A robust, real‑time telemetry protocol for miniaturized neural implants using off‑the‑shelf hardware

Abstract

Background

Wireless telemetry from fully implanted, millimeter-scale neuromodulation devices is constrained by tight power budgets, inefficient antennas, and in-body attenuation. Standard protocols (e.g., BLE) offer mature ecosystems but exhibit throughput shortfalls and reduced robustness under non-ideal operating conditions. This study introduces the Neural Real-Time Telemetry Protocol (NRTP), a 2.4 GHz, half-duplex protocol designed to address the unmet need for reliable, real-time neural telemetry from miniaturized implants using off-the-shelf hardware, without custom electronics or ASICs.

Methods

NRTP was implemented on commercial 2.4 GHz hardware with static-length packets, immediate acknowledgments, bounded retransmissions, and single RF channel operation. We evaluated three mitigation strategies—retries, sample-level interleaving, and data overlapping—individually and in combination, and defined a quantitative evaluation metric that prioritizes data quality and power draw. Using identical hardware for NRTP and BLE, we performed controlled sweeps of received signal strength, tested multiple payload lengths and timing configurations, and measured throughput, data loss, and current draw.

Results

NRTP sustained zero data loss down to -75 dBm, whereas BLE performance degraded below -55 dBm due to throughput shortfalls under interference and deferred unlimited retries. Interleaving converted contiguous gaps into half-rate segments, delaying score decline at lower received signal strength; overlapping improved robustness but its doubled packet rate requirement was power-prohibitive for implant constraints. Across variants, NRTP delivered higher scores and lower variability over a wider operational range than BLE; BLE’s greater scores at high signal strength were driven by lower current consumption but fell off earlier with attenuation. The observed link-margin advantage for NRTP (up to ~ 23 dB at first loss; ~ 11 dB at 0.5% loss) implies ~ 3.2 × range in air, ~ 2.5 cm greater implant depth in tissue, or equivalently lower TX power for similar performance.

Conclusions

NRTP provides robust telemetry for miniaturized implantable devices and is readily adoptable on commodity 2.4 GHz hardware. Its immediate, bounded retries and optional interleaving sustain throughput and minimize true data gaps under attenuation and interference, outperforming BLE across operating conditions relevant to small-animal implants. The resulting link-margin gains translate into practical benefits in coverage, implant depth, and power consumption, lowering barriers to chronic, closed-loop studies.

Supplementary Information

The online version contains supplementary material available at 10.1186/s42234-026-00201-3.

Introduction

Preclinical studies play an important role in the discovery and development of new bioelectronic medicine therapies. Large animal models are typically used for the final stages of translational work, where therapies can be tested by adapting existing clinical devices. Initial discovery work is primarily performed in small animals, where interventions can be studied in the context of widely available disease models. However, the technologies for performing such experiments remain nascent because of the challenges in developing the necessary miniaturized devices. Recent advances have enabled long-term evaluation of interventional strategies, allowing researchers to study chronic treatments over biologically-relevant timescales in small animal models. These advances include both tethered approaches (Mughrabi et al. 2021) and wireless platforms (Wright et al. 2022; Wright et al. 2019), with the latter permitting more advanced paradigms by allowing animals to behave freely and reducing stress from repetitive handling. Such wireless implantable devices not only offer stimulation capabilities but can also facilitate the monitoring of physiological and neural signals. This real-time telemetry allows assessment of intervention efficacy, enabling closed-loop approaches. However, reliable wireless streaming of data from these miniature, resource-constrained implantable devices presents significant challenges due to size and power constraints (Aghagolzadeh et al. 2011).

Here we present Neural Real-Time Telemetry Protocol (NRTP), a 2.4 GHz, half-duplex protocol tailored to miniaturized fully implantable devices, incorporating static-length framing, immediate acknowledgements, bounded retransmissions, single-channel operation, and heartbeat-based discovery. We formalize an evaluation framework and implement loss mitigation strategies—retries, sample-level interleaving, and data overlapping—for systematic comparison. We introduce a quantitative metric that prioritizes net data preservation (minimizing gaps in time), data quality (full- versus half-sampling-rate segments), and decreased current draw, enabling objective ranking between Bluetooth Low Energy (BLE) and across variants of NRTP. We perform controlled Received Signal Strength Indicator (RSSI) sweeps to isolate protocol effects, and report that NRTP sustains zero data loss down to approximately − 75 dBm, whereas BLE degrades below − 55 dBm due to throughput shortfalls under unlimited retries. NRTP variants with mitigation strategies further improved performance in poor operating conditions. Critically, NRTP addresses a key unmet need: reliable, real-time neural telemetry from millimeter-scale, fully implantable devices. By operating on commercial 2.4 GHz hardware and requiring no custom electronics or application specific integrated circuits (ASICs), NRTP is immediately adoptable and reproducible, lowering barrier to entry while meeting the power, size, and attenuation constraints of small-animal implants.

Background and related work

Wireless telemetry for implantable devices

Wireless telemetry enables freely behaving small-animal studies (Wright et al. 2022; Datta-Chaudhuri 2021) but imposes stringent constraints relative to wired systems due to miniaturization requirements. Limited battery capacity (Ghovanloo 2006) and sub-wavelength implanted antennas reduce link quality, while in-body placement compounds losses via tissue attenuation (Chow et al. 2013; Lin et al. 2024). At 2.4 GHz, attenuation in tissue analogues is ~ 0.4 dB/mm (Christoe et al. 2021), diminishing range and increasing susceptibility to interference from coexisting devices. Orientation-dependent coupling further degrades performance and cannot be controlled in freely moving animals. Neural telemetry demands substantial throughput: a single 16-bit channel at 20 kS/s requires 320 kb/s. For context, uncompressed 16-bit, 44.1 kS/s audio is ~ 700 kb/s, but is typically compressed to 96–256 kb/s (Zyka 2020) before transmission and originates from high-power devices with efficient antennas—conditions that do not apply to fully implanted systems.

