Developing clinical grade flexible implantable electronics

Implantable electronic sensors and systems are utilised in an increasingly broad array of clinical applications, such as pacemakers, neuro-modulators and bioelectronic vagal nerve controllers. Advances in microelectronics, materials, and bio-interfaces allow for new clinical applications and support fundamental research. However, a longstanding issue with such devices has been the mismatch between the relative stiffness of such structures compared to tissue softness. This disparity has led to tissue rejection in the form of scar tissue around implantable probes, leading to loss of function and/or capability. This review, therefore, explores the field of implantable electronics and neuroprosthetics with a particular focus on developments in soft, flexible devices. We include advancements in materials and device topologies as well as the current understanding of their long-term efficacy in biological tissue.


Introduction
This review aims to explore the latest advances in implantable electronic devices for clinical use. We first explore the application space and requirements. We then provide a technical view of bio-interfaces before exploring advancements in specific flexible and semiflexible implants.

Application space
Implantable devices comprise many forms-from simple structures such as near-field identification passes to complex brain interfaces to treat neurological diseases. Surgical risks mean that, in general, electronic implants for humans are generally limited to medical use at the current time. The most common clinical application is to interact with the human bioelectrical system-i.e. neuroprosthetics. The impact of this field is significant, and growing. The cardiac pacemaker and cochlear prosthesis markets are worth around $4.5B [1] and $1.9B [2], respectively. The success of these two applications is the driving force for others, such as deep brain stimulation and peripheral nerve bioelectronics.
Most clinical implants follow a common architecture comprising three primary components as shown in figure 1(a): (i) the control unit, which is typically placed subcutaneously, (ii) the bio-interface, which can be: planar, penetrating or cuffed around a nerve, as shown in figures 1(b)-(d), and (iii) the flexible connection between them-usually a Cooper cable [3]. This architecture can be mapped to nearly all historical and modern neuroprosthetics systems, from pacemakers to sensory prosthetics.
The historical evolution of this field has been provided by Norman in 2007 [4] and more recently by Vassanelli et al [5] in 2016 and Lebedev et al [6] in 2017. Briefly, the earliest clinical application was for cardiac pacemaking, which first evolved in the 1950s with Arne Larsson as the first patient [7]. In the 1970s, cochlear prosthetics [8,9] were developed to restore auditory function to the deaf. The bio-interface for this device is a cuff electrode to match the curved shape of the auditory nerve.
Another form of sensory prosthesis device has been visual brain prosthesis. In 1968, Brindley and Lewin [10] first developed a planar bio-interface into the right hemisphere of the visual cortex of a 52 yearold blind female subject. With 80 active electrodes and primitive electronics, they restored 'phosphenes' (flashes of apparently light) and mapped their positions in space. Dobelle and colleagues continued the efforts in the 1970s [11], but only recently, a commercial system has become available, the Orion system [12], by Second Sight Inc. An alternative modality to the planar bio-interface has been the penetrating probe interface, developed by Norman and colleagues at the University of Utah in 1997 [13]. The Utah array was first developed to treat blindness [13,14] but has also been explored in other clinical research projects, such as robotic control for subjects with locked-in syndrome [15].
In the 1990's, it was discovered that in a niche blinding condition called Retinitis Pigmentosa, the communication cells and some processing cells in the eye were still intact. Many researchers then started to develop retinal prosthetics to restore sight by stimulating these remaining layers. The most successful of these to date have been the Argus II [16] implant by Second Sight and the Alpha IMS [17,18] implant developed by Retina Implant AG. Unfortunately, these both failed commercially. So currently, Pixium, another retinal prosthesis company, is targeting a different, more prevalent condition called dry age-related macular degeneration. In all of these cases, the implant is a planar device. However, it should be noted that the Prima system by Pixium [19] acts as a solar cell array with an external headset rather than an implantable control unit.
Also, in the 90 s, the pacemaker architecture used for cardiac resynchronisation was retooled and adapted for brain disorders. It was found that electrodes placed deep in the brain and pulsed at frequencies above 100 Hz could suppress tremors and stiffness in patients with Parkinson's disease [20]. Thus, the field of deep brain stimulation was formed [21]. There are now several major medical technologies operating in this space across the world. More recently, there has been significant interest in the domain of bioelectronic medicine [22]-which aims to treat chronic diseases through the stimulus of the peripheral nerve system.
The field of neuroprosthetics has also seen many new applications in tandem with advances in neuroprobe technology in recent years. Higher stimulation and recording density allow for improvement in spatial and temporal resolution. Improved flexibility also allows new applications. For example, neuroprosthetics has made significant progress in restoring motor function, which is significant for people with paraplegia. Borten et al [23] and Minev et al [24] have successfully demonstrated a limited restoration of ambulation through the use of spinal neuroprosthetics in a brain-spine interface [25].
In summary, there have been impressive clinical implementations of implantable bio-electronics. These have created real improvements in the lives of patients suffering from a variety of neurological conditions. The most important success stories are cardiac pacemaking and cochlear prosthetics, which have no alternative treatment. In addition, niche devices such as bladder controllers for spinal cord injury have an equally important role [26]. Nevertheless, history is not linear, and the commercials failures in the retinal prosthetics field demonstrate that further significant improvement in the bio-interface is required.

