.3, pp.205–218, Sep.2011, doi: 10.1016 / j.stem.2011.07.014.

[0028] B. Q. Lai et al., “Tissue‐Engineered Neural Network Graft Relays Excitatory Signal in 5 the Completely Transected Canine Spinal Cord,” Adv. Sci., vol.6, no.22, p.1901240, Nov.2019, doi: 10.1002 / ADVS.201901240.

[0029] P. Gad et al., “Development of a multi-electrode array for spinal cord epidural stimulation to facilitate stepping and standing after a complete spinal cord injury in adult rats,” J. Neuroeng. Rehabil., vol.10, no.1, pp.1–18, Dec.2013, doi: 10 10.1186 / 1743-0003-10-2 / TABLES / 1.

[0030] D. McCreery, V. Pikov, A. Lossinsky, L. Bullara, and W. Agnew, “Arrays for chronic functional microstimulation of the lumbosacral spinal cord,” IEEE Trans. Neural Syst. Rehabil. Eng., vol.12, no.2, pp.195–207, Jun.2004, doi: 10.1109 / TNSRE.2004.827223. 15

[0031] J. D. Yeager, D. J. Phillips, D. M. Rector, and D. F. Bahr, “Characterization of flexible ECoG electrode arrays for chronic recording in awake rats,” J. Neurosci. Methods, vol. 173, no.2, pp.279–285, Aug.2008, doi: 10.1016 / j.jneumeth.2008.06.024.

[0032] U. J. Jeong et al., “A minimally invasive flexible electrode array for simultaneous recording of ECoG signals from multiple brain regions,” Lab Chip, vol.21, no.12, pp. 20 2383–2397, Jun.2021, doi: 10.1039 / D1LC00117E.

[0033] D. Wesp et al., “Analysis of PMMA versus CaP titanium-enhanced implants for cranioplasty after decompressive craniectomy: a retrospective observational cohort study,” Neurosurg. Rev., vol.45, no.6, p.3647, Dec.2022, doi: 10.1007 / S10143-022- 01874-5. 25

[0034] K. S. Katiyar et al., “Tissue Engineered Axon Tracts Serve as Living Scaffolds to Accelerate Axonal Regeneration and Functional Recovery Following Peripheral Nerve Injury in Rats,” Front. Bioeng. Biotechnol., vol.8, p.532857, May 2020, doi: 10.3389 / FBIOE.2020.00492 / BIBTEX.

[0035] K. Aquilina, D. Graham, and N. Wimalasundera, “Selective dorsal rhizotomy: An old 30 treatment re-emerging,” Aug.01, 2015, BMJ Publishing Group Ltd. doi: 10.1136 / archdischild-2014-306874.

[0036] T. S. Park and J. M. Johnston, “Surgical techniques of selective dorsal rhizotomy for spastic cerebral palsy: Technical note,” Neurosurg. Focus, vol.21, no.2, p. E7, Aug. 2006, doi: 10.3171 / FOC.2006.21.2.8. 35

[0037] T. R. Makin and H. Flor, “Brain (re)organisation following amputation: Implications for phantom limb pain,” Neuroimage, vol.218, p.116943, Sep.2020, doi: 10.1016 / J.NEUROIMAGE.2020.116943.

[0038] E. C. Lechelt, “Sensory substitution systems,” Percept. Mot. Skills, vol.62, pp.356–8, 1986, Accessed: Mar.26, 2025. [Online]. Available: http: / / www.ncbi.nlm.nih.gov / pubmed / 17945986

[0039] P. Bach-y-Rita and S. W. Kercel, “Sensory substitution and the human-machine 5 interface,” Trends Cogn. Sci., vol.7, no.12, pp.541–546, 2003, doi: 10.1016 / j.tics.2003.10.013.

[0040] P. Ferrari, F. Giofrè, and P. De Filippo, “Intermuscular pocket for subcutaneous implantable cardioverter defibrillator: Single-center experience,” J. Arrhythmia, vol. 32, no.3, pp.223–226, Jun.2016, doi: 10.1016 / J.JOA.2016.01.005. 10

[0041] T. Liu, Y. Inoue, and K. Shibata, “Development of a wearable sensor system for quantitative gait analysis,” Measurement, vol.42, no.7, pp.978–988, Aug.2009, doi: 10.1016 / J.MEASUREMENT.2009.02.002.