Standard wireless protocols such as BLE and Wi-Fi offer mature ecosystems and features that reduce development time and cost for neural recording systems. BLE has supported single-channel spike streaming in small animals and non-human primates (NHPs) (Idogawa et al. 2021) and dual-channel spike and local field potentials (LFPs) in mice (Wang et al. 2022) typically as externalized head-stages that avoid in-tissue attenuation. Although BLE 5.x specifies a 2 Mb/s physical layer, practical throughput is lower. Nordic Semiconductor’s proprietary Enhanced ShockBurst (ESB) runs on the same hardware as BLE, trading BLE’s richer features for potentially lower latency and power. Both ESB (Gagnon-Turcotte et al. 2017) and BLE (Liu et al. 2016) have been paired with on-board compression to stream up to 32 channels, but this requires high-power electronics or a custom ASIC for real-time compression. Wi-Fi (2.4/5 GHz) has been used in NHP head-stages (Lee et al. 2023), overcoming BLE’s bandwidth limits and delivering data rates greater than 10 times BLE’s theoretical maximum, but requires substantially more power: typical BLE maximum transmit (TX) power is ~ 6.3 mW, whereas Wi-Fi peaks can exceed 100 mW. Consequently, Wi-Fi systems generally rely on externalized power sources and are implemented as head-stages rather than fully implanted devices.

Custom radios utilizing unlicensed operation in the Industrial, Scientific, and Medical (ISM) radio band can be built with off-the-shelf components or ASICs. High-GHz links offer substantial bandwidth (Calado et al., 2021): a 5.8 GHz wideband Frequency Modulation (FM) system transmitted 15 analog neural channels without digitization (Roy et al. 2011) and a 60 GHz architecture achieved 6 Gb/s during benchtop testing (Kuan et al. 2015), potentially supporting thousands of digitized channels. Sub-GHz designs trade bandwidth for better tissue penetration, longer range, lower power, and reduced crowding; e.g., a 900 MHz ASIC-based radio on an insect head-stage supported multi-channel telemetry (Harrison et al. 2011).

There are also hybrid approaches which combine standards with custom links: BLE has been used to provide reliable control/configuration while telemetry is carried by a custom radio (e.g., 434 MHz (Lee et al. 2016; Jia et al. 2017) or 6.5 GHz (Kanchwala et al. 2018). Multi-radio systems can select bands per use case; one device integrated three radios from sub-GHz to 10 GHz to meet range and bandwidth requirements (Kassiri et al. 2016). For systems with Wireless Power Transfer (WPT), the WPT channel itself can carry data (Jegadeesan et al. 2015; Rajavi et al. 2017; Yang et al. 2014; Yoo et al. 2021). These data links can be high bandwidth but are constrained by the short range of inductive coupling, typically a few mm. Joint optimization of power and data often entails tradeoffs because optimal design choices may diverge. A practical design therefore requires selecting a compromise in operation that balances delivered power, transfer efficiency, and telemetry rate given an application’s specific requirements.

Categorization of techniques to handle data loss: PMR

Data loss is unavoidable in real world wireless links, so protocols employ complementary strategies that fall into three categories: Prevention, Mitigation, and Recovery (PMR). Prevention reduces the probability of loss by improving link conditions through hardware, environmental, or power decisions — for example, increasing TX power, minimizing potential sources of interference, and optimizing antenna alignment. Mitigation maximizes the amount of unique data delivered despite losses by changing how data are sent. Mitigation techniques include packet retransmissions (retries), interleaving to spread samples across packets, and overlapping to introduce structured redundancy. Mitigation techniques require specific packet handling before and after transmission. Recovery approaches reconstruct or estimate missing data during or after reception. Such approaches include parity packets for error correction and signal processing (e.g., regression) to estimate missing samples.

Existing strategies for overcoming data loss

BLE incorporates strategies for dealing with packet loss. One approach BLE utilizes is called Adaptive Frequency Hopping, commonly known as “channel hopping”. While BLE is said to operate at 2.4 GHz, the exact carrier frequency is one of 40 channels operating from 2.4 GHz to 2.4835 GHz, with each channel spaced 2 MHz apart (the channel bandwidth). Channel hopping is when two connected BLE transceivers agree to switch channels at the start of each connection event, spreading communication across the spectrum and reducing the probability of collisions. A collision is when two or more transmissions coincide on the same channel, resulting in packet corruption. A channel map keeps track of reliable channels and is frequently updated and exchanged between the transceivers to avoid switching to unreliable channels. A channel is considered unreliable when two sequential corrupt packets occur. This spread spectrum technique can be seen as a prevention strategy. Additionally, BLE leverages two mitigation strategies in the form of retries and data whitening. Regarding retries, BLE utilizes an error detecting checksum, Cyclic Redundancy Check (CRC), and acknowledgement responses (ACK) to ensure transmitted packets are correctly received. A device will retransmit a packet in the subsequent connection event if the acknowledgement from the other device is not correctly received. BLE performs unlimited retries if there are repeated failures but will eventually terminate the connection between the two devices if no packet is correctly received after a set amount of time. Data whitening applies a filter to the packet payload and CRC immediately before transmission, scrambling bits to produce a more uniform distribution; the term derives from white noise’s flat spectrum. At the receiver, the transformation is inverted prior to CRC verification using the same whitening key. A whitened bit stream reduces long runs of identical bits, lowering susceptibility to environmental distortion (Tipparaju et al. 2021), yields a more even spectral spread favorable for regulatory compliance, and improves clock recovery and synchronization by increasing transition density.

Aside from those incorporated in BLE, other strategies exist to overcome data loss. An adaptive wireless communication strategy presented in (Aghagolzadeh and Oweiss 2011) seeks to optimize data transmission depending on the movement of the implanted subject. More movement induces a higher likelihood of error in transmissions. This strategy reduces radio on-air time of individual transmissions by decreasing the amount of data sent at a time during periods of greater movement and increasing them during periods of less movement. This would be considered a prevention strategy. Another loss prevention scheme for real-time audio streaming over wireless IP networks was presented in (Huang et al. 2006). Network congestion was estimated by measuring Round Trip Time (RTT), the time it takes for a packet to travel between the host and a gateway. This estimate determines whether a higher or lower bandwidth transmission scheme is appropriate. A lower bandwidth transmission is one that transmits data with a lower sampling rate, allowing the system to maintain real-time transmission at the cost of its quality. The rationale for such a scheme is that for the given application, lower quality data is preferred over periods of complete stalling, the latter occurring during a sufficient congestion event if lower bandwidth is not utilized.

