High-density neural recordings bring diverse probe geometries and complex, neuron-specific drift. Most methods still assume a fixed geometry and a single drift dimension. KIASORT works without those assumptions. It tracks each neuron's drift on its own and trains channel-specific classifiers in a hybrid linear and nonlinear embedding space.
Key innovations
The ideas that set KIASORT apart from existing spike sorters.
Geometry-free, per-neuron drift tracking
Detailed biophysical simulations show that even sub-micron probe shifts (under 1 µm) produce neuron-specific waveform changes that standard drift models cannot correct. KIASORT tracks each neuron independently, without assuming anything about how channels are arranged on the probe.
Hybrid linear and nonlinear embedding
Where other methods rely on linear embeddings alone, KIASORT pairs PCA with UMAP to keep both the global and local geometry of spike waveforms. Dropping the assumption of purely linear variability gives cleaner clusters when waveforms change in nonlinear or time-dependent ways.
Modular design built for real time
Three modules handle sampling, clustering and training, and sorting. Once the classifiers are trained on initial data, the sorting module runs on incoming streams in real time, which suits closed-loop experiments, brain-computer interfaces, and clinical use.
Ready for flexible, next-generation probes
Because it makes no geometric assumptions, KIASORT handles the non-rigid, shifting channel layouts of flexible probes now entering human trials, including those from Neuralink, Synchron, and Paradromics, as readily as it handles high-density arrays like Neuropixels.
Key features
Geometry-free tracking
The first per-neuron tracking system that drops the fixed-geometry assumption and follows each neuron's drift on its own.
Hybrid dimensionality reduction
PCA and UMAP together preserve both the global and local structure of waveforms, instead of assuming variability is linear.
Future-ready design
Works with flexible probes from Neuralink, Synchron, and Paradromics, where channel layouts deform and move.
Modular architecture
A three-stage design that supports real-time processing for closed-loop experiments and brain-computer interfaces.
Integrated platform
One GUI for automated sorting and manual curation, so the path from raw data to results stays in a single place.
Runs on a standard CPU
GPU acceleration is supported, but KIASORT reaches real-time speed on CPUs alone, which comparable tools cannot do.
KIASORT in action
Hover to see after
Denoising
Noise is removed while the features that matter for sorting are preserved. Hover to compare the recording before and after denoising.
Neuron-specific drift tracking
Each neuron is followed independently, so sorting stays accurate even when probe motion is complex and differs from one neuron to the next.
Real data example
Performance of KIASORT across different high-density probes. Over a range of geometries and recording conditions, it recovers more high-quality units than current state-of-the-art methods.
Integrated GUI
From raw-data inspection to curated results in one interface, with automated sorting and manual curation tools side by side.
Open source on GitHub
Get the source, follow development, and join the community of researchers using KIASORT.
github.com/banaiek/KIASORTTutorials
Guides to install KIASORT and run your first sort, plus a video walkthrough.
Tutorial
A full walkthrough of the workflow, from loading data to sorting and curating results.
Download PDFInstallation guide
Step-by-step setup, including MATLAB toolboxes, Python packages, and system requirements.
Download PDFVideo walkthrough
Get in touch
Questions about KIASORT, or interested in collaborating? Send a note.
Kianooshbb@gmail.com