[0042] E. Rocon, J. C. Moreno, A. F. Ruiz, F. Brunetti, J. A. Miranda, and J. L. Pons, “Application of inertial sensors in rehabilitation robotics,” 2007 IEEE 10th Int. Conf. 15 Rehabil. Robot. ICORR’07, pp.145–150, 2007, doi: 10.1109 / ICORR.2007.4428420.

[0043] R. Kobetic et al., “Development of hybrid orthosis for standing, walking, and stair climbing after spinal cord injury,” J. Rehabil. Res. Dev., vol.46, no.3, pp.447–462, 2009, doi: 10.1682 / JRRD.2008.07.0087.

[0044] T. Gordon and J. Mao, “Muscle Atrophy and Procedures for Training After Spinal 20 Cord Injury,” Phys. Ther., vol.74, no.1, pp.50–60, Jan.1994, doi: 10.1093 / PTJ / 74.1.50.

[0045] R. J. Willia, “Simple Statistical Gradient-Following Algorithms for Connectionist Reinforcement Learning,” Mach. Learn., vol.8, no.3, pp.229–256, 1992, doi: 10.1023 / A:1022672621406. 25

[0046] M. Harib, H. Chaoui, and S. Miah, “Evolution of adaptive learning for nonlinear dynamic systems: a systematic survey,” Intell. Robot., vol.2, no.1, pp.37–71, 2022, doi: 10.20517 / ir.2021.19.

[0047] R. M. Hochmuth, “Micropipette aspiration of living cells,” J. Biomech., vol.33, no.1, pp.15–22, 2000, doi: 10.1016 / S0021-9290(99)00175-X. 30

[0048] G. G. Naples, J. T. Mortimer, A. Scheiner, and J. D. Sweeney, “A Spiral Nerve Cuff Electrode for Peripheral Nerve Stimulation,” IEEE Trans. Biomed. Eng., vol.35, no. 11, pp.905–916, 1988, doi: 10.1109 / 10.8670.

[0049] F. J. Rodríguez et al., “Polyimide cuff electrodes for peripheral nerve stimulation,” J. Neurosci. Methods, vol.98, no.2, pp.105–118, Jun.2000, doi: 10.1016 / S0165- 35 0270(00)00192-8.

[0050] A. I. Basbaum, “Effects of central lesions on disorders produced by multiple dorsal rhizotomy in rats,” Exp. Neurol., vol.42, no.3, pp.490–501, Mar.1974, doi: 10.1016 / 0014-4886(74)90073-9.

[0051] R. Kuwabara and H. Iwata, “Bioabsorbable device to prepare subcutaneous pockets for 5 islet transplantation,” J. Biomed. Mater. Res. - Part B Appl. Biomater., vol.107, no.4, pp.1107–1112, May 2019, doi: 10.1002 / jbm.b.34203.

[0052] R. E. Floyd, “RFID in Animal-Tracking Applications,” IEEE Potentials, vol.34, no.5, pp.32–33, Sep.2015, doi: 10.1109 / MPOT.2015.2410308.

[0053] “ISO 11785:1996(en), Radio frequency identification of animals — Technical 10 concept.” Accessed: Mar.27, 2025. [Online]. Available: https: / / www.iso.org / obp / ui / #iso:std:iso:11785:ed-1:v1:en

[0054] G. Benelli and A. Pozzebon, “RFID Under Water: Technical Issues and Applications,” 2013, doi: 10.5772 / 53934.

[0055] T. P. Lillicrap et al., “Continuous control with deep reinforcement learning,” 4th Int. 15 Conf. Learn. Represent. ICLR 2016 - Conf. Track Proc., Sep.2015, Accessed: Mar. 17, 2025. [Online]. Available: https: / / arxiv.org / abs / 1509.02971v6

[0056] H. Zhang, F. Wang, J. Wang, and B. Cui, “Robot grasping method optimization using improved deep deterministic policy gradient algorithm of deep reinforcement learning,” Rev. Sci. Instrum., vol.92, no.2, Feb.2021, doi: 10.1063 / 5.0034101. 20

[0057] Y. Guo, S. Gok, and M. Sahin, “Convolutional networks outperform linear decoders in predicting EMG from spinal cord signals,” Front. Neurosci., vol.12, no. OCT, 2018, doi: 10.3389 / fnins.2018.00689.