A mitigation strategy was presented in (Claypool 2003) to address packet loss due to network congestion in the context of video streaming. Data packets were interleaved prior to transmission to spread out losses. This trades complete time gaps in video streams for lower quality periods. Such techniques can also be used in neural data streaming because lower quality data for a longer period is preferable over a complete loss of data for a shorter period.

A recovery strategy named Periodic Estimation of Lost Packets (PELP) was presented in (Dastin- Rijn et al. 2022). PELP addresses the issue of inaccurate time-domain reconstruction of neural data following packet loss. A common practice for determining the number of samples lost following packet losses is the utilization of time stamps within packets. Assuming a fixed transmission rate and packet size, one can deduce the number of samples lost by finding the difference between time stamps. While this method works for relatively short timespans, over longer timespans, inaccuracy eventually creeps in due to drifts in the timing elements that produce these time stamps. This proves to be an issue for post-processing, since filters applied to misaligned data can yield unwanted artifacts. PELP aims to solve this by utilizing stimulation artifacts for precise timing. While shown to be highly effective, this strategy is reliant on continuous stimulation with static parameters during recording. As such, it isn’t broadly applicable, as many devices window their stimulation and recording, often to avoid stimulation artifacts. Moreover, this approach would not be compatible with closed-loop stimulation, which can make use of dynamic stimulation parameters.

System and protocol

Testing platform architecture

The radio protocol introduced here was developed for an implantable neuromodulation platform capable of sensing, real-time telemetry, and electrical stimulation. It was designed specifically for use in in mice. With a volume of ~ 1.5 cm³, this system is a smaller version of the platform previously described in (Wright et al. 2022), and was designed using only readily available, off-the-shelf components to facilitate easy reproduction and adoption by researchers. As such, there are no custom integrated circuits, and the radio protocol is implemented using the popular nRF52840 architecture available from Nordic Semiconductor. The nRF52840 is capable of BLE, ESB, and custom radio implementations allowing for comparison of different approaches. The particular system in package (SIP), HongJia 840, was chosen to facilitate the high degree of miniaturization required for this application. At only 6.2 × 7 mm, the SIP incorporates an antenna and other peripherals required for the implant system.

The platform, shown in Fig. 1, incorporates both recording and stimulation capabilities, with one channel dedicated to each, and a third channel that may be switched between either functionality. The platform supports arbitrary waveform stimulation up to 10 kHz, calculation of derived values such as heart rate, and on-board responsive closed-loop functionality. The implant is powered by an 18.5 mAh lithium polymer battery (Wyon, Switzerland) and can be wirelessly charged using a resonant coupling system. The high component density and tight integration, combined with a four-layer stackup (two PCBs, ferrite, and coil) and polymer shell encapsulation, present significant challenges for telemetry by limiting antenna radiation efficiency. The preliminary screening presented in the Methods section was performed with this implant to establish typical operational conditions, with only the recording and telemetry functions enabled.

Fig. 1.

Block diagram of implant system architecture. It uses an internal chip antenna within the SIP as shown by the arrow in the top right photo of the unhoused system

Overview of radio operation

Digital wireless links convey information as discrete packets to share a congested medium efficiently. Packetization bounds each transmission’s on-air time, reduces overlap with other devices compared to continuous streams, and enables framing, addressing, and integrity checks. In wireless channels where many transceivers coexist on the same frequency and antennas radiate broadly, shorter packets statistically reduce collision probability and improve robustness under interference.

A packet comprises a preamble, an address, an optional length field, a payload, and a CRC checksum. The preamble is a known bit pattern that marks the start of a packet and assists the receiver in bit and symbol synchronization. The address identifies the communicating endpoint and supports filtering at the receiver. When payload size is dynamic, a length field delimits reception; it is omitted when both ends use a fixed payload size to minimize overhead and parsing latency. The payload carries user data, here neural samples alongside metadata such as packet number and sample indices, as shown in Fig. 2, which are reconstructed at the receiver. The CRC provides an integrity check computed over the payload using a shared polynomial; the transmitter appends the CRC, and the receiver accepts the packet only if its computed CRC matches. A single bit error in the payload or CRC suffices to invalidate the packet. Because CRC length is shorter than typical payloads, certain types of errors can in principle produce false accepts or rejects, though the probability is lower with longer CRCs.

Fig. 2.

On-air representation of a radio packet for both NRTP and BLE. Information is transmitted and received in the order shown from left to right. The content in the payload is user-defined and is specific to the experiment discussed in this paper, namely the payload transmitted by the implant during each trial

Bit errors and packet misses originate primarily during the packet’s on-air time due to interference and attenuation. Interference from background noise, colliding transmissions on the same channel, or inter-symbol interference (ISI) arising from multipath propagation and limited bandwidth can flip bits and corrupt frames. Attenuation from propagation loss and antenna inefficiencies lowers received signal to noise ratio (SNR); once below the receiver sensitivity threshold, demodulation fails and packets are not detected. Invalid packets are typically discarded by the receiver as if never received.

Data loss has two distinct forms. Source-side loss (failure to transmit) occurs when acquisition outpaces the telemetry link, finite buffers fill, and new samples are not enqueued. This is a critical failure mode in real-time implants with limited memory and tight latency budgets; mitigation requires sustaining telemetry throughput above the acquisition rate, minimizing processing delay, and sizing buffers to absorb short-term variability. Channel-side loss (failure to receive) arises when packets are corrupted or missed due to interference and attenuation; packets with invalid CRCs are discarded, and signals below sensitivity thresholds blend with noise and are not decoded.

Operational choices shape these loss characteristics. Longer payloads increase efficiency but spend more time on-air, raising collision and interference probability; shorter payloads reduce packet error rate at the expense of packet header overhead. Static payload lengths simplify parsing and reduce on-air header bytes, whereas dynamic lengths add flexibility and complexity. Immediate acknowledgements with bounded retransmissions reduce latency and buffer backlogs relative to deferred or unlimited retries, trading some delivery probability for sustained throughput. Channel selection outside crowded frequencies lowers collision probability, and techniques such as whitening or forward error correction, if enabled, can further reduce packet error rate with added power and latency costs.

