The decoding side of BCIs has made remarkable progress. Kalman filters—originally developed for aerospace tracking—work surprisingly well for cursor control, correlating decoded kinematics with actual movement at 0.8-0.99. More recently, LSTM recurrent neural networks match or exceed Kalman performance by leveraging temporal dynamics (https://arxiv.org/pdf/1812.09835).

Signal features matter. Single-unit spiking activity provides the highest spatial and temporal resolution, but it degrades. Electrodes shift, glial scarring builds up, and the neurons we were recording from die or move away. Local field potentials (LFPs) and ECoG are more stable but coarser—better for long-term use, worse for precise control.

The clinical translation challenges are:

1. Signal drift requiring frequent recalibration—users have to stop and retrain the decoder
2. Inter-subject variability—each person needs a personalized model
3. Computational demands—sophisticated algorithms are hard to run on implantable hardware

Testable predictions:

1. Adaptive decoders that track signal drift in real-time will outperform static models in multi-month trials
2. LFP-based BCIs will achieve clinical viability before single-unit systems due to stability
3. Hybrid approaches combining spike and LFP features will offer the best accuracy-stability tradeoff

The future is not just better algorithms—it is better electrodes and biological integration. The Neuralink approach of more, thinner electrodes may help, but the fundamental materials problem remains.

Research synthesis via Aubrai.