[0058] Y. Bengio, P. Lamblin, D. Popovici, and H. Larochelle, “Greedy Layer-Wise Training of Deep Networks,” NIPS 2006 Proc.19th Int. Conf. Neural Inf. Process. Syst., no.1, 25 pp.153–160, 2006, doi: 10.7551 / mitpress / 7503.003.0024.

[0059] Y. Guo, R. A. Foulds, S. V. Adamovich, and M. Sahin, “Encoding of forelimb forces by corticospinal tract activity in the rat,” Front. Neurosci., vol.8, no.8 MAY, pp.1– 11, 2014, doi: 10.3389 / fnins.2014.00062.

[0060] J. Cha, C. Heng, D. J. Reinkensmeyer, R. R. Roy, V. R. Edgerton, and R. D. De Leon, 30 “Locomotor ability in spinal rats is dependent on the amount of activity imposed on the hindlimbs during treadmill training,” J. Neurotrauma, vol.24, no.6, pp.1000– 1012, Jun.2007, doi: 10.1089 / neu.2006.0233.REFERENCES (other than Examples 1, 3 and 4) [1] B. B. Lee, R. A. Cripps, M. Fitzharris, and P. C. Wing, “The global map for traumatic spinal cord injury epidemiology: update 2011, global incidence rate,” Spinal Cord 2013522, vol.52, no.2, pp.110–116, Feb.2013, doi: 10.1038 / sc.2012.158. 5 [2] G. E. Loeb, R. A. Peck, W. H. Moore, K. Hood, and A. E. Mann, “BIONTMsystem for distributed neural prosthetic interfaces,” Med. Eng. Phys., vol.23, pp.9–18, 2001, Accessed: Apr.16, 2024. [Online]. Available: www.elsevier.com / locate / medengphy [3] D. B. Popović, “Advances in functional electrical stimulation (FES),” Journal of Electromyography and Kinesiology, vol.24, no.6. Elsevier, pp.795–802, Dec.01, 10 2014. doi: 10.1016 / j.jelekin.2014.09.008. [4] A. G. Feldman and M. F. Levin, “The equilibrium-point hypothesis - Past, present and future,” Adv. Exp. Med. Biol., vol.629, pp.699–726, 2009, doi: 10.1007 / 978-0-387- 77064-2_38 / COVER. [5] A. Feldman, S. Adamovich, and M. Levin, “The relationship between control, 15 kinematic and electromyographic variables in fast single-joint movements in humans,” Exp. Brain Res., vol.103, no.3, Nov.1995, doi: 10.1007 / BF00241503. [6] A. G. Feldman and M. L. Latash, “Testing hypotheses and the advancement of science: recent attempts to falsify the equilibrium point hypothesis,” Exp Brain Res, vol.161, no.1, pp.91–103, 2005, doi: 10.1007 / s00221-004-2049-0. 20 [7] F. Mussa-Ivaldi, N. Hogan, and E. Bizzi, “Neural, mechanical, and geometric factors subserving arm posture in humans,” J. Neurosci., vol.5, no.10, pp.2732–2743, Oct. 1985, Accessed: Feb.25, 2015. [Online]. Available: http: / / www.jneurosci.org / content / 5 / 10 / 2732.short [8] A. G. Feldman, “Once More on the Equilibrium-Point Hypothesis (λ Model) for Motor 25 Control,” J. Mot. Behav., vol.18, no.1, pp.17–54, Mar.1986, doi: 10.1080 / 00222895.1986.10735369. [9] M. Kawato, “Internal models for motor control and trajectory planning,” Curr Opin Neurobiol, vol.9, no.6, pp.718–727, 1999, [Online]. Available: http: / / www.ncbi.nlm.nih.gov / pubmed / 10607637

[0010] B. J. Pfister, J. H. Huang, N. Kameswaran, E. L. Zager, and D. H. Smith, “Neural engineering to produce in vitro nerve constructs and neurointerface,” Neurosurgery, vol.60, no.1, pp.137–142, Jan.2007, doi: 10.1227 / 01.NEU.0000249197.61280.1D.

[0011] A. Iwata, K. D. Browne, B. J. Pfister, J. A. Gruner, and D. H. Smith, “Long-Term 5 Survival and Outgrowth of Mechanically Engineered Nervous Tissue Constructs Implanted Into Spinal Cord Lesions,” https: / / home.liebertpub.com / ten, vol.12, no.1, pp.101–110, Feb.2006, doi: 10.1089 / TEN.2006.12.101.