Need for flexibility
Implantable devices need to operate stably in the body for several years or even decades. To ensure such long-term stability, it is necessary to consider the biocompatibility of the probe to the biological tissues it contacts.
Of the three primary components of the typical electronic implant, the impact of the bio-interface on its immediate surroundings is the most important because it will affect the therapeutic function as well as the comfort of the device. This impact is affected by surgical trauma during insertion and subsequent tissue reactions post-implantation.
The first form of reaction to an inserted device will be an acute inflammatory response. These can be mediated by microglia and astrocyte activation as per figure 2(a). The second type of reaction results from the probe's existence in the tissue, as per figure 2(b). If no inherent toxicity of the device's surface is present, i.e. chemical biocompatibility is achieved, the primary form of irritation is mechanical. We can define this as mechanical/structural biocompatibility.
The biological tissue, whether the central nervous system or elsewhere, is constantly moving. Muscle movements and flexing constantly change the shape of the body, and from the inside, the heart and lungs are in constant motion. At a smaller scale, the pulsing of the arteries creates constant micro motions. As such, if the probe does not move in unity with the tissue, the subsequent forces can cause micro-abrasions and tearing between the probe surface and the tissue. In turn, this can lead to increased astrocyte and glial response. A detailed exploration of these effects can be found in Mols et al [27] and Salatino et al [28]. But in short, they occur over weeks and months subsequent to probe implantation.
The barrier formed by astrocyte/glial granulation tissues and scars can seriously impact implantable probe functionality, Ersen et al [29] have previously documented the response of astrocytes to implants. Typical scar tissue layers vary from tissue to tissue and probe to probe in the order of 50-200 µm thick [27,28]. Such a layer can insulate electrical stimulation and recording electrodes and cause scattering of light emission used for probing tissue. Most importantly, glial layers around the probe can be devoid of the original nerve cells. Therefore, any stimulation or recording must be performed at a greater distance than when the probe was first inserted.
When a recording probe is first inserted, recordings from individual neuron cells near the electrode are possible. However, as glial layers build up around the probe, the nearest functioning neurons become further away. As a result, recordings need to be increasingly taken from bulk tissue (many neurons) via so-called local field potentials. Similarly, from the stimulus perspective, increasing stimulus strength may be required to provide a therapeutic benefit, which may accelerate the glial/astrocytic processes. From a global therapeutic perspective, this means that the intervention control system needs regular retuning and adaptation.
Since the 1990s, advances in semiconductor manufacturing technology and MEMS (micro-electricomechanical systems) have steadily improved the capacity to manufacture new types of bio-interface. However, traditional semiconductor processing techniques are based on the utilisation of rigid, inflexible crystalline structures.
In order to assess the relative mechanical biocompatibility of a given material, we need to compare their 'hardness' with that of tissue. The Mohs scale determines hardness based on whether one material can scratch another. Similarly, the Rockwell scale measures hardness on the basis of indentation. However, neither are suitable for determining biological 'hardness' . As such, a more helpful way to non-biological and biological materials is by considering material flexibility. A material's (in)flexibility is characterised by its Young's modulus. Thus, a high value for Young's modulus represents stiff, inflexible materials and vice versa. For example, Young's modulus of silicon is 130-185 GPa [30] which contrasts strongly with the softness and flexibility of brain tissue ∼2 × 10 −6 GPa [31]. The range of Young's modulus of materials used in implantable probes vs. neural tissue can be seen in figure 2(c).
It is worth noting that the surface of probes, in many cases, is not necessarily the same as the core structural material. To improve longevity and prevent electrolytic degradation of electrical structures, probes are often coated with parylene or silicone (PDMS-polydimethylsiloxane). The Young's modulus of these materials is much closer to that of biological tissue, as per table 1 and figure 3 below. Structural biocompatibility and tissue reactions (a) the conceptual mechanism for early tissue reaction due to probe insertion (b) long term scar formation due to mechanical incompatibility between probe and tissue (c) the elasticity scale for different materials. Encapsulation materials are listed in red.  [50] a Depending on tissue type and measurement method [51]. Young's modulus has been taken for compressive approach, which is typically 1000x higher than tensile measurements for soft materials.
Therefore, depending on the thickness of their coating, some increased softness at the micro-scale can be engendered. However, if the probe still has a hard substrate, it will still be relatively stiff at the structural scale compared to the surrounding tissue.
Negative tissue reactions can be significantly reduced when probes are smaller and better probe mechanical compatibility can be achieved. For example, Luan et al [32] created a flexible probe shaft just under 100 µm wide and 10 µm thick, and observed no significant gliosis in rodent animal models.

