Biomedical researchers have introduced a custom hardware-in-the-loop analog resampling platform that converts previously recorded digital neural signals into microvolt-level analog voltages, transmits them through a conductive saline bath, and re-records them under controlled adverse environments to augment training datasets for neural motor decoders. Developed by a team from the University of Minnesota and Fasikl Incorporated, the system addresses a critical bottleneck in neural interface development: training high-capacity deep learning decoders demands massive, diverse datasets that are exceptionally laborious and costly to gather through direct human experimentation. By establishing a physical pathway to replay authentic physiological recordings through targeted physical distortions, the platform provides machine learning models with abundant corrupted data alongside exact clean reference signals, enabling systematic evaluation and training of artifact mitigation algorithms.
Signal corruption from environmental noise, electronic circuitry, stimulation feedback, and ambient wireless transmissions presents a persistent challenge across neural recording applications such as electroencephalography, deep brain stimulation, and peripheral motor decoding. In peripheral nerve interfaces designed for motor prosthetics, maintaining high classification accuracy requires exceptional signal fidelity across multiple movement classes, yet collecting the extensive datasets needed for each user is severely constrained by experimental time and human fatigue. Concurrently, emerging neuroprosthetic designs increasingly combine recording with biomimetic stimulation feedback or compact wireless transmitters. These additions inject substantial interference, such as stimulation artifacts and radio-frequency bursts, which can saturate front-end amplifiers, cause prolonged ringing, or mimic genuine physiological activity, frequently causing downstream decoders to fail completely.
To overcome the limitations of purely synthetic numerical simulations, the research team constructed an experimental apparatus that physically replicates the analog recording environment. Digital motor control signals stored on a computer are streamed over a serial universal serial bus link to an Artix-7 field programmable gate array, which buffers the source samples and drives a 16-bit digital-to-analog converter at the original recording clock rate. Because commercial converters typically operate across multi-volt spans and suffer severe quantization distortion near their noise floor when generating microvolt biological potentials, the circuit pairs the converter with precision operational amplifiers, differential-to-single-ended conversion stages, and an instrumentation attenuation stage. This hardware scales output voltages to one microvolt per least significant bit across a 65.536-millivolt span, delivering an active bandwidth exceeding 100 kilohertz to maintain flat magnitude and linear phase responses across the target 10-kilohertz biological frequency range.
The synthesized analog signal is delivered through a stainless steel monopolar needle electrode into a saline bath designed to approximate the electrical conduction properties of biological tissue. A differential recording setup, built around the Scorpius neural interface, senses the replayed waveform by comparing the immersed active electrode against a reference electrode placed in an isolated phosphate-buffered saline container. Both solutions are maintained at a 600-millivolt bias using wet-gel sodium chloride electrodes, while custom 3D-printed containers secure the spatial positions of all electrodes. Power isolation via a nine-volt battery and dedicated low-dropout regulators supplies stable direct-current rails, shielding sensitive analog stages from 60-hertz mains coupling and digital switching noise.
The underlying human biological recordings used to validate the platform were obtained from three able-bodied adult male participants aged 24, 27, and 30 under protocols approved by the Institutional Review Board at WCG Clinical, Inc. Data acquisition employed four surface channels positioned over the median and ulnar nerves, capturing composite electroneurogram and electromyogram potentials spanning 30 to 1,000 microvolts during five discrete finger flexion gestures. Each subject executed ten trials per gesture, yielding 50 distinct movements per recording session. Signals were preprocessed through antialiasing filtering at 2.5 kilohertz, decimation, a 60-hertz notch filter, and a fourth-order Butterworth bandpass filter spanning 25 to 600 hertz. To resample multichannel data using the single-channel hardware pipeline, individual channels were replayed sequentially with a fixed sinusoidal calibration tone for precise temporal alignment before being recombined into synchronized multichannel datasets.
Baseline validation demonstrated that physical resampling introduced negligible distortion to the underlying biological signals. Time-domain evaluation across all gestures yielded Pearson correlation coefficients exceeding 0.90 across all three participant datasets, ranging from 0.91 for thumb flexion to 0.93 for pinky flexion, while normalized root mean squared error remained below 0.45. In the frequency domain, log-spectral distance remained beneath 0.24 decibels across all datasets, representing less than a 5.6 percent variation in power spectral density. Magnitude-squared coherence peaked strongly in the functionally vital 100-to-200-hertz band before rolling off gradually toward 600 hertz, reaching minimum coherence values of 0.70 and 0.65 in two datasets due to minor high-frequency gain shifts and electrode interface characteristics.