The NRTP protocol

NRTP is a half-duplex 2.4 GHz protocol between an implant (primary transmitter) and a hub (primary receiver). The hub acquires an implant’s address and radio frequency (RF) channel to “tune in”; otherwise, it listens on a specified channel for packets bearing the target address. Each implant uses a single RF channel for both TX and receiver (RX), accepts requests from any hub, and publicly broadcasts requested data on its channel. Idle implants periodically emit short heartbeat packets for discovery. Multiple implants may share a channel, but a hub demultiplexes by address and listens to one implant at a time.

A successful transaction consists of an implant data packet immediately followed by a hub ACK packet. Radios on both sides alternate roles (TX/RX) per packet. Bidirectional data is supported by attaching a payload to the hub’s ACK. NRTP uses bounded retransmissions: if the implant does not receive a valid ACK, it retransmits the same packet up to a user-defined maximum number of attempts; packets that still lack a valid ACK are counted as dropped, and the implant moves on to the next packet. The hub may similarly retry ACK-attached payloads. The hub does not solicit retransmission when it receives an invalid packet; it continues listening for the next packet. If the hub received the data but its ACK failed, the implant’s retransmission is detected as a duplicate and discarded. Supplementary Fig. 1 illustrates different transaction scenarios.

NRTP employs a single packet buffer in static memory. The application overwrites this packet with new data at each timer-driven transmission interval and sends it immediately, rather than queueing multiple distinct packets as in BLE. Payload length is static and identical at both ends for both directions. The implant transmits at a fixed interval set to meet the required data rate, but retransmissions occur immediately after a failed ACK and do not wait for the next interval. In contrast, BLE may send multiple packets per connection event and defers retransmissions to subsequent connection intervals. When idle (no active request), NRTP maximizes the transmission interval to reduce power while maintaining infrequent heartbeats for discoverability. To initiate a stream, the hub embeds a request in its ACK payload—either in response to an active transmission or to a heartbeat from an idle implant. The hub must first receive a packet from the implant before issuing a request or retrying an unacknowledged request.

NRTP variants for evaluation

We evaluated three mitigation techniques within the protocol—overlapping, interleaving, and retries—individually and in combination, and compared their performance to BLE. Each technique or combination defines an NRTP variant. We also assessed two payload lengths for both protocols to characterize the effect on packet-loss likelihood and power cost.

Interleaving rearranges data non-contiguously across packets to mitigate burst errors. Interleaving can operate at multiple levels (bit, symbol, sample, packet); here, sample-level interleaving was applied prior to packetization. Using a block interleaver modeled as an m × 2 matrix (m = samples per packet), samples are written row-wise and read column-wise, yielding packets that contain either all even-indexed or all odd-indexed samples in time (Fig. 3A). This spreads a given time span across two packets so that loss of one packet produces a segment of half-sampling-rate (HSR) data rather than a contiguous gap, whereas loss of both packets in an even–odd pair yields 100% data loss for that span. If two consecutive losses are not an even–odd pair, the affected span again appears at HSR. Higher-order interleaving (spanning > 2 packets) can tolerate longer burst losses but reduces the preserved sampling rate; 2-packet interleaving was selected because brief HSR segments are acceptable at a 20 kS/s target, whereas < ½ rate was not. Supplementary Figs. 2–3 illustrate the effects of scattered and consecutive packet loss on the data stream when interleaving is employed.

Fig. 3.

Visual representation of the different strategies that were evaluated. Each of the 3 diagrams shows how the series of consecutive samples are used to construct radio packets for transmission. A Interleaving of samples between packets. B Overlapping samples between neighboring packets. C Combination of interleaving and overlapping

Overlapping introduces static redundancy by repeating a portion of the previous packet within the current packet alongside new data, creating overlap between adjacent packets (Fig. 3B). This increases resilience to scattered (nonconsecutive) packet loss but offers no benefit against consecutive packet loss. Overlapping trades bandwidth for robustness and adds preprocessing and memory overhead. Here, 50% overlap was used (half of each packet redundantly mirrors the prior packet). Supplementary Figs. 4–5 illustrate the effects of scattered and consecutive packet loss on the data stream when overlapping is employed.

Retries provide dynamic redundancy by retransmitting a packet when an ACK is not received. Retries increase robustness but reduce net theoretical maximum throughput because airtime is spent on repeats rather than new data; excessive retries can therefore cause source-side loss when acquisition outpaces transfer. In this evaluation, the maximum amount of retries was limited to one (two total transmit attempts per packet).

Combining interleaving followed by overlapping, (Fig. 3C), yields robustness greater than either alone. This ordering increases tolerance to consecutive losses (up to five consecutive packets compensable under our 50% overlap), shortens the duration of HSR segments compared to interleaving alone for the same loss pattern, and removes the constraint that two lost packets forming an even–odd pair must result in total loss. The trade-off is reduced effective bandwidth due to redundant transmission. Supplementary Figs. 6–7 illustrate these behaviors.

Methods

Evaluation metric

A quantitative metric was defined using Eq. 1 to compare BLE to NRTP and to evaluate NRTP mitigation techniques individually and in combination, prioritizing minimal loss (defined as periods of time for which there are no samples), then data quality, and finally average current draw; higher scores indicate better performance. The score increases with the percentage of received data, weighting full-sampling-rate (FSR) data more than HSR data. HSR data is valued at half the weight of FSR data because half as many samples are received. It is important to note that depending on the experimental paradigm, HSR data may not preserve adequate information (e.g. neural spikes sampled at 20 kS/s), potentially making it less useful. However, the 0.5 weight was chosen for HSR data based on quantity of samples received relative to FSR data because the objective was to determine which protocol provided the most reliable data transmission.

The score drops sharply with increasing data loss, due to the inversely proportional relationship; to avoid divergence when loss is zero, the loss term is floored at 0.001. The score also decreases with higher current draw, but not as sharply as with loss because the relative changes in average current are much lower than those of the loss term which can range from 0 to 1. FSR, HSR, and Loss are computed relative to the total expected samples (1,000,000 for 50 s at 20 kS/s). This hierarchy reflects the design objective and enables ranking of variants.