The bio-interface design and functionality
The main classes of bio-interface form-factors have been presented in figure 1. Penetrating probes (figure 1(a)) are generally used for interfacing with the brain tissue and can come in the form of deep probes [52] and intracortical probes such as the Utah array [53]. For less invasive actions such as ECoG (electrocorticography), planar probes can record from the surface of the tissue. Finally, cuff electrodes can be used to record from or stimulate nerve fibres. Examples, in this case, include cochlear nerve prosthetics for the deaf [8,9] or vagus nerve prosthetics for bioelectronic medicine applications [54].
There are many different ways to interrogate biological tissue-neural tissue or otherwise. A recent review has been provided by Walton et al [55] and will be summarised here. Figure 4(a) shows the primary interface modalities: electrical, ultrasonic, optical, magnetic, and chemical methods. All can be used for acute sensing and stimulation. However, only a few of these modalities are practical for chronically implantable probes, as per figure 4(b). Electrical, optical and magnetic probes have been successfully used to stimulate nerve cells, and electrical probes are the primary method for neural recording. A comparison between different stimulation methods is presented in figures 4(c)-(e), with details described in the sections below.

Electrical bio-interfaces
Nerve, muscle and pancreatic beta cells all have electrical operation via trans-membrane ionic flows. Historically, the primary method for interfacing with such cells has been via electrical stimulation and recording. The first scientific description of electrical stimulation dates back to Luigi and Lucia Galvani's frog leg stimulus in 1780 [56]. On the recording side, Edgar Adrian recorded the discharge activity to nerve fibres for the first time through a Lippmann electrometer in 1928 [57]. Hodgkin and Huxley subsequently developed a full electrical model for the nerve in 1952 [58]. The critical advantage of electrical interfaces is their simplicity. They only require an electrode and electronic circuit, with no further transducer.
The earliest electrical stimulation probes were planar, including the first pacemaker in 1958 [7] and implantable visual prosthesis in 1968 [10]. The first cochlear prostheses in the 1970s [8,9] employed planar electrodes incorporated into a cuff to wrap around the cochlear nerve. In the 1990s, the Utah array [53] was developed in the form of a penetrating array fabricated from silicon with electrodes at the tips. Variants of this include creating probes out of metal [59] and high-density active probes such as Neuropixels, which have large numbers of electrodes on their shaft and internal electronics [60]. A simplified historical timeline for electrodes and electrode arrays is given in figure 5.
Despite the variation of stimulation topologies, the basic mechanism is the same: a biphasic charge pulse i.e. with current stimulus comprising balanced negative and positive phases will act to depolarise (and thus stimulate the neurons). The stimulation can be set up with a distant tissue grounding electrode for unipolar stimulation or a nearby counter electrode for bipolar stimulation and current steering [61]. An additional constraint is that the charge density of each of the pulses with respect to the   [73]. Mesh electronics [74]. Neuralink electrodes [75]. Luigi Galvani probes in 1780 [56]. Hans Berger recorded the first human EEG in 1924 [76], The first Michigan silicon probe [77], The Utah Array [78] Neuropixels active probes [60]. electrode area must not exceed a certain limit. If the stimulation charge density exceeds Shannon's empirical limit, then tissue damage can occur. If the stimulation charge density exceeds the charge injection capacity of the electrode material, then the electrode can rapidly degrade. A detailed review by Cogan [62] explains these effects in more detail.
Electrical recording, as the name suggests, directly determines voltage fluctuations in the neural tissue. The type of recording is determined by the distance of the electrode to the neurons. Individual action potentials (or unit potentials) can be determined when electrodes are close enough to neuron cells. This can be achieved using penetrating electrodes such as figure 1(b). At a further distance, local field potentials, the aggregate activity of local neurons. Brain surface recordings are measured at the surface of the brain. These can be achieved with planar electrodes such as figure 1(d). Recording of nerve bundle activity is typically achieved using cuff electrodes such as in figure 1(c).
Neurologists such as Hermann von Helmholtz [56], Gasser and Erlanger [63,64], Hodgkin and Huxley [58] have continued to update the recording  [83] Klein et al with LEDs on a Flexible probe [84]. Stretchable optoelectronic neural interface by Ji et al [85]. Graphene electrodes with blue LED by Park et al [86]. Flexible substrate dual-color micro-LED probes by Lizhu et al [87] with a microLED array by Grossman et al [88], Utah array integrated with an optic fibre by Wang et al [89], waveguide probes by Zorzos et al [90] silicon probes with integrated LEDs by Wu et al [91], semi-flexible LEDs integrated on an inductive coil by Shin et al [92].
performance of neuroelectrodes in the early stage. In particular, in the twentieth century, advances in manufacturing processes led to the Utah array and Michigan array, which enabled the manufacture of high-density neuroelectrodes in bulk, allowing for high spatial and temporal resolution signal acquisition, and were widely used in clinical and research applications [65].
Recent advances in CMOS technology [66] have further expanded the possibilities of recording electrodes. Based on the development of Michigantype probes, the Neuropixel probe [66,67] enables programmable probes that can be switched and addressed between electrodes through CMOS circuitry. Neuropixels 2.0 probes have now been tested experimentally on humans [67].
In addition to recording electrodes to improve spatio-temporal resolution, there has also been an exploration of new forms of electrode probes. Of notable novelty has been the Stentrode developed by Oxley et al [68]. These are intravascularly implanted to simulate or record from target brain regions from the blood vessels. The device was implanted via minimally invasive surgery in four participating patients with amyotrophic lateral sclerosis and has just received FDA approval [69]. The researchers followed the patients for a year, who were able to use the interface to control computers to assist them in their daily lives.
In conclusion, the ideal implantable electrode should have small size, high density, good biocompatibility, and good electrical properties (including low impedance, high signal-to-noise ratio, and high conductivity stability). In addition to traditional metallic materials, there have been many advances in the research of conductive polymers, graphene and other carbon materials [70].