[0012] J. H. Huang et al., “Long-Term Survival and Integration of Transplanted Engineered Nervous Tissue Constructs Promotes Peripheral Nerve Regeneration,” 10 https: / / home.liebertpub.com / tea, vol.15, no.7, pp.1677–1685, Feb.2009, doi: 10.1089 / TEN.TEA.2008.0294.

[0013] Y. Li and G. Raisman, “Long axon growth from embryonic neurons transplanted into myelinated tracts of the adult rat spinal cord,” Brain Res., vol.629, no.1, pp.115–127, Nov.1993, doi: 10.1016 / 0006-8993(93)90489-A. 15

[0014] A. Amirali, L. Mu, J. M. Gracies, and D. M. Simpson, “Anatomical localization of motor endplate bands in the human biceps brachii,” J. Clin. Neuromuscul. Dis., vol.9, no.2, pp.306–312, Dec.2007, doi: 10.1097 / CND.0B013E31815C13A7.

[0015] R. J. Triolo and J. Uhlir, “Selectivity of intramuscular stimulating electrodes in the lower limbs,” Artic. J. Rehabil. Res. Dev., 2001, Accessed: Apr.07, 2024. [Online]. 20 Available: https: / / www.researchgate.net / publication / 11623103

[0016] T. Deng et al., “Scalable Generation of Pseudo-Unipolar Sensory Neurons from Human Pluripotent Stem Cells,” bioRxiv, p.2022.03.24.485622, Mar.2022, doi: 10.1101 / 2022.03.24.485622.

[0017] R. Kobetic et al., “Development of hybrid orthosis for standing, walking, and stair 25 climbing after spinal cord injury,” J. Rehabil. Res. Dev., vol.46, no.3, pp.447–462, 2009, doi: 10.1682 / JRRD.2008.07.0087.

[0018] Y. Guo, S. Gok, and M. Sahin, “Convolutional Networks Outperform Linear Decoders in Predicting EMG From Spinal Cord Signals,” vol.12, no. October, 2018, doi: 10.3389 / fnins.2018.00689.

[0019] N. Meuleau, L. Peshkin, K.-E. Kim, and L. P. Kaelbling, “Learning Finite-State Controllers for Partially Observable Environments,” Jan.2013, Accessed: Apr.15, 2024. [Online]. Available: https: / / arxiv.org / abs / 1301.6721v1

[0020] M. Harib, H. Chaoui, S. M.-I. Robot, and undefined 2022, “Evolution of adaptive 5 learning for nonlinear dynamic systems: A systematic survey,” f.oaes.cc, vol.2, no.1, p.37, 2022, doi: 10.20517 / ir.2021.19.

[0021] N. Weidner, A. Ner, N. Salimi, and M. H. Tuszynski, “Spontaneous corticospinal axonal plasticity and functional recovery after adult central nervous system injury,” Proc. Natl. Acad. Sci., vol.98, no.6, pp.3513–3518, Mar.2001, doi: 10 10.1073 / PNAS.051626798.

[0022] D. N. Loy et al., “Functional Redundancy of Ventral Spinal Locomotor Pathways,” J. Neurosci., vol.22, no.1, pp.315–323, Jan.2002, doi: 10.1523 / JNEUROSCI.22-01- 00315.2002.

[0023] S.-W. You et al., “Spontaneous recovery of locomotion induced by remaining fibers 15 after spinal cord transection in adult rats,” Restor. Neurol. Neurosci., vol.21, no.1–2, pp.39–45, Jan.2003.

[0024] D. M. Serlin and M. H. Schieber, “Morphologic Regions of the Multitendoned Extrinsic Finger Muscles in the Monkey Forearm,” Acta Anat. (Basel)., vol.146, no.4, pp.255–266, Apr.1993, doi: 10.1159 / 000147465. 20

[0025] F. J. Rodríguez et al., “Polyimide cuff electrodes for peripheral nerve stimulation,” J. Neurosci. Methods, vol.98, no.2, pp.105–118, Jun.2000, doi: 10.1016 / S0165- 0270(00)00192-8.

Claims

CLAIMS I claim:

1. An implantable functional electrical stimulator (FES) comprising: an electrode configured for implantation and to functionally integrate with a target 5 biological tissue; and a feedback control system comprising a processor in electronic communication with the electrode, wherein the processer executes processor-executable instructions provided by a machine learning algorithm, to cause the electrode to interface with the target biological tissue.