To evaluate whether resampled signals preserved functional decoding characteristics, the researchers tested them against deep neural networks trained exclusively on the original clean human data. Input feature matrices representing 50 time steps across four channels with 14 features each were processed by a network combining an initial convolutional layer, a recurrent hidden layer containing 128 gated recurrent units, and intermediate linear layers optimizing cross-entropy loss over 50 stochastic training runs with randomized initialization seeds. When evaluating resampled signals without weight updates, downstream classification accuracy closely tracked clean baselines, showing an average test deviation of only 0.11 percent in Dataset 0 and 0.57 percent in Dataset 1, with per-gesture accuracy differences consistently held under one percent for most movements.
The team demonstrated the platform’s utility for dataset augmentation by exposing the physical playback path to uncontrolled wireless interference. A commercial smartphone transmitting continuous UDP network packets governed by 802.11 Wi-Fi protocols at a throttled throughput of 1,000 kilobytes per second was placed ten centimeters from the recording electrode, aimed at a router two meters away. Because Wi-Fi operates via carrier sense multiple access protocols with collision avoidance, it generates unpredictable bursts that couple into recording leads, producing sharp interference spikes with positive peak amplitudes averaging 1,031.18 microvolts, negative peaks averaging minus 720.45 microvolts, and post-spike ringing recovery times averaging 9.61 milliseconds. A total of 5,012 wireless interference spikes were cataloged across the resampled recording sets.
To mitigate these distortions, the researchers trained a specialized one-dimensional U-Net denoising architecture using five-fold cross-validation partitioned strictly at the file level into training, validation, and test sets. The network processed 51.2-millisecond windows through four downsampling stages of double convolutions, batch normalization, and max-pooling, linked to an attention bottleneck containing an eight-head multi-head self-attention module, followed by transposed-convolution decoding stages with skip connections. The denoiser substantially suppressed wireless artifacts, elevating signal-to-noise ratios from negative starting levels between minus 13.9 and minus 6.9 decibels up to positive values between 6.3 and 10.2 decibels, representing net improvements of 14.2 to 22.3 decibels. Concurrently, mean absolute error fell from over 200 microvolts to between 16.57 and 36.77 microvolts, while Pearson correlation for active movements recovered from baseline values below 0.39 to over 0.83, peaking at 0.94 for index flexion.
This signal reconstruction directly rescued downstream motor decoding performance that had been completely crippled by radio-frequency interference. On unmitigated corrupted signals, resting-state classification F1 scores plummeted from 95.4 percent to 42.9 percent, thumb flexion fell from 70.0 percent to 29.1 percent, and both middle and ring flexion collapsed to zero percent accuracy. Passing the corrupted signals through the U-Net restored the resting F1 score to 94.7 percent and brought active gesture F1 scores back to within five to eight percent of their original clean baselines. Furthermore, the decoder’s agreement with its original clean predictions recovered from as low as 8.6 percent up to 98.1 percent for resting periods and between 80.4 and 89.8 percent across active movements.
Despite these favorable outcomes, the authors identified several key constraints in the current platform implementation and experimental scope. The validation relied on surface motor recordings collected from a small cohort of three healthy individuals, meaning results do not capture the complete variability of implanted neural interfaces, diverse user populations, or varying tissue impedances. Additionally, the current circuit implementation replays only a single channel at a time, preventing simultaneous multichannel noise injection where spatial coupling patterns vary across recording sites. The physical environment of an exposed metallic needle within an aqueous saline bath also differs from the complex electromagnetic shielding and electrical double-layer impedance encountered in living tissue. Future work will focus on expanding the platform to multi-channel hardware arrays, implementing higher-resolution converters, and validating artifact removal pipelines against in vivo recordings under realistic clinical conditions.
Subject of Research: Synthetic neural data augmentation and artifact removal for deep learning motor decoders
Article Title: Synthetic neural data augmentation for robust deep learning-based neural decoding
Article References: Drealan, M. W., Nguyen, A. T., Lim, B. Z. H., Awada, J., Hoang, L., Luu, D. K., Zhao, Q., & Yang, Z. (2026). Synthetic neural data augmentation for robust deep learning-based neural decoding. Frontiers in Neuroscience, 20, 1954119. https://doi.org/10.3389/fnins.2026.1954119
Image Credits: AI Generated
DOI: 10.3389/fnins.2026.1954119
Keywords: neural decoding, biosignal processing, data augmentation, hardware in the loop, analog resampling, electromyography, electroneurogram, deep learning, 1D U-Net, wireless interference, artifact removal, neuroprosthetics, signal to noise ratio, digital to analog converter, motor control, biomedical engineering
News Source: Cassandra Pierce. (November 20, 2026). Analog Neural Resampling Cleans Wireless Noise to Restore Motor Decoder Accuracy. Scienmag.