Optogenetic stimulation
Most multicellular organisms have some form of optical mechanism-whether for energy conversion or visual sensing. In 2003, Nagel and coworkers [79] demonstrated the genetic expression of a depolarising channel (Channelrhodopsin) into non-neuronal cells, allowing for optical stimulation. Additional expression of hyperpolarising channels into neurons was subsequently performed by Boyden et al in 2005 [80]. These discoveries created the field of optogenetics-the genetic altering of cells to be controllable by light. Later, it also became possible to genetically engender fluorescence sensing of ionic activity, e.g. with GCAMP [81]-though this latter function is for acute scientific interest and less likely to be used in chronic human-grade implants.
The essential advantage of optogenetic technologies is the genetic targeting of specific classes of neurons allowing for specific stimulation as per figure 4(a). Over the past two decades, researchers have developed probes that generate light locally or transmit light into the targeted area via light guides. Figure 6 provides a historical perspective of the improvements in the optogenetics field.

Magnetic stimulation
The connection between electricity and magnetism dates back to Maxwell's equations. Functional magnetic resonance imaging monitors neural activity indirectly via changes in blood oxygenation. However, conversely, transcranial magnetic stimulation can be used to modulate neural activity [93] via largearea stimulation of the brain surface. Bonmassar et al [94] and Lee et al [95] have also shown that it is possible to incorporate micro-coils onto probes that could be used for chronically implantable stimulators. In these cases, the stimulation could be much more local and focused. There is also some evidence that the direction of the field can selectively stimulate certain cell types-e.g. pyramidal cells, which are perpendicular to the cortical surface as opposed to interneurons which are parallel [96], as illustrated in figure 4(e). This domain is still very new but provides a lot of interesting scope for the future.

Flexible probe substrates
Active implantable devices are expected to survive for many years in the body. So, ensuring their longevity is essential. As such, any (typically organic/polymer) coating must ensure that there is no water ingress from which electric fields can begin the process of electrolysis and device failure. This can take many years of detailed accelerated lifetime testing (e.g. see Donaldson et al [97]) and is related to the manufacturing method (and any defect formation) and material properties.
Many types of materials can act as either an underlying substrate or an encapsulant to separate electronic components from tissue/fluid contact. The applications and properties of various materials and various types of probes are reviewed by He et al [98]. However, there are only five materials that have either been used or are being seriously proposed for active neural implants: Polyimide, PDMS, polyurethane, parylene, and liquid crystal polymer (LCP). Their details are provided in table 2, and in the sections below. Kipping in 1904 [125]. It is a silicone polymer comprising alternate silicon and carbon atoms interspaced with oxygen. The silicone chains are methylated, resulting in a highly hydrophobic material. PDMS is one of the most commonly used medical polymers because it has very good biocompatibility-it is soft and causes very limited foreign body reactions. As such, it has received The U.S. Pharmacopeial Convention Class VI clinical approval for unrestricted use in chronic implants. PDMS is available in the form of a viscous liquid that needs to be crosslinked with a curing agent and heat, as per Griscom et al [126] and Zaaimi et al [127].

Polyimide
Polyimide has been widely used in biomedicine in the past 30 years, especially for nerve implants. Polyimide's insulation, thermal oxidation stability and chemical resistance can protect nerve implants from the erosion of the human environment. It has good optical clarity, and good chemical biocompatibility [128]. However, polyimide is significantly harder than PDMS-with a Young's modulus of around 2.5 GPa compared to ∼1 MPa for PDMS. Polyimide comes in many different chemical variants. A method for in-situ patterning is to deposit dissolved precursors, such as polyamic acid, followed by UV and thermal curing [117,129]. Patterning can be achieved through direct writing technology [130,131], or O 2 gas plasma [130].

Polyurethane
Polyurethane is the polymerised form of urethane. The urethane monomer is both toxic and carcinogenic and is used as a terminal anaesthetic component in animal studies [132]. In contrast, the polymer is commonly used for many different industrial and consumer work surfaces and is considered non-toxic [133]. However, there have been no long term safety studies specifically exploring its use in implantables. The reason for the interest in this material is that it closely matches the acoustic impedance of saline/tissue [118]. As such, it is an ideal coating material for ultrasonic implants [134]. It has also been considered for the plastic header of pacemaker pulse generator units [135]. Polyurethane, similar to silicone, can be deposited as a two part resin and moulded.

