Neural Noise

Invasive neurophysiology and whole-brain connectomics for neural decoding in brain implant patients

2 min read

  • Invasive neurophysiology is linked with whole-brain connectomics and neural decoding in patients with brain implants (Nature).1
  • The integration directly informs implant-based decoding pipelines and network-level readouts.1 1

Gardner updates

  • Invasive neurophysiology combined with whole-brain connectomics supports neural decoding in patients with brain implants and informs implant-based decoding pipelines and network-level readouts. 1

Weekly enrichment (2026-07-20)

  • The study (Merk et al., Nature Biomedical Engineering 2025) analyzed 123 hours of invasively recorded brain data from 73 neurosurgical patients treated with brain implants for movement disorders, depression, and epilepsy.2
  • Movement decoding computed FFT features across eight frequency bands from 4–400 Hz in 1,000 ms segments updated at 10 Hz (100 ms resolution), using ridge-regularized logistic regression with 3-fold cross-validation; theta, high-beta, and high-gamma bands carried the most information.2
  • Movement decoding was validated across four cohorts spanning the US, Europe, and China: Berlin (n=12), Beijing (n=11), Pittsburgh (n=16), and Washington (n=19).2
  • Decoding performance negatively correlated with Parkinson’s motor severity (UPDRS-III; Spearman’s rho = −0.36, P = 0.02) and was reduced by therapeutic 130 Hz STN-DBS, flagging disease severity and stimulation as clinical confounds.2
  • An across-patient decoder used functional/structural connectivity fingerprints (4 mm spherical seed ROIs via the Lead-DBS toolbox) plus a five-layer CNN trained with the InfoNCE contrastive loss (CEBRA), enabling a priori channel selection without patient-specific training.2
  • The CEBRA contrastive approach outperformed grid-point and connectomic baselines (leave-one-participant-out within cohort: P = 0.0038 vs grid points, P = 0.0027 vs connectomics) and replaced roughly 3 h 47 min of cumulative individual training time.2
  • Emotion decoding in eight DBS depression patients peaked ~600 ms post-stimulus (balanced accuracy 0.62 ± 0.05), was driven by high-frequency (200–400 Hz), low- and high-gamma activity in prefrontal/cingulate circuits, and correlated with DBS-induced BDI improvement (rho = 0.79, P = 0.01).2
  • For epilepsy, over 100 hours of NeuroPace RNS recordings from nine focal-epilepsy patients (mean age 35.3 ± 8.2 years) were used to optimize the embedded bandpass seizure detector, raising the F1 score from 0.41 ± 0.12 (original settings) to 0.92 ± 0.06 (optimized) by cutting false positives.2
  • The methods are released as the open-source py_neuromodulation Python platform, with connectome templates for channel selection shared openly (e.g., via Zenodo).3

Footnotes

  1. https://news.google.com/rss/articles/CBMiX0FVX3lxTE1uVUNvRUI5MjZfOXpPUTZNRmhSdmd4ZHQzdzg5WWtGcG9haVhTZmpETlo1V1NSQ0ItcTlPRGtGMTlLbTJzVXVsSFNMQjJIV1Z0RnRQanktMHFyUUo3V0V3?oc=5 ↩ ↩2 ↩3 ↩4

  2. https://www.nature.com/articles/s41551-025-01467-9 ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8

  3. https://neuromodulation.github.io/py_neuromodulation/ ↩

Graph View