An Implant-to-Wearable IR-UWB Transmitter-Receiver Architecture and Layered Protocol for High-Density Brain-Computer Interfaces

G. D. Su; Y. Mo

eess.SPSep 1, 2026

An Implant-to-Wearable IR-UWB Transmitter-Receiver Architecture and Layered Protocol for High-Density Brain-Computer Interfaces

Authors: G. D. Su, Y. Mo

Organizations: Hangzhou Dianzi University · BroadLink Co., Ltd.

Abstract

High-rate neural telemetry requires an implant-towearable link whose circuit partition, waveform mapping, packet processing, and failure semantics are mutually consistent. This paper specifies a chip-oriented, one-way impulse-radio ultrawideband (IR-UWB) uplink comprising an implanted packet engine and free-running transmitter, a tissue-proxy channel, and a wearable mixed-signal receiver and digital baseband.

Figures & tables

Fig. 1: System-level application context for the implant-to-wearable IR-UWB transmitter–receiver architecture. Multiple high-density neural interfaces feed an implanted recording and telemetry SoC; the wearable receiver recovers the layered neural-data stream and forwards it to a host. The use case follows recent neural IR-UWB research [ 9 , 5 ] ; the artwork and technical partition are original.

Fig. 2: End-to-end chip-oriented partition of the one-way implant-to-wearable uplink. The implant RO, selected-phase edge-combining, pulse-generation, and PA organization is adopted from Lei et al. [ 9 ] ; the packet/lane interfaces, wearable receiver partition, decoder flow, and cross-layer service semantics are specified in this work. W1 3/5-cycle PWM is shown as the baseline programmable width profile. The tissue blocks denote a synthetic channel boundary, not an anatomical or measured propagation model.

Fig. 3: Normalized simulated TX power spectral density. The shaded 3.1–10.6-GHz region is frequency-allocation context only. The curves are not in dBm/MHz and do not establish FCC-mask compliance, radiated power, antenna efficiency, or a measured implant output spectrum.

Fig. 4: Wearable external-receiver partition. One envelope path supplies acquisition, timing, and Base-width evidence; one time-shared burst-I/Q path supplies optional differential-phase evidence. All templates and lane decisions are digital.

Fig. 5: Predictive NCO sampling window. Envelope evidence from an admitted observation corrects future windows; it does not reveal the transmitted event time to the runtime detector.

Fig. 6: Logical transmit modulation and receive demodulation flow. This original block diagram follows the paired modulator/demodulator presentation convention of [ 9 ] , while replacing its D16PPM hybrid modem with the present fixed-slot PWM Base lane (illustrated with the W1 3/5-cycle baseline), optional DBPSK Enhancement lane, predictive receive timing, soft LLRs, and explicit erasure/fallback path.

Fig. 7: Deterministic W1/M1 sample-domain modulation and demodulation example. Base bits select three- or five-cycle bursts; Enhancement bits update differential polarity; the synthetic receiver envelope and gated I/Q observations produce signed soft metrics. All amplitudes and metrics are normalized and no measured waveform is implied.

Fig. 8: Packet timeline and the independently verified Base and Enhancement lanes. Enhancement failure can produce labeled Base-only service but cannot be counted as strict complete delivery.

Fig. 9: WIDTH_TRAIN quality gating and conservative fallback. The frozen campaign reports gated and ungated online-template receivers separately, because rejecting difficult packets must not be mistaken for improved strict delivery.

Dimension	Primary settings
Width profile	W1 (3/5 cycles), W3 (4/7 cycles)
PHY mode	M0, M1
Nominal SNR	6, 9, 12, 15, 18 dB
Payload FEC	Conv. $K=7$ , rate $1/2$
Detector	Scalar, fixed, online, gated
Population	In-distribution and shifted hold-out

TABLE I: Primary Held-Out Evaluation Matrix

Fig. 10: Renderer convergence. The selected 256-sample/symbol rate removes cycle slips in the convergence experiment and reduces sample-derived timing RMSE to 146 ps; residual Base-width BER is therefore not attributed to renderer synchronization failure. DBPSK phase-metric convergence is not evaluated by this figure.

Fig. 11: Synthetic tissue-proxy population distributions. These are generated statistical proxies, not human or animal measurements.

Fig. 12: Strict Complete PSR of the ungated online-template receiver. Error bars are 95% hierarchical cluster-bootstrap intervals. Nominal SNR is an integrated simulation reference, not TX power, RSSI, sensitivity, or range, and cannot be converted to those quantities without measured, jointly calibrated hardware and channel data.