Parylene
Parylene is the common name for polyparaxylylene. Parylene is a linear, non-crosslinked and semicrystalline polymer. Parylene was first discovered by Micheal Szwarc in 1947 through the thermal decomposition of p-xylene H [136][137][138]. Parylene is available in several forms. The most widely used is parylene-C, but for optical applications, parylene-H/HT has better transparency at shorter wavelengths. It is deposited as a vapour, then pyrolysed into a monomer gas as per Ratier et al [139][140][141]. Subtractive patterning (e.g. to open electrodes) can then be achieved with photolithography and plasma etching as per Yeh et al [142] and Ratier et al [139]. However, the vapour phase polymerisation method means that it is relatively crystalline. So it is a relatively hard encapsulant with a Young's modulus of around 3 GPa.

Liquid crystal polymer
Liquid crystal polymer (LCP) is a separate category of polymers. It is a thermoplastic polymer that consists of both rigid and flexible monomers to create a liquid crystal structure. It has been explored by Jeong et al [143] for retinal implants because of its very low water absorption. It is thereby hoped that such implants can achieve exceptional longevity. Prévôt et al [144] demonstrated its biocompatibility with biological tissue in the form of 3D scaffolds. As a thermoplastic, the primary form of handling is through  [115], UV curing [116] UV exposure [117] Moulding/hot pressing [118] Vapour deposition polymerisation [119] Hot pressing [120] Glass transition temperature (T4,  hot embossing and moulding [120], with subtractive removal via laser [145].

Flexible probes
Here we review four notable architectures using soft substrates suitable for cortical surface, intracortical, and deep brain operation. Their summary is provided in table 3 below. Khodagholy et al [146] presented an all-polymer approach to the measurement of electrical activity from the cortical surface i.e. ECoG. Their electrode architecture was called the 'NeuroGrid' . This was an ultra-thin 4 µm flexible surface probe made of parylene C, which encapsulated platinum gold internal wiring. High resolution (10 × 10 µm electrodes with a 30 µm inter-electrode spacing) surface electrodes comprised of PEDOT:PSS-a polymeric, synthetic metal electrode. The device demonstrated stable operation in rodents for 10 d as well as acute operation in humans. However, it should be noted that ECoG recording is typically used as a short-term diagnostic tool for epilepsy rather than a long term implant.
For intracortical measurements, Chung et al [147] created a penetrating polyimide-based probe. Their probe comprised a polymeric polyimide substrate with Titanium/gold metal lines and PEDOT:PSS surface electrodes. The probe was in a fork arrangement with four shafts; 5 mm long 80 µm wide, and 14 µm thick. Each shaft had 16 electrodes at the tip of each. The team also demonstrated arranging eight systems in parallel for a total of 512 channels per stack. This arrangement is primarily suitable for cortical recordings and testing in rats that demonstrated stability for up to 1 week of testing. These probes were then developed further by the Neuralink team [75], who developed the core architectures into 'threads' . i.e. individual shafts with 32 electrodes which could be sequentially inserted into the cortex. Using this approach, they increased their total electrode count to 3072. The paper stated that recordings were achieved from a 'chronically' implanted device, but no further details were provided.
Structure to be inserted into the deeper brain regions typically need to be initially compressed into a tube and then released to expand out into the target region. As such, Xie et al [148] created a mesh framework using polyimide and SU-8 (a non-medical grade photoresist). Interestingly their structure utilised a mechanism to have 'arms' bend away from  [74,75,82,155,161,162,165] the mesh so that once unsheathed individual arms would potentially penetrate through tissue. They did not present long term histology effects or recording stability but could show tissue integration with the probes.

Making hard materials flexible
Existing hard technologies have been presented in figures 5 and 6. Common technologies such as the Michigan probe [126], Utah array [47], and the Neuropixels probes [60,149] have achieved highdensity probing. But they are based on hard silicon substrates. As such, there can be issues with long term mechanical biocompatibility, as described above. As such, one possibility is to migrate manufacturing towards polymeric materials. However, polymeric electronics do not perform to the same degree as silicon. Similarly, hard photonic technologies such as GaN and AlGaP can achieve much higher radiance densities than their polymeric equivalents. As such, there is still a need to connect hard components to soft probes and, where possible, make those hard components flexible. In contrast to its bulk stiffness, when glass is thinned in both dimensions down to a narrow (optic) fibre, it becomes flexible. It is still brittle and has a minimum bend angle, but it is much more flexible than before. As such, two primary strategies can be performed to engender flexibility in hard materials.
The first approach is to super-thin the hard material. Bedell et al [150] reported a controlled spalling technology to make otherwise rigid CMOS chips flexible. This approach effectively splits off the top part of the chip from the bulk substrate by creating a stress fracture and steadily peeling off the top layer. The same approach can be used to thin LED materials such as GaN [151]. Alternative approaches include chemical etching, plasma etching and perhaps mechanical or laser thinning.
Another approach is to pattern the hard material to have flexibility in given directions. For example, Dinyari et al [152] patterned a silicon substrate to have islands suspended between serpentine (See Gutruf et al [153] for mechanical analysis) connectors. This allowed the otherwise rigid planar surface to stretch and bend to be conformal with surfaces. Such approaches can be taken to the extreme by super thinning. For example, Hong et al [74] created a highly flexible mesh network of silicon nanowires to make a new type of neural probe.

