Critical Review of Mohan et al.’s Neuromorphic Compression for Neural Sensing in Next-Generation Wireless Brain-Machine Interfaces

1. Compression Efficiency:

Mohan et al. report approximately 15×–265× compression when all generated events are transmitted. A higher compression range of 200×–50,000× is reported when transmission is selectively restricted to spike-related events. These two ranges therefore represent different transmission modes and should not be interpreted as equivalent measures of demonstrated compression performance. Interpretation: This demonstrates that event-driven representation can substantially reduce the volume of data requiring wireless transmission. The approach is particularly attractive for future high-density neural interfaces where continuous transmission would impose significant power and bandwidth requirements. Limitation: Higher compression is associated with coarser signal reconstruction. PCM4, for example, produces lower transmission data rates but sacrifices more information about spike shape. Implication: Future systems should optimise compression according to the downstream task rather than treating maximum compression as the primary objective.  

2. Spike Information Preservation

The study evaluates spike-detection performance using sensitivity (TP/(TP+FN)). Sensitivity varies across datasets and modes, ranging from 71.81%–99.86% for synthetic data, 88.38%–96.91% for NHP data, and 87.97%–91.04% for Neuropixel data. Thus, spike preservation depends on dataset and compression mode. For NHP and Neuropixel data, sensitivity was based on “potential spike” times generated using AT-SPD rather than independently verified labels. Interpretation: The results indicate that neural spike timing can remain relatively robust despite substantial compression. Critical assessment: The results are encouraging, but preservation of spike shape is not uniform across all compression levels. APM and PCM1 generally perform better than PCM2 and PCM4 because higher PCM bin widths produce coarser representations. Therefore, the study supports the feasibility of compression-based neural sensing, but the optimal compression level depends on whether the intended application requires detailed spike morphology or primarily spike timing.

3. Collision Handling and Scalability

A key technical strength of the study is its detailed simulation of collisions in the AER readout system. Using Cadence Virtuoso circuit modelling, the authors demonstrate that the toggle-tree arbitration process can handle simultaneous events across a simulated 10,000-channel array with negligible data loss. However, this evaluation remains limited to architectural simulations and synthetic data replication, without proving full 10,000-channel operation in physical silicon or in vivo environments. The study also reports real-data collision rates in the NHP and Neuropixel datasets. Approximately 25% of events corresponding to spike samples were involved in collision scenarios, highlighting why collision management is important for dense neural arrays. Critical interpretation: The 10,000-channel experiment provides useful evidence of architectural scalability. However, the 10,000-channel case is based on synthetic replication rather than a physically recorded 10,000-channel neural dataset. The authors themselves describe it as a worst-case simulation based on replicating a single-channel signal. Thus, the scalability claim should be interpreted as architectural scalability demonstrated through simulation, rather than experimental validation of a 10,000-channel implant. Comparison with the Literature Comparisons with related studies provide context for Mohan et al. rather than a uniform performance ranking. Differences in datasets, metrics, experimental conditions, and validation methods mean that reported values should be interpreted as context-specific results rather than directly equivalent benchmarks.

Study	Approach	Reported Performance	Validation Context
Mohan et al.	Neuromorphic AER sensing and compression	15–265×; higher values under selective spike transmission	Synthetic data, experimental datasets, and 10,000-channel simulation
Yan et al.	Neural signal compression	Reported compression performance	Dataset/experimental conditions differ from Mohan et al.
Wang et al.	Neural data compression/processing	Reported study-specific metrics	Different signal and evaluation conditions
Ke et al.	Neural sensing/detection architecture	Detector-level performance	Hardware validation at detector/macro level
Krishna et al.	Neural signal processing/compression	Study-specific metrics	Different dataset and validation conditions
He et al.	Neuromorphic neural interface	Study-specific performance	Hardware/architecture-specific validation

Overall, Mohan et al. demonstrates a distinctive integration of neuromorphic sensing and event-based compression, but its reported performance should be interpreted within the specific datasets, transmission modes, and simulation conditions used in the study rather than as a directly comparable benchmark against all existing approaches.

The practical value of the proposed wireless implantable brain-machine interface

Based on extrapolations from prior ADM hardware measurements, the authors estimate a front-end surface power density of approximately 9–13 mW cm⁻², below the general iBMI thermal budget of ~80 mW cm⁻². They suggest that over 100× data compression could reduce communication power requirements, making RF or inductive telemetry more feasible. However, this value is an extrapolated front-end estimate rather than a measured power density of a fully integrated, fabricated implant..

Another practical advantage is that the event representation can potentially be directly interfaced with low-power spiking neural networks for spike detection, spike sorting and neural decoding.

However, an important limitation is that the simulations assume an ideal transmission channel. Noise interference and signal coupling effects that could occur in real wireless implants were not investigated.

Therefore, the transition from simulation to an implantable system requires further investigation of wireless channel reliability, interference, circuit variability, thermal constraints and long-term operation.

 

The linked anchor leads to an unrelated biomaterials call-for-papers page. Replace it with a relevant, live internal BMI resource or remove the link.

 

State that this is an estimate extrapolated from prior ADM hardware measurements, not a measured full-system implant power density.