1

NRTP is designed to address data loss when using standard 2.4 GHz radio hardware in an implant setting, as such, it was evaluated for an example bandwidth of 0.320 Mb/s using an uncompressed, single channel, 16-bit, 20 kS/s data stream. Such a bandwidth can be utilized to transmit multiple channels recorded at lower resolution and or sampling rates. Similarly, depending on use-case, data compression schemes and digital signal processing techniques can be used to further expand the amount of information transferred, and on-board storage can be employed to add additional flexibility when real-time operation is not a requirement. The example bandwidth used for protocol evaluation was chosen given the constraints of the implantable platform such as power consumption, ADC sampling rates, available compute resources, and reasonable expected performance of both NRTP and BLE under the testing conditions. Since the average current draw captures the time that the radio spends on/off, a higher bandwidth will have higher average current and thus lower score than a lower bandwidth (e.g., with the same FSR and HSR values in the numerator, a 1 Mb/s bandwidth will have a lower score than 0.320 Mb/s). Therefore, the evaluation metric is not generalizable and can only evaluate telemetry protocols operating at equal bandwidths for their relative scores to be comparable.

Firmware implementation

BLE was tested across 14 configurations combining two payload lengths (244 bytes, the BLE maximum excluding three opcode bytes; 122 bytes) and seven connection intervals (7.5, 8.75, 10, 11.25, 12.5, 15, 20 ms). For both payloads, 16 bytes encoded packet metadata (e.g., packet number, count of new samples, first sample acquisition index). This yields up to 114 samples per 244-byte packet and 53 per 122-byte packet (2 bytes/sample). The implant advertised until connected, then acquired samples and queued packets for transmission (up to four queued). Radio operations and timing were managed by Nordic Semiconductor’s SoftDevice (S140 v7.2.0, nRF5 SDK v17.1.0).

NRTP used the same payload lengths but reserved 18 bytes for metadata, supporting up to 113 samples (244-byte payload) and 52 samples (122-byte payload). The 16 NRTP variants comprised all combinations of the two payload lengths and three mitigation features (retries, interleaving, overlapping), each enabled or disabled (2⁴ possibilities). Without a BLE-style connection interval, NRTP’s packet transmission interval was set based on payload length and overlap setting, and slightly faster than the data acquisition rate to absorb retry latency and prevent buffer overflow. With overlap disabled, intervals were 5 ms (200 Hz) for 244-byte payloads and 2.5 ms (400 Hz) for 122-byte payloads; with 50% overlap, transmit rate was doubled to maintain necessary throughput because the new-sample capacity per packet was halved, resulting in higher power consumption due to increased radio activity. Transmission was interrupt-driven via a hardware timer; radio actions were likewise interrupt-driven. NRTP runs without a SoftDevice, providing full control subject to radio timing constraints shown in Supplementary Fig. 8; the RX window is ~ 5% longer than the packet’s on-air time to allow tolerance for delays. NRTP can operate on any of 83 channels each 1 MHz wide centered from 2.401 to 2.483 GHz (within the unlicensed 2.4 GHz ISM band limits). Tests used channel 82 (2.482 GHz) to avoid the crowded Bluetooth Classic/BLE (upper limit at 2.480 GHz) and Wi-Fi (upper limit predominantly at 2.473 GHz) spectrum and reduce collisions. The state machine diagram of Supplementary Fig. 9 illustrates the implant behavior for NRTP.

Hub firmware for NRTP and BLE was identical in structure: an interrupt handler processed each received packet, serialized packet data and measured RSSI, and wrote them to alternating USB transfer buffers (double buffering to avoid contention). PC software reconstructed the stream and updated throughput and loss statistics. The state machine diagram of Supplementary Fig. 10 illustrates the hub firmware.

For interleaving in NRTP, the implant maintained two sample buffers (even and odd indices in time). At each acquisition interval, samples alternated between buffers; at each transmission interval, packets extracted exclusively from the buffer selected by packet parity (even packet → even samples; odd packet → odd samples). For overlapping in NRTP, each packet wrote new samples into half the payload while preserving the other half as redundancy from the prior packet; the overwritten half alternated by packet parity (odd packets overwrite the first half, even packets the second). Packets included the count and starting acquisition index for both redundant and new samples to enable correct reconstruction at the PC.

Preliminary screening

We evaluated all 16 NRTP variants against 14 BLE configurations across a range of RSSI values using the implant and an nRF52840DK as the hub. RSSI served as the received‑signal attenuation metric. The objectives were to establish typical operational conditions of the implant, and to down‑select to the three best NRTP variants and three best BLE configurations for detailed testing under tighter control, reducing overall test time while preserving real‑world conditions. The implant was mounted on a programming fixture; a Keithley 2450 source meter supplied 3.8 V and measured current, and a J‑Link programmed firmware per trial configuration. TX power was fixed at + 8 dBm; RSSI was varied by changing separation and orientation. Tests were conducted in an open environment with people, obstacles, and coexisting Wi‑Fi/BLE devices. Overlapping variants of NRTP scored ~ 50% lower than non‑overlapping variants due to their doubled packet rate and corresponding greater current draw—a ~ 75% increase compared to non-overlapping variants. The greater current draw is a consequence of the increased radio on-time associated with the higher packet rate. This significant current increase can be impractical for millimeter‑scale implants with strict power budgets. For example, with the 18.5 mAh battery capacity of the test implant, the higher discharge rate would reduce runtime by 42%, accelerate battery degradation, and increase implant temperature. Overlapping would have been the highest scoring option if current were excluded from the score, as it does offer a slight increase in successful data transmission at very low RSSI due to its redundancy as shown in Supplementary Fig. 11-C. For BLE, 122‑byte payloads scored about half as well as 244‑byte payloads did due to insufficient throughput, and connection intervals > 11.25 ms underperformed shorter intervals. Supplementary Fig. 11 shows results from the preliminary screening.

Main testing: setup and procedure

We evaluated the top three NRTP variants and BLE configurations with fixed geometry to stabilize environmental interference across RSSI sweeps. As illustrated in Fig. 4, two nRF52840DK boards were used (implant and hub), spaced 1.22 m apart on a level surface with co-aligned, vertically oriented antennas. Both radios operated at a constant − 8 dBm TX power to keep current draw consistent. RSSI was varied by adding attenuation on the implant side: the implant board used an external antenna via a MXHS83QE3000 probe (disconnecting the on-board antenna) and one or more attenuators (Mini-Circuits 15,542 VAT-A) in series; the hub used its on-board PCB antenna. Current was measured with a Keithley 2450 in series with the microcontroller’s current sense pins, averaging 500 readings per each ~ 50 s trial. Both boards were USB-powered at 5 V, where the microcontroller’s voltage measured 2.96 V.