Incorporation of microelectronics and MEMS
With traditional passive (whether flexible or rigid) probes, increasing the number of sensors and stimulators also increases the number of control lines. These become the limited feature as thick cables can exert undesirable forces on the bio-interface, exacerbating mechanical biocompatibility issues. As such, active electronic probes such as the Neuropixel probe [75] constitute a significant advance in the capability of bio-interfaces. Incorporating micro-electronics onto the probe can allow for multiplexing of large numbers of sensors/stimulators and thus have a relatively small control cable. However, such probes are rigid.
One method is to try to combine both worlds by attaching silicon control chips onto existing probes. For example, Ramezani et al [154] used CMOS chips bonded to a (rigid) optical probe to provide local control electronics and multiplexing. In contrast, The Neuropixel probe [75] connected CMOS chips with 256 contact pads (+ 96 pads for power and external communications) each to connect to eight individual threads. Multiple chips were used to connect to the full-thread array. In addition to CMOS, MEMS is also of interest to implantable electronics. Jeong et al [155] reported thinning a piezoelectric energy harvester to make it sufficiently flexible for implantation. Thematerial composition of the various semi-flexible, hardflexible probes issummarised in table 4.

Incorporation of optical emitters
Optogenetic tissue stimulation requires intense tissue irradiation in wavelengths between 470 nm and 570 nm, depending on the opsin used. The typically accepted 'threshold' for channelrhodopsin-2defined as the irradiance required to achieve 50% of the maximum neural response is 0.7 mW mm −2 [156]. However, as light disperses through the tissue, light radiance from the emitter needs to be in the range of 100 s of µW to a few mW depending on stimulation volume [157]. This is significant and needs to be delivered either via light guides or light-emitting diodes.
The earliest two-dimensional arrays demonstrating optical emission were by Poher et al [158] and Grossman et al [88]. Later, Mac Alinden et al developed a penetrating optrode based on the sapphire substrate of the LED substrate [159]. Similarly, Ramezani et al [154] proposed an optrode based on mini-LEDs attached to a silicon substrate. However, Kwon et al [82] developed a flexible planar ECoG probe using micro LEDs on a PDMS substrate.

Semi-flexible probes
As per section 2.2, we can think of probes in terms of cortical surface, intracortical and deep penetration. For the former, Kwon et al [82] developed a flexible planar ECoG probe using micro LEDs on a PDMS substrate. Jeong et al [155] reported a similar surface structure with an additional flexible energy harvester.
Hong et al [74] demonstrated a mesh electronics structure whereby waveguides and electrical wires are encapsulated in a mesh-akin to a fishing net. This system could be delivered using a hollow tube guide, which is then retracted, leaving the mesh in place. The team showed stable action potential for 4 months and recorded no significant immune response after 1 year of implantation. The structure of the internal electronics was not detailed but was based on previous work with silicon nanowires passivated with silicon nitride [160]. These were then coated with biomolecules such as biotin.

