BCIwiki (bciwiki.com) — A spike-sorting system that pairs deep learning with fuzzy logic classified neurons with 96.4% accuracy and 45.2 ms amortized latency across four rat and monkey recording datasets, Firas Almukhtar reported in Applied Computational Intelligence and Soft Computing on September 24, 2026.
The backbone is a dual-output convolutional neural network-bidirectional long short-term memory (CNN-BiLSTM) architecture that assigns spikes to neurons and estimates event-conditioned waveform morphology. Three soft-computing layers sit around it: multiobjective Pareto screening selects the architecture, interval Type-2 fuzzy membership functions quantify latent-space validity, and fuzzy adaptive thresholding detects spikes when signal quality is uncertain. The author says the multitask loss is motivated by a mutual-information lower bound derived under conventional boundedness assumptions.
The datasets include rat cortical, hippocampal CA1 and monkey motor cortex recordings. Under full supervision the system reached 96.4% accuracy and 94.8% precision, which the author says meets the computational requirement for closed-loop neuroprosthetic control. With 10% labeled data, accuracy was 88.9%, which the author describes as statistically equivalent to unsupervised state-of-the-art methods while adding calibrated confidence ratings and morphology estimates absent from competing pipelines.
The author notes the framework could suit acute clinical settings where calibration data are available, but prospective validation is required before clinical use. For neurons recorded across multiple electrodes, single-electrode processing may cause 12% over-splitting.