2. The implantable FES of claim 1, wherein the processer implements an interface scheme with the target biological tissue for one or more of: adaptive control; feedback control; and / or predictive control.

3. The implantable FES of claim 2, wherein the feedback control comprises one or more motion sensors connected to an external system of the FES for tracking a location of a limb or limb extremity.

4. The implantable FES of claim 2, wherein the adaptive control is provided by an adap- tive control algorithm implemented via artificial intelligence.

5. The implantable FES of any one of claims 1-4, wherein the processor is provided in a computing device that further comprises a memory device communicatively coupled to the processor.

6. The implantable FES of claim 5, wherein the memory device includes a non-transi- tory computer readable medium storing processor-executable instructions encoded as software, which, when executed by the processor, cause the processor to implement an interface scheme with the target tissue.

7. The implantable FES of claim 6, wherein the interface scheme is AI-generated to pro- vide predictive control of the target biological tissue.

8. The implantable FES of any one of claims 1-7, wherein the processor and electrode are configured to: monitor an electrical parameter of the target tissue; monitor a target tissue motion; electrically actuate the target tissue; ormonitor the electrical parameter of the target tissue and electrically actuate the target tissue.

9. The implantable FES of any one of claims 1-8, wherein the electrode is a biohybrid electrode comprising a grafted or seeded neuronal cell in electrical contact with a 5 synthetic portion of the electrode.

10. The implantable FES of any one of claims 1-9, further comprising: a network of electrodes electrically connected to hermetically sealed implantable electronic components; and an external processing element containing at least the processor.

11. The implantable FES of any one of claims 1-10, further comprising a pod configured to be wearable by a user, wherein the pod comprises a motion sensor for position tracking.

12. The implantable FES of any one of claims 1-11, wherein the processor is contained in a handheld electronic device, and is in wireless communication with the electrode.

13. The implantable FES of any one of claims 1-12, further comprising an assistive exo- skeleton in electrical communication with the processor, wherein the assistive exo- skeleton is configured to provide physical support of a biological system and a meas- ure of a kinetic parameter.

14. The implantable FES of any one of claims 1-13, wherein the electrode comprises: a spinal cord electrode configured to be positioned upstream of a spinal cord lesion and to implantably connect to living cells; a periphery nerve electrode configured to be positioned downstream of the spinal cord lesion and to implantably connect to living cells; and the processor is configured to process, store and / or transmit electrical spike trains.

15. The implantable FES of any one of claims 1-14, further comprising a muscle elec- trode configured to electrically stimulate a muscle based on a signal from the proces- sor and / or a cortical electrode configured to electrically stimulate a cortex layer for tactile sensation.

16. The implantable FES of any one of claims 1-15, further comprising: an external sys- tem, wherein the external system comprises: a communication pod configured to spatially align with a hub implanted under the skin, wherein the hub is in electronic communication with the electrode(s), a computing core communicatively coupled to the communication pod,a plurality of motion sensors; and a software system operably connected to the computing core, the software system comprising: an adaptation onboard neural network for a subject / encoding adaptation; and 5 a simulation onboard neural network for simulation of a local spinal circuit.

17. The implantable FES of any one of claims 1-16, further comprising: a hub configured for: implantation beneath skin, a hub in electrical contact with the electrode(s); inductive power charging of an inductively charged battery; and inductive or radio communication with an externally-positioned processor or a communication pod.

18. The implantable FES of any one of claims 1-17, wherein the electrode comprises a cellular electrode array (CEA).

19. The implantable FES of any one of claims 1-18, wherein at least one electrode is a conformable electrode for conformal contact with the target tissue having a shaped surface for positioning of biological cells to follow a peripheral nerve to the target tis- sue comprising a destination muscle.

20. The implantable FES of claim 19, provided in a cuff configuration for implantation around a major peripheral nerve, including at a lesion to bridge a nerve gap.

21. The implantable FES of any one of claims 1-20, wherein the electrode is a spinal cord electrode configured for implantation onto a spinal cord of the patient.

22. The implantable FES of claim 21, wherein the spinal cord electrode is seeded with: (i) cells that are configured to attract axon collateral from upper motor neurons and / or project dendrite directly onto the upper mountain neurons; (ii) sensory neuronal cells; (iii) neural precursor cells; and / or (iv) embryonic neural cells, including provided in an array pattern.