Manufacturing flexible probes
The standard method of the semiconductor industry is to utilise photolithography, as per figure 7(a). Fundamentally, this technique allows the patterning of thin layers of polymers using UV polymerisation to make them selectively soluble. In turn, this allows selective deposition or removal of metals and other materials. It is essentially a 2.5D process, i.e. individual 2D layers can be removed or added until the structure is complete. Typically this process would begin on a hard silicon substrate which would then be removed after the completion of the probe.
The primary method of polymer layer deposition in photolithography is spin coating-which provides a flat, even layer. Alternatives include simple dipcoating with centrifugation (image not shown) or vapour phase deposition (See figure 7(b)), which is the case for parylene deposition. In this latter case, the advantage is the ease with which the material can underfill gaps or cavities [166].
To pattern polymer layers using photolithographic patterns, they must either be converted from soluble to insoluble form in the required pattern. Or subsequently removed in some way. However, apart from polyimide, the medical-grade polymers described above are not photo-polymerisable (also known as photo-curable). Therefore, one alternative is to use inkjet deposition methods to model desired polymers, as per figure 7. For example, Nana Kokubo [167] showed a 6 µm thick flexible sheet probe based on inkjet printing technology. More recently, Hyunwoo Yuk [168] demonstrated a conductive polymer PEDOT: PSS-based 3D printing ink material in 2020, which can realise co-printing with PDMS ink to manufacture a full flexible probe.
Alternatively, dry contact printing methods can be used. This is where polymers are deposited on a flat surface, and a patterned stamp is used to remove selectively transfer layers. This is achievable by temporarily making the stamp more adhesive to assist the transfer process. Meitl et al [169] introduced a transfer printing technology in 2005, which can arrange inks composed of multiple materials on a polymer material substrate through a soft elastic film. This was further demonstrated by Chen et al in 2008 [170]. This is based on setting the ink on the donor substrate, then picking up the ink on the donor substrate by stamp, and finally, using the stamp to print and transfer the ink to the substrate of the receiver. This approach requires the target to have greater adhesion than the stamp. One approach is to use a plasma treatment to temporarily increase adhesion in the stamp [170]. Further methods include: Modifying the surface chemistry [171] glue assisted transfer printing [172], kinetically controlled transfer printing [169], laser-driven non-contact transfer printing [169], gecko-inspired (van der Waals forces on fibrillar surfaces) transfer printing [173] and aphid-inspired (contact area variation) transfer printing [174].
The final method for patterning polymers is through direct moulding and/or hot embossing, as per figure 7(e). This is suitable for polymers that do not dissolve in solvents-e.g. PDMS is an epoxyi.e. it forms when its monomer and mixing agent are combined. For example, Jaafar et al [118], moulded polyurethane around their ultrasonic implant to act as both an encapsulant and to support ultrasonic communication transfer to the tissue.
Deep cutting is required to cut out probe shapes. There are several methods to achieve this. For traditional hard probes such as the Utah array [53,175], dicing saws could be used to cut out evenly separated trenches. By doing this in two directions, pillars can be left, forming penetrating probes. However, there needs to be patterned cutting for flexible and semiflexible systems. For example, Dinyari et al [152] created a flexible silicon array for potential uses in retinal prosthetics. This patterned cutting can be achieved with vapour phase removal-known as deep reactive ion etching. This is analogous to figure 6(d) image, albeit with layer removal rather than deposition. As it is a ballistic process, it can anisotropically etch deep trenches through unprotected surfaces. Other alternatives are to use laser ablation or mechanical cutting [176]. However, undesirable toxic compounds can be formed for some types of laser cutting if the settings are incorrect [177]. The final consideration is the attachment of electronic components to target substrates. Traditional methods include conductive adhesives such as silver epoxies as per figure 7(g). For example, in earlier work, we utilised silver epoxy for flip-chip bonding of both LED arrays [88,178] and LEDs onto single probes [157]. However, adhesives yield and reliability can be challenging depending on the application. An alternative to adhesion for conductive contacts is to use metallic bonding as per figure 7(h). In the latter case, metallic bonding can be achieved via gold wire bonding (e.g. Ji et al [179]) or direct flip-chip bonding techniques (e.g. Soltan et al [180]). However, an important consideration is that the high temperatures involved can be problematic for soft polymer substrates, typically with glass transition temperatures below 100 • C (see table 2).
Furthermore, due to different expansion coefficients, layers bonded at high temperatures can create stresses as they cool. In the long term, such stresses can lead to failure or device rupture. To counter these challenges, lower temperature solders and metals such as indium can be used.

Implantation strategies for flexible probes
Planar probes that simply sit on the target tissue or cuff electrodes that wrap around a target nerve (see figures 1(c) and (d)) are relatively straightforward to implant. For example, planar ECoG arrays are regularly implanted temporarily in the brains of epilepsy patients to determine subsequent treatments [181]. Retinal prostheses are also planar probes which are inserted in the retina (e.g. Argus II retinal prosthesis system [182]). Similarly, vagal nerve (cuff) implants are regularly used for a variety of conditions [183].
Rigid probes are also commonly implanted into human brain for applications such as deep brain stimulation [184] or cortical recording [185]. However, in these instances, micro-motion around implanted rigid probes is thought to play an important role in chronic inflammation and astrocyte scar formation [186]. He [187] discusses the mechanics of the relevant shuttle devices related to the analysis, so micro-motion more gentle flexible probes became popular. However, flexibility adds complexity for probes that need to be inserted into the brain. Parittotokkaporn et al [188] showed that for larger probes such as deep brain stimulators (1.3 mm diameter), insertion forces required to cut the tissue could be as high as 1 N. This compares to ∼2 mN for extraction from already displaced tissue. Smaller probes will have reduced forces scaling with surface area, but by design, these forces exceed the buckling strength of flexible probes [189].
As such, for comparison with rigid probes, stereotactic equipment is used to position the probe to the target location (as defined by e.g. MRI imaging). Inserting a single probe, such as a DBS probe, can then be slow and steady to ensure minimal damage. However, this is not always been effective. For example, the insertion of Utah arrays can cause the brain to depress rather than allow penetration. As such, surgical devices have been developed for fast insertion. For example, Rousche and Normal [78] developed a method of high-speed insertion between 1 m s −1 and 11 m s −1 (figure 8(a)).
A standard method to insert flexible probes is to use a rigid shuttle to transfer the probe into the  [78,198]. (b) shuttle systems for inserting individual probes. Image right is from Kozai et al [190]. (c) Robotic insertion methods for multi-shuttle probe insertion. Right, the robot developed by Neuralink [75]. (d) Stiffeners are used to provide temporary hardness during probe insertion. Right, image from Felix et al [194]. target and then retract the shuttle leaving the flexible probe. For example, Kozai and Kipke [190] demonstrated shuttle insertion of retinal prosthetic devices into the retina (albeit, strictly speaking, the retinal stimulator was inflexible whilst the ribbon cable to the external device was flexible). Another example is proposed by Joo et al [191], who demonstrated a silicon shuttle device that uses a single-handled silicon shuttle. The flexible probe can be implanted into the brain through the dura mater. This has been illustrated in figure 8(b), with an image by Kozai et al [190]. Another shuttle variant is to fold the flexible probe and then inject the probe into the corresponding part through a micro-syringe. The flexible mesh electrode can slowly unfold and penetrate into the tissue after entering the tissue [74,192].
A single shuttle solution can work with individual probes but can become laborious and difficult with multiple probes. As such, Neuralink [75] demonstrated a robot that can fully automate multi Flexiprobe insertion at and high speed, with automated cleaning and sterility functions. This is illustrated in figure 8(c). The robotic arm can insert six flexible probes (192 electrodes) per minute.
Finally, an alternative to shuttle insertion is to engender temporary stiffness to the probe. Xie et al [148] reported in 2015 a solution to freeze the probe and then implant it. After the probe enters the human body, it would gradually thaw in the tissue. A more common method in the literature [193] is to coat the probe with a hard biodegradable material, which temporarily changes the stiffness of the flexible probe during surgery. For example, Xiang et al [161] demonstrated a temporary maltose coating for probe insertion. Similarly, Felix et al [194] and Takeuchi et al [195] used poly-ethylene glycol, while Wu et al used silk [196], and Kozai et al [190,197] used carboxymethyl cellulose. This has been illustrated in figure 8(d), with an image by Felix et al [194].