Fig. 4.

Illustration of the experimental setup. The implant and hub devices are connected to the PC as USB peripherals, as well as the Keithley 2450 Sourcemeter for current measurement

Each trial acquired 1,000,000 16-bit samples at a target 20 kS/s and attempted to transmit as many as possible in real-time. Samples recorded a 16-bit hardware timer value via an interrupt-driven routine (i.e., timestamps rather than time-dependent signals). Samples were buffered on the implant (capacity 10,000 samples) until transmission. If the buffer filled due to inadequate transmission rate, new samples were not saved until the buffer emptied; these unsaved samples were never transmitted and were counted as lost time. The state machine diagram of Supplementary Fig. 9 illustrates sampling behavior. Packets included metadata (e.g., packet number, the count of contained samples, and the acquisition index of the first sample) to enable correct ordering at the PC; samples within a packet were consecutive as shown in Fig. 2.

A Windows batch script automated configuration cycling, re-flashing both devices with individual trial configuration-specific firmware and measuring implant current during each trial. Trials at a fixed attenuation formed a session, yielding the same average RSSI within that session. NRTP sessions and BLE sessions were run at the same nominal attenuation but produced slightly different average RSSI due to RF channel hopping in BLE. Multiple sessions at varied attenuation levels produced a wide RSSI sweep. Per-trial results and statistics were logged for analysis.

Post-trial processing on the PC sorted received samples by acquisition index and computed differences between adjacent samples. With interleaving disabled, differences > 1 indicated data loss; with interleaving enabled, differences > 2 indicated data loss and differences = 2 indicated a single missing sample (classified as HSR data), while differences = 1 indicated gapless FSR data. The PC also inspected adjacent packet numbers to detect dropped packets; up to 255 consecutive drops were detectable given a 1-byte packet counter. Non-transmitted loss was inferred when acquisition indices advanced more than expected without any dropped packet. Note that FSR data and HSR data were defined by sample indices. Interrupt-driven delays reduced the effective sampling rate and introduced gaps (NRTP delays spanned single-digit microseconds up to 150 μs; BLE up to 500 μs), but the scoring assumed constant inter-sample time intervals and thus relied solely on acquisition indices.

Results

Data from comparison tests of the top scoring BLE configurations and NRTP variants are presented in Fig. 5. All data is plotted against RSSI which is a measurement of received power. With respect to score (Fig. 5A) NRTP performed better than BLE over a wider range of RSSI. All NRTP variants had greater area under the curve than any BLE configuration, indicating superior performance at lower received signal strength. In terms of maximum score, NRTP-244-RI (packet length 244, retries and interleaving) was nearly identical to NRTP-244-R (packet length 244, only retries), and NRTP-122-RI (packet length 122, retries and interleaving) ranked third. NRTP-122-RI’s lower overall score was driven by higher current draw (Fig. 5B). NRTP-244-R’s score declined earlier than other NRTP variants due to higher data loss than the interleaved variants (Fig. 5E, data loss).

Fig. 5.

Results of all trials with average best fit curves. Markers represent individual trial results within each session. A Scores vs RSSI. B Average Current vs RSSI. C Percentage of FSR Data Received vs RSSI., D Percentage of HSR Data Received vs RSSI. E Percentage of Data Loss vs RSSI with logarithmic scale (inset plot with linear scale)

At high RSSI, above –55 dBm, two BLE configurations (packet length 244, 7.5 ms and 10 ms connection intervals) initially achieved higher average scores (best fit lines) than the NRTP variants, and the third BLE configuration (packet length 244 and 11.25 ms connection interval) did so above − 45 dBm due to lower average current consumption (Fig. 5B). However, BLE showed greater trial to trial variability early in the RSSI range, and scores began declining below − 55 dBm for the best performing BLE configuration, BLE-244–7.5, whereas NRTP scores remained high until − 75 dBm.

Current consumption (Fig. 5B) was lower for BLE compared to NRTP. This behavior was maintained across the RSSI sweeps. At poor RSSI, starting at approximately –75 dBm, current consumption for all BLE configurations declined, while it increased for all NRTP variants. However, the difference between BLE (~ 6.8 mA) and NRTP (~ 7.1 mA) average current consumption at high RSSI levels was minimal except for NRTP-122-RI (~ 7.5 mA) which had higher current consumption in all testing conditions.

When considering data quality, NRTP outperformed BLE in all cases. FSR data (Fig. 5C) showed significant variability in BLE performance starting at –65 dBm, whereas NRTP had no variability at all until past –75 dBm. Amongst the BLE configurations, BLE-244–7.5 maintained FSR better than others, but only by a slim margin of RSSI. Similarly, NRTP-122-RI maintained FSR better than other variants, but again by only a few dBm of RSSI. HSR data (Fig. 5D) showed that NRTP variants with interleaving allowed half-sampled-data to replace otherwise lost data when operating at poor RSSI, beyond –75 dBm. NRTP-244-RI achieved higher HSR data than NRTP-122-RI but the former decreased earlier. Lost data (Fig. 5E) was calculated to consider only gaps in time when no data was present, which only occurred when neither FSR nor HSR data were present. Here NRTP, including variants without interleaving, again outperformed BLE in all conditions, maintaining zero data loss until near –75 dBm. The two NRTP variants with interleaving outperformed the one without since HSR data was able to contribute for cases where data would have been lost. Amongst the two NRTP variants with interleaving, NRTP-122-RI showed the least data loss overall.

Discussion

NRTP’s superior robustness stems from design choices that prioritize sustained throughput under attenuation and interference: static framing, immediate bounded retries, and operation on channels with minimal congestion. Immediate bounded retries avoid the throughput collapse seen in BLE’s deferred unlimited retry model, which reduces packet loss but increases source-side data loss (non-transmitted sample loss) whenever acquisition outpaces effective transfer rates. Supplementary Fig. 12 shows BLE has zero dropped packets for the entire RSSI sweep despite having considerably high data loss. Bottlenecking by the SoftDevice’s packet queue produce delay induced losses that worsen with longer connection intervals, demonstrating BLE’s optimization for lower bandwidth to save power via less radio activity.

