MICCAI 2026 · Strasbourg, France

SpikeMamba

Spike-Driven State Space Models for
Energy-Efficient Biomedical Sequence Modeling

Si Yong Lee¹Ryangjin Lee¹Hawon Park¹Yoora Kim¹Byungkon Kang²Yoon Seok Yang²

¹ Stony Brook University² The State University of New York Korea

The core idea

Let the signal decide
when to compute.

SpikeMamba maps selective state space dynamics onto spiking neurons. Sparse binary events carry information, while a state-dependent gate suppresses updates during inactive segments.

5Biomedical benchmarks
across ECG and EEG

~75KParameters in the
Tiny configuration

12–16%Firing rates reported on
PTB, PTB-XL, and ADFTD

01 / Method

A state space model.
A spiking core.

Continuous inputs. Binary events.
Selective recurrent updates.

Original architecture figure: stacked SpikeMamba layers, temporal spike encoding, LIF dynamics, state-dependent gating, and a linear output head. — FIGURE 01The SpikeMamba architecture. A spiking recurrent core sits within residual layers; the input projection and classifier retain continuous arithmetic.SVG ↓

Encode the signal

A linear projection, batch normalization, and SiLU prepare continuous biomedical signals. Depthwise convolution mixes local temporal information.

Update through spikes

Leaky Integrate-and-Fire dynamics maintain a membrane state and emit binary spikes. Downstream spike-weight products become conditional accumulations.

Gate inactive steps

A state-dependent gate uses firing activity and membrane-potential variance to determine whether to update or preserve the state.

Energy accounting

E = E_MAC · FL_Input+MLP + E_AC · R_eff · FL_Spike

R_eff = firing rate × gate activation ratio

02 / Results

Explore the trade-off.

Accuracy and estimated energy,
across five biomedical datasets.

Accuracy vs. estimated energy

Higher & further left is better ↖

SpikeMamba BaselinesEnergy axis: logarithmic

Evaluation: PTB · 2-class ECG classification · subject-independent protocol.

Exact values Accuracy, F1, AUROC & estimated energy

PTB results · Paper, Table 2
Model	Accuracy (%)	F1 (%)	AUROC (%)	Energy (mJ)
iTransformer	83.89	79.06	91.18	51.51
MTST	76.59	67.38	86.86	8803.84
Reformer	77.96	69.65	91.13	3866.06
Medformer	83.50	79.18	92.81	7564.93
SpikeMamba-L	91.72	94.00	94.92	322.91
SpikeMamba-M	89.00	92.27	92.73	46.78
SpikeMamba-T	80.34	86.72	84.35	4.69

Reading these results. Energy values are analytical estimates under a 45 nm CMOS operation model (MAC: 4.6 pJ; AC: 0.9 pJ), not on-chip measurements. PTB, PTB-XL, and ADFTD use subject-independent evaluation. CHB-MIT and MIT-BIH use segment-level random 6:2:2 splits. Values are transcribed from Tables 2–3 of the paper.

03 / Evaluation

From biosignals
to benchmarks.

Three model scales.
Two biomedical signal modalities.

Experimental pipeline linking ADFTD and CHB-MIT EEG and PTB, PTB-XL, MIT-BIH ECG to three model sizes and evaluation metrics. — FIGURE 02Five biomedical benchmarks, three model scales, and a shared operation-based energy analysis.SVG ↓

Tiny

~75K parameters

4 layers · 64 hidden dimensions

Medium

~1.1M parameters

4 layers · 256 hidden dimensions

Large

~8.6M parameters

8 layers · 512 hidden dimensions

04 / Reference

Build on this work.

If you find SpikeMamba useful in your research,
please consider citing the paper.

Paper ↗GitHub ↗BibTeX ↓

@inproceedings{lee2026spikemamba,
  title = {SpikeMamba: Spike-Driven State Space Models
    for Energy-Efficient Biomedical Sequence Modeling},
  author = {Lee, Si Yong and Lee, Ryangjin and
    Park, Hawon and Kim, Yoora and Kang, Byungkon
    and Yang, Yoon Seok},
  booktitle = {Medical Image Computing and
    Computer Assisted Intervention -- MICCAI 2026},
  year = {2026}
}

← Back to all research Questions about SpikeMamba? Get in touch ↗