Input-Adaptive Direct Neural ADC via Level-Crossing Quantization: Feasibility, Statistics-Driven Design, and HDR Directions

Level-crossing (LC) ADCs are attractive for neural interfaces because their event-driven operation naturally adapts conversion activity and output data rate to the input waveform dynamics. However, reported LC-ADC neural front ends still rely on a preceding analog front end (AFE), limiting systemlevel benefit because AFE power is largely input-independent. This paper investigates whether an LC-ADC can operate as a direct neural ADC with the input tied to the electrode, i.e., without a preceding amplifier. We first derive closed-form relations that connect LC-loop speed, comparator bandwidth and gain, and input-referred-noise constraints to the maximum instantaneous input slope, yielding practical lower bounds on front-end power for both continuoustime (CT) and discrete-time (DT) LC implementations. The analysis shows that worst-case sizing based on full-scale amplitude at maximum bandwidth can make direct LC quantization impractical, whereas DT comparators with regenerative latches substantially relax the gain-delay bottleneck. Next, motivated by the shaped spectrum and bounded slope statistics of real neural signals, we show that LC-loop requirements should be specified using the empirical slope distribution of raw EEG/iEEG rather than flat-spectrum worst-case assumptions. As a transistor-level proof point in the normal dynamicrange regime, we present a 180 nm CMOS LC-ADC designed using this statistically informed methodology and report postlayout performance. Finally, we examine high-dynamic-range (HDR) scenarios with large artifacts, explain why conventional LC-ADCs lose efficiency as slope demands grow, and outline architectural directions that combine LC adaptivity with HDRefficient noise shaping.