NRTP variants with interleaving further shift loss from contiguous gaps to HSR segments, preserving information content under sparse drops and some burst losses; its benefit appears in delayed score decline at lower RSSI. As consecutive drops grow, contiguous losses reemerge and HSR data proportion falls. Larger payloads yield broader HSR data spans per dropped packet, increasing HSR data percentage but also raising drop probability due to longer on-air exposure—explaining the earlier HSR data decline for 244-byte payloads compared to 122-byte payloads. Supplementary Fig. 12 shows NRTP-122-RI has the fewest dropped packets among NRTP variants. But shorter payloads require a higher packet rate to meet the transfer demand, raising the average current, which the metric penalizes. In terms of power, interleaving showed negligible cost.

BLE’s higher scores at very high RSSI are explained by lower current consumption, but its earlier score decline reflects sensitivity to coexistence of other devices in the environment. Figure 5E shows BLE’s nonzero data loss at higher RSSI relative to NRTP which caused its earlier score decline, while NRTP had zero data loss until very low RSSI. Although a prevention strategy, channel hopping within the crowded BLE band doesn’t eliminate collision risk. The observation that BLE exhibits data loss even at higher RSSI values, relative to NRTP, is consistent with packet collisions from coexisting devices; NRTP’s ability to operate on a channel not utilized by BLE while still remaining within the unlicensed 2.4 GHz ISM band largely mitigates this, even with its lack of channel hopping. While BLE is generally designed for low bandwidth intermittent telemetry (e.g. sporadic sensor data) in crowded network environments, NRTP is designed for higher bandwidth continuous telemetry (e.g. streaming) in controlled environments. With the appropriate license, NRTP can operate in channels which BLE and Wi-Fi cannot, further isolating it from potential interferences. However, even when operating at the edges of the unlicensed ISM band, NRTP is still subject to interference from Bluetooth and Wi-Fi sidelobes, and direct interference from Zigbee radios and consumer appliances such as microwave ovens. As such, standard procedures involve scanning for the clearest channel to use and removing potential sources of interference.

Inter-session variability for score remained low for NRTP until very low RSSI and was higher for BLE across the range. This suggests NRTP’s immediate retry and static interval behavior produces steadier throughput, whereas BLE’s connection-event scheduling and coexistence with other devices introduces trial-to-trial fluctuations. At poor RSSI, both systems show increased variability due to heightened interference and sensitivity to small geometry changes.

NRTP’s data loss begins significantly later than BLE’s across the RSSI sweeps. When looking at the first occurrence of any data loss, the best performing NRTP variant exceeds the best performing BLE configuration by a link margin of 23 dB; if data loss of 0.5% is considered as a threshold, the margin drops to 11 dB. There are significant practical implications of this result when designing resource constrained implantable systems, even when considering a conservative nominal value of 10 dB improvement. Given that the free space path loss in air is proportional to 1/R², NRTP will operate with similar performance to BLE at ~ 3.2 × the distance between implant and hub, using the same radio output power. With regard to RF attenuation by tissue, the same 10 dB performance improvement corresponds to ~ 2.5 cm (Christoe et al. 2021) deeper implantation depth for the same level of performance. Alternatively, with distance being the same, NRTP will operate with similar performance to BLE at lower radio output power, significantly reducing system power consumption from increased link margin. The total system power consumption dropped by ~ 33% when decreasing TX power from the maximum value of + 8 dBm down to 0 dBm for both BLE and NRTP. Supplementary Fig. 11-B shows the average current draw using + 8 dBm TX power during the preliminary screening. Though BLE typically has lower power consumption than NRTP at a given TX power setting, the greater link margin provided by NRTP allows operation at a lower TX power while still providing the same level of performance.

NRTP’s increased reliable operation down to −75 dBm makes it more suitable than BLE for use in implantable devices in freely moving animals; where RF connections must tolerate attenuation through device packaging, tissue, and air. The implant’s sub-optimal antenna performance is further compounded by orientation changes, limiting the effective range. A typical operating distance between such an implant and hub is 1 m, which introduces a path loss through air of approximately 40 dB for 2.4 GHz signals. To give a quantitative example, when using −8 dBm TX power at 1.22 m from the hub with co-aligned antennas, the implant device used in the preliminary screening produced an RSSI of −72 dBm, whereas the nRF52840DK with a more optimal antenna produced an RSSI of −42 dBm. Rotating the implant was found to further decrease RSSI by ~ 10 dB. Given these operating conditions, NRTP’s greater link margin can be used as a buffer to overcome the non-idealities in real-world use cases.

Power behavior aligns with protocol mechanics: BLE current consumption decreases at poor RSSI due to fewer transmissions per second when repeated corrupt packets cause the protocol to wait until the next connection events and shift channels; NRTP current consumption increases at poor RSSI because immediate retries raise the transmission rate. These dynamics reinforce the trade-off between robustness, throughput, and power that the metric captures: configurations that minimize true data gaps and maintain usable segments of data across a wider RSSI range, without excessive current, achieve the highest scores.

When comparing these findings to prior works listed in Table 1, specifically in the last four rows which show comparable radio systems designed for implantation, NRTP enables the highest data rate for fully implantable systems using 2.4 GHz radio hardware. As tested, the best NRTP variant maintains the target throughput of 0.320 Mb/s and its radio operation has a power draw of 21.0 mW under most operating conditions (running all system functions simultaneously increases power draw to 39.1 mW). As tested, the power draw per data rate is 65.6 mW/(Mb/s), the lowest amongst the fully implantable 2.4 GHz radio systems in Table 1, and similar to the non-implantable 2.4 GHz radio systems (head-stages) which benefit from more advantageous operating conditions. Prior works typically report limited radio results—either omitting empirical performance or citing protocol-level theoretical maxima (Oh et al. 2024). In contrast to these narrow evaluations, we systematically characterized radio behavior across operating conditions via controlled RSSI sweeps, quantitatively comparing NRTP and BLE over a broad range. To our knowledge, no published work to date presents an RSSI-sweep-based, end-to-end throughput and loss analysis of implant telemetry at this level of detail. In general it seems that BLE has not been successfully used to transmit high sampling rate neural data from fully implanted devices, with one study noting that though the theoretical limit is 2 Mb/s, the practical limit of BLE is closer to 0.120 Mb/s (Oh et al. 2024).

