dance-eeg-event-detection

name

dance-eeg-event-detection

description

DANCE (Detect And Classify Events) — deep learning pipeline for joint event detection and classification from continuous, unaligned EEG signals. Frames neural decoding as a set-prediction problem, eliminating the need for pre-aligned event windows. Achieves SOTA on seizure monitoring and matches onset-informed BCI models. Use when working with: EEG event detection, continuous neural decoding, seizure detection, asynchronous BCI, set prediction for time series, or real-time neural monitoring. Trigger words: DANCE, EEG event detection, continuous decoding, set prediction EEG, asynchronous neural decoding.

DANCE: Detect And Classify Events in EEG

Based on arXiv:2605.10688

Core Problem

Traditional EEG decoding classifies fixed windows aligned to known event onsets — fine for controlled experiments, but onset labels are unavailable in continuous real-world monitoring.

DANCE solves this by framing neural decoding as a set-prediction problem: jointly detect AND classify events directly from raw, unaligned signals.

Architecture

Raw EEG (continuous) → Feature Encoder → Set Prediction Head → {Events: (time, class)}

Key Design

• Set prediction formulation: Model outputs a set of (timestamp, class) pairs — no fixed windows needed
• End-to-end asynchronous: No alignment to stimulus markers required
• Handles variable duration: Events from milliseconds (ERPs) to minutes (seizures) in one model

Set Prediction Loss

Uses bipartite matching (Hungarian algorithm) between predicted and ground-truth events:

from scipy.optimize import linear_sum_assignment

def set_prediction_loss():
    
    
    cost_matrix = build_cost_matrix(pred_events, gt_events)
    row_ind, col_ind = linear_sum_assignment(cost_matrix)
    
    
    matched_loss = compute_matched_loss(pred_events, gt_events, row_ind, col_ind)
    
    
    unmatched_loss = compute_unmatched_loss(pred_events, (((pred_events))) - (row_ind))
    
     matched_loss + unmatched_loss