Condition	Header	Base	Enh.	Strict Complete [95% cluster CI]	Service	Goodput	Width BER
ID M0 W1	89.8	88.1	–	88.1 [85.9, 90.2]	88.1	42.6	4.46
ID M0 W3	78.2	75.4	–	75.4 [72.2, 78.5]	75.4	36.4	5.99
ID M1 W1	89.1	81.5	82.2	75.8 [70.0, 80.7]	81.5	47.1	5.50
ID M1 W3	79.0	70.2	75.6	68.3 [62.0, 73.7]	70.2	42.0	7.08
Shift M0 W1	69.4	66.1	–	66.1 [61.7, 70.5]	66.1	31.9	11.01
Shift M0 W3	56.8	53.0	–	53.0 [48.2, 57.7]	53.0	25.6	13.00

TABLE II: Online-Template Held-Out Results at Nominal 18 dB

Fig. 13: Layer reachability for the online-template receiver at nominal 18 dB. Service can exceed strict Complete only when a verified Base lane is delivered with Enhancement explicitly missing. The nominal SNR is the internal simulation reference of ( 7 ), not a real-world link-budget point.

Fig. 14: Aggregate soft-information evidence. Negative mismatched-LLR information scores indicate harmful confidence scaling; they are retained rather than clipped from the evidence.

Receiver	Ops/sym.	State B	PSR	Brier
Scalar-width soft LLR	18	32	21.7	0.173
Fixed dual template	280	512	16.2	0.224
Online dual template	320	1024	24.4	0.158
Quality-gated online	360	1152	21.6	0.174

TABLE III: Deployable Detector Complexity and Sweep-Aggregate Outcome

Mode	SNR (dB)	Variant	Complete PSR (%)	Goodput (Mbit/s)	Model energy (nJ)	Latency ( $\mu$ s)
M0	12	Baseline conv 1/2	0.0	0.00	2.088	3.974
M0	12	TBCC + fixed blind $\times$ 2	85.0	16.54	5.146	9.867
M0	12	Repeated conv + fixed blind $\times$ 2	92.2	11.90	7.880	14.881
M0	15	Baseline conv 1/2	85.4	41.27	2.088	3.973
M0	15	TBCC + fixed blind $\times$ 2	100.0	19.46	5.145	9.866
M0	15	Repeated conv + fixed blind $\times$ 2	100.0	12.90	7.879	14.879

TABLE IV: Robust reference sensitivity results at selected nominal SNR values.

Width	SNR	Enh. BER (%)		LLR $\rho$	Disagree. (%)
	(dB)	256	512	256/512	256/512
W1	12	9.52	9.15	0.793	12.88
W1	18	2.81	3.65	0.942	3.88
W3	12	8.19	7.33	0.867	8.90
W3	18	1.86	1.88	0.972	1.96

TABLE V: Paired DBPSK sampling-rate sensitivity.

Parameter group	Frozen value or rule
Campaign	Hash 593f2287c9b49ead ; root seed 2026082699; 40 cells; 2000 packets/cell; 4000 cluster-bootstrap replicates; sequential stopping disabled.
Calibration	Hash 48cdcc968e591308 ; 250 training and 250 validation packets per algorithm; one immutable artifact applied to both final populations.
Renderer and AFE	300-MHz symbol rate; 256 samples/symbol (76.8-Gsample/s numerical grid); 6-bit uniform I/Q quantization over $[-1,1]$ ; causal one-pole 750-MHz front end; AGC target 0.35, gain range 0.25–512, frozen after 64 symbols.
Acquisition and timing	Acquisition correlation $\geq 0.10$ ; refined SFD correlation $\geq 0.30$ ; event peak/noise $\geq 3$ ; search half-window 0.75 ns; $K_{p}=0.10$ , $K_{i}=0.002$ ; timing-error clamp $\pm 0.10T$ ; integration interval $[-0.30T,+0.62T]$ about the prediction.
Online width rows	64 resampled envelope-power values; subtract row 20th percentile and clip below zero; direct valid-class means ( $\lambda_{0}=\lambda_{1}=1$ ); no per-row amplitude fit or sample weights.
Width statistic	Shape-distance normalization by $\max\{2\,\operatorname{median}_{\mathcal{T}}\min_{b}d_{b},10^{-5}\}$ ; pooled scalar-variance floor $10^{-5}$ ; component clipping $\pm 24$ ; fusion $0.75s_{\mathrm{E}}+0.25s_{\mathrm{shape}}$ ; final scale 2 and clipping $\pm 8$ . No separately calibrated $\sigma^{2}$ is used.

TABLE VI: Frozen campaign and receiver parameters used for the reported held-out results.