Table 1.

Comparison of other published work regarding wireless miniature scale neuromodulation devices

Reference	Wireless Band & Protocol	Format/Animal	Data Rate (Mb/s)	Power Draw (mW)	(mW/Mb/s)
Idogawa et al. ( 2021)	2.4 GHz BLE	Headstage/mouse	0.12	28.6	238.3
Wang et al. (2022)	2.4 GHz BLE	Headstage/rat	0.48	10*	20.8
Su et al. (2016)	2.4 GHz ESB	Headstage/NHP	1	~ 57	57
Gagnon-Turcotte et al. ( 2017)	2.4 GHz ESB	Headstage/mouse	1.4	~ 120	85.7
Roy et al. (2011)	5.8 GHz FM (analog)	Headstage/NHP	N/A	55	N/A
Harrison et al. ( 2011)	900 MHz FSK	Headstage/insect	0.24	1	4.2
Lee et al. (2016)	915 MHz FSK	Head-stage/rat	2.76	51.4	18.6
Yang et al. (2014)	13.56 MHz inductive	Implantable/rat	8.3	N/A	N/A
Kanchwala et al. (2018)	6.5 GHz UWB-IR	Implantable/rat	0.64	45.5	71.1
Kassiri et al. (2016)	UWB1/UWB2/916 MHz FSK	Implantable/mouse	N/A	N/A	N/A
Liu et al. (2016)	13.56 MHz inductive, BLE	Implantable/rat	3.2	N/A	N/A
Pederson et al. ( 2019)	2.4 GHz BLE	Implantable/rat	0.05	50	1000
Liu et al. (2022)	2.4 GHz BLE	Implantable/PNS	< 0.1152	N/A	N/A
Ouyang et al. (2023)	2.4 GHz BLE	Implantable/mouse	.0092	< 35	3804.3
This work	2.4 GHz NRTP	Implantable/mouse	0.320	21.0/39.1**	65.6/122.2

*It is reported that the 3.7 V 30 mAh battery was depleted in 1.5 h, yielding 74 mW system power draw

**Telemetry only/All system components running simultaneously

Future directions

In future studies, we aim to explore adding dynamic payload lengths to shrink ACK packets and increase theoretical throughput, preventing buffer overflow by temporarily accelerating transfer when delays occur. To assess performance in more lossy conditions, we will introduce controlled interference using a secondary transmitter on the same RF channel as a means of emulating packet collisions and multi‑implant coexistence. Subsequently, we aim to evaluate channel hopping and error correction to improve robustness under these conditions, as well as investigate the use of data whitening. Finally, we will utilize tissue phantoms to replicate attenuation and environmental interference, as well as conduct in‑vivo tests using the implant to quantify how motion and placement affect interference.

Conclusion56

NRTP is a neural telemetry protocol for miniaturized implants that addresses a key unmet need: reliable, real-time data streaming from millimeter-scale, fully implantable devices without custom electronics or ASICs. It is designed for implementation on commercially available 2.4 GHz radio hardware. We characterized its performance against BLE across diverse operating conditions and found NRTP to be more robust under weak links and environmental interference—even before enabling loss mitigation features. BLE underperformed due to insufficient effective throughput under interference: unlimited retries increased per packet delivery but delayed transmission relative to acquisition, increasing net data loss. In contrast, NRTP’s immediate, bounded retries sustained throughput and minimized true gaps in the data stream. Across NRTP variants, all three mitigation strategies reduced data loss; our evaluation metric quantified their tradeoffs: overlapping improved robustness but was unsuitable for our implant platform because of significantly reduced battery runtime due to the higher power draw arising from the doubled packet rate, whereas single retries and sample-level interleaving were effective with negligible power overhead. The resulting link margin gains compared to BLE yield practical benefits: expanded coverage range, greater tolerable implant depth in tissue, and the ability to reduce transmit power for equivalent performance,extending battery life and easing thermal constraints. Collectively, NRTP’s robustness, power efficiency, and adoptability on commodity hardware lower the barrier to deploying chronic, closed-loop telemetry in small animal implants.

Abbreviations

ACK

Acknowledgment

ASIC

Application Specific Integrated Circuit

BLE

Bluetooth Low Energy

CRC

Cyclic Redundancy Check

dB

Decibels

dBm

Decibels relative to 1 milliwatt

DK

Development Kit

ESB

Enhanced ShockBurst (Nordic Semiconductor proprietary protocol)

FM

Frequency Modulation

FSK

Frequency-Shift Keying

FSR

Full-Sampling-Rate (data)

GHz

Gigahertz (e.g., 2.4 GHz, 5.8 GHz, 60 GHz)

HSR

Half-Sampling-Rate (data)

ISM

Industrial, Scientific, and Medical (radio band)

kS/s

Kilo-samples per second

Mb/s

Megabits per second

MHz

Megahertz (e.g., 900 MHz, 915 MHz, 434 MHz, 13.56 MHz)

mW

Milliwatts

μs

Microseconds

NHP

Non-Human Primate

NRTP

Neural Real-Time Telemetry Protocol

PELP

Periodic Estimation of Lost Packets

RF

Radio Frequency

RSSI

Received Signal Strength Indicator

RTT

Round Trip Time

RX

Receive / Receiver

SIP

System-in-Package

SNR

Signal-to-Noise Ratio

TX

Transmit / Transmitter

USB

Universal Serial Bus

UWB

Ultra-Wideband

WPT

Wireless Power Transfer

cm³

Cubic centimeters

mm

Millimeters

Authors’ contributions

M.E and M.R wrote the manuscript, prepared figures, and designed and performed experiments. J.W edited the manuscript and designed experiments. T.D.C wrote and edited the manuscript and designed experiments. All authors reviewed the manuscript.

Funding

This work was funded by the Feinstein Institutes for Medical Research.

Data availability

Data sets generated during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.