Discussion and conclusion
This review focuses on the exploration of flexible electronic devices for the clinical application of neural prostheses.
Since the early neural implants in the 1960s, their designs have undergone several iterations. In tandem, the performance and precision of neural implants have improved with advances in manufacturing processes and materials science. Several mainstream neural probes have evolved: penetrating, cuff/stent, and plane/surface. Stimulation modalities include electrical, optogenetic, and magnetic field stimulation.
Probe morphology has evolved in response to the clinical challenges of surgery and tissue response. As such, from the initial planar Pt contacts, there are now different shapes to fit various target tissues. Flexible probes with better mechanical-biocompatibility have also started to receive more attention and preference from researchers. Flexibility causes a reduced mechanical perturbation in the tissue, and thus can reduce inflammatory responses. Furthermore, the ductility of flexible probes can allow better fitting of the probe at the tissue for better spatial resolution.
Currently, flexible probes still face a number of challenges: The fabrication of flexible probes faces many difficulties. Flexible probes commonly use a mixture of different rigid materials for their structure, where flexible materials such as organics are used for substrates and encapsulation, and rigid materials such as metals and LEDs are needed for neural stimulation and signal transduction. In addition to the challenge of joining and combining these different materials, flexible probes also tend to be smaller in size for better ductility, which further increases the difficulty of manufacturing probes. The composite structure described above also poses additional challenges for large-scale fabrication.
Semiconductor fabrication processes currently used in the fabrication of rigid silicon probes can only be partially transferred to flexible probe manufacture. Furthermore, probe insertion into tissue still requires optimisation-via temporary hardening layers or guide structures. A particularly interesting system has been the surgical robot presented by Hanson et al [199] and the Neuralink Inc. team. It has the capacity can implant a large number of neural probes in a short period of time, which increases what is possible within a defined surgical time window.
The interaction between neural tissue and neural implants can affect the longevity of neural implants [200]. The body's process of repairing chronic neurological tissue damage from brain micro-motion can isolate the neural implant from the brain tissue. Greater distance from the target tissue, and encapsulation by glial cells can reduce the therapeutic efficacy of the probe (as per Turner et al [186]). The organic substrates widely used in flexible probes generally have better biocompatibility, and the softer mechanical properties of flexible organics can reduce damage to tissue [201,202]. The unbalanced electric field caused by the pulsed stimulation of the electrodes [203] causes a change in pH in the vicinity of the electrodes, which accelerates the corrosion of the electrodes and renders them ineffective [202,204].
A further aspect to probe flexibility is the incorporation of transducers directly on the probe. Neural dust [205][206][207][208]-although not directly flexible have passive (backscattering) ultrasonic transducer system which could be incorporated onto wireless flexible probes. Similarly, near-field RF approaches similar to contactless card reading can be used to probe implants. These would be most applicable to subwirelcutaneous systems such as diabetes monitoring.
In summary, in the future, flexible probes will be an important tool in both medical Neuroprosthetics and fundamental neuroscience research. Better mechanical properties and mechanical biocompatibility of neural implants will greatly reduce mechanical damage. Finally, a major objective is how to improve the spatial resolution of neural recording and stimulation. Such advances would lead to better long-term exploration of the brain in fully awake moving animals, as well as advance therapies to restore lost neural function.

Data availability statement
No new data were created or analysed in this study.