23. The implantable FES of any one of claims 1-22, wherein the periphery nerve elec- trode is a bi-directional electrode.

24. The implantable FES of any one of claims 1-23, wherein the periphery nerve elec- trode is a cuff electrode.

25. The implantable FES of any one of claims 1-24, wherein the periphery nerve elec- trode is seeded with: (i) cells that are configured to attract axon collateral from lowermotor neurons and / or project dendrite directly onto the lower motor neurons; and / or (ii) sensory neuronal cells.

26. The implantable FES of any claims 1-25, wherein the periphery nerve electrode is seeded with sensory and motor neuronal cells. 5 27. The implantable FES of any one of claims 1-26, wherein the at least one muscle elec- trode comprises a uni-directional electrode.

28. The implantable FES of any one claims 1-27, wherein the at least one muscle elec- trode is seeded with lower motor neuronal cells.

29. The implantable FES of any one of claims 1-28, wherein the at least one spinal cord electrode, the at least one periphery nerve electrode, and / or the at least one muscle electrode are connected to the at least one hub via an electrically insulated ribbon ca- ble that is biocompatible, flexible and / or stretchable, such as a polyimide ribbon ca- ble.

30. The implantable FES of any one of claims 1-29, wherein the hub comprises one or more of: a battery pack; a pulse train generator; a microcontroller; an inductive coupling transceiver; a radio transceiver; an amplifier; an analog-to-digital converter; an inductive charger; a spike detector or threshold circuit; a memory device; and / or an integrated circuit.

31. A method of electrically interfacing a target tissue with an implantable FES, the method comprising the steps of: providing the implantable FES of any one of claims 1-30; implanting the electrode into a patient; integrating the electrode with the target tissue; operably connecting the feedback control system to the implanted electrode to thereby provide electronic communication between the feedback control system and the implanted electrode; andtraining the feedback and control system using the machine learning algorithm; thereby electrically interfacing the target tissue with the implantable FES.

32. The method of claim 31, wherein the training comprises a modified DDPG (deep de- 5 terministic policy gradient) for FES control and / or a trust region method for DDPG.

33. The method of any one of claims 31-32, wherein the processor processes and trans- mits neural signals as pulse trains, thereby reducing a total required power.

34. The method of any one of claims 31-33, for treating paralysis in a patient by indirect or direct stimulating of a muscle of the patient; and / or receiving a signal from a pe- ripheral nerve.

35. The method of any one of claims 31-34, the method further comprising the step of: training the patient with a defined motor skill to thereby train both neural networks and the patient in a closed-loop configuration.

36. The method of any one of claims 31-35, the method further comprising: transmitting processed and / or synthetic sensations back to the patient from the periphery nerve electrode and the plurality of motion sensors worn by the patient.

37. A method of making an implantable FES, the method comprising the steps of: establishing a bioconnection between biological cells and a conformable array of synthetic electrodes to provide a conformable biohybrid electrode array, wherein the conformable biohybrid electrode array is conformable to a target tissue; and providing a feedback control system comprising a processor in electronic communication with the biohybrid electrode.

38. The method of claim 37, wherein the target tissue is muscle, spinal cord, or a superfi- cial cortex layer.

39. The method of claim 38, wherein the conformable biohybrid electrode array im- planted in a patient is configured to: align cells to follow an existing peripheral nerve pathway to a destination muscle and thereby form a specific connection to a muscle motor unit; attract a target axon to form a specific connection with a biohybrid electrode of the conformable biohybrid electrode array; or provide cells to form a specific contact with target cells in a superficial cortex layer.

40. The method of any one of claims 37-39, wherein each biohybrid electrode of the bio- hybrid electrode array is provided with a single biological cell, the method further comprising the step of: microfluidically introducing the single biological cell to well configured to op- 5 erably connect to a single biohybrid electrode; and repeating the microfluidically introducing step for each biohybrid electrode of the conformable biohybrid electrode array, wherein the repeating step is op- tionally simultaneous with all other biohybrid electrodes.

41. The method of any one of claims 34-37, further comprising the step of using a ma- chine learning algorithm to train the implantable FES to generate an interface scheme provided from the processer to the biohybrid electrode to generate one or more bio- logical outcomes.

Images