Term	Distribution or deterministic rule	Sampling/correlation level
RO residual	$A_{\mathrm{dev}}\sim\mathcal{N}(0,6000^{2})$ ppm	One draw per packet; shared by every symbol.
Startup and drift	$A_{\mathrm{start}}\sim\mathcal{N}(0,2500^{2})$ ppm with $e^{-(k+96)/48}$ decay; $A_{\mathrm{end}}\sim\mathcal{N}(0,220^{2})$ ppm with linear packet ramp	Independent packet draws; correlated over symbols through the stated curves.
RO random walk	$\eta_{k}\overset{\mathrm{iid}}{\sim}\mathcal{N}(0,2.5^{2})$ ppm/symbol and $\epsilon_{\mathrm{rw}}[k]=\sum_{i=0}^{k}\eta_{i}$	Independent increments; accumulated within a packet.
Data pushing	$120(w_{k}-\bar{w})$ ppm	Deterministic from the packet width sequence.
Timing/width perturbations	Boundary jitter $\mathcal{N}(0,4^{2})$ ps; packet-shared width offset $\mathcal{N}(0,8^{2})$ ps; per-symbol width jitter $\mathcal{N}(0,2^{2})$ ps; RX-LO offset $\mathcal{N}(0,2^{2})$ ppm	Independent draws except for the packet-shared width term; the RO state jointly sets symbol period, carrier offset, width, and accumulated phase.
ID proxy hierarchy	Cluster path loss $\mathcal{N}(48,3^{2})$ dB; proxy increment $\mathcal{N}(0,1.5^{2})$ dB; packet increment $\mathcal{N}(0,0.6^{2})$ dB. Cluster RMS delay $0.85e^{\mathcal{N}(0,0.30^{2})}$ ns and proxy multiplier $e^{\mathcal{N}(0,0.12^{2})}$ .	32 final clusters, eight proxies/cluster; cluster and proxy terms recur, packet terms are redrawn.

TABLE VII: Frozen stochastic RO and tissue-proxy model.

Wireless Communications

Wireless communications studies how information is sent over radio links and how networks manage signals, spectrum, and resources reliably.

See all 403 papers

DBNN: Neural Spike Classification Using a Deep Binarized Neural Network
Jul 6, 2026Binyi Ren, Luca M. Meyer, Majid ZamaniEfficient Neural Network Inference Brain-Computer Interfaces
The School of Electronics and Computer Science (ECS), University of Southampton, Southampton SO17 1BJ, U.K.
Spiking Personalized Federated Learning for Brain-Computer Interface-Enabled Immersive Communication
Mar 24, 2026Chen Shang, Dinh Thai Hoang, Diep N. Nguyen +1Federated Learning Brain-Computer Interfaces
School of Electrical and Data Engineering, University of Technology Sydney, Australia · Thrust of Internet of Things, The Hong Kong University of Science and Technology (Guangzhou), China
SwitchBraidNet: Quantisation-Aware Lightweight Architecture for Hybrid Brain-Computer Interface
Jun 17, 2026Gourav Siddhad, Yogesh Kumar MeenaEfficient Neural Network Inference Brain-Computer Interfaces
Human-AI Interaction (HAIx) Lab, Indian Institute of Technology Gandhinagar, Gujarat, 382055, India

Jul 6, 2026J/K move · Enter open · S save

DBNN: Neural Spike Classification Using a Deep Binarized Neural Network

Binyi Ren, Luca M. Meyer, Majid Zamani

The School of Electronics and Computer Science (ECS), University of Southampton, Southampton SO17 1BJ, U.K.

Efficient Neural Network Inference Brain-Computer Interfaces Binary Neural Networks

Open paper AskPDF arXiv

Implantable brain-computer interfaces require on-node spike sorting to reduce telemetry bandwidth and power while maintaining reliable neural decoding. This paper presents a hardware-oriented deep binarized neural network (DBNN) spike-sorting system with two binarized hidden layers with 256 neurons and a fixed-point output layer to enable multiplier-free inference dominated by sign-controlled accumulation and bit-wise logic. The proposed classifier operates on compact 16-sample spike waveforms to reduce the implementation cost (16-256-256-3) and achieves a median classification accuracy of 98.7% on both synthetic and in-vivo datasets. An FPGA prototype on a Cyclone V device operates at 50 MHz and requires 528 cycles per spike, corresponding to a 0.01 ms compute latency, while consuming 828 ALMs and 1023 registers with zero DSP blocks. For ASIC feasibility, the DBNN is implemented using FreePDK45-based flow; synthesis in Synopsys Design Compiler indicates an estimated silicon area of 0.014 mm2 and an operating power of 122 nW at 20 kHz under a 1.1 V supply. These results demonstrate that the proposed DBNN spike sorter offers a favorable trade-off between accuracy and implementation cost, supporting low-power, implantable neural interfaces. Overall, the proposed DBNN spike sorter achieves high accuracy (98.7%) with extremely low hardware cost (0.014 mm2, 122 nW at 20 kHz) and multiplier-free operation, making it suitable for low-power, implantable neural interfaces. This paper introduces the first DBNN designed for real-time neural spike sorting, striking an excellent balance between input data size and network complexity.

Abstract

Wireless Communications

Explore similar work

DBNN: Neural Spike Classification Using a Deep Binarized Neural Network

Spiking Personalized Federated Learning for Brain-Computer Interface-Enabled Immersive Communication

SwitchBraidNet: Quantisation-Aware Lightweight Architecture for Hybrid Brain-Computer Interface