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(pred_events, gt_events, num_classes):
    """
    pred_events: [(time, class_logit), ...] — N predictions
    gt_events: [(time, class_label), ...] — M ground truth
    """
    # Build cost matrix: time distance + classification cost
    cost_matrix = build_cost_matrix(pred_events, gt_events)
    row_ind, col_ind = linear_sum_assignment(cost_matrix)
    
    # Matched pairs: regression + classification loss
    matched_loss = compute_matched_loss(pred_events, gt_events, row_ind, col_ind)
    
    # Unmatched predictions: objectness penalty
    unmatched_loss = compute_unmatched_loss(pred_events, set(range(len(pred_events))) - set(row_ind))
    
    return matched_loss + unmatched_loss

Performance

Evaluated on 10 datasets spanning:

• Cognitive tasks (ERP detection)
• Clinical monitoring (seizure detection — new SOTA)
• BCI applications (matches onset-informed model accuracy)

When to Use

• Use DANCE when: Continuous EEG monitoring without onset labels, real-time seizure detection, asynchronous BCI, mixed event-type pipelines
• Don't use when: You have perfectly aligned event windows and only need classification (simpler models suffice)
• Keywords: DANCE, EEG event detection, continuous decoding, set prediction, asynchronous BCI, seizure monitoring, unaligned neural signals

Comparison to Traditional Approaches

Aspect	Window-based Classification	DANCE (Set Prediction)
Onset labels needed	Yes	No
Event duration range	Fixed	Variable
Real-time capable	Requires alignment	Direct from raw signal
Seizure detection	Suboptimal	SOTA
BCI accuracy	Good (with alignment)	Matches aligned models

arXiv Reference

• Paper: "DANCE: Detect and Classify Events in EEG"
• ID: 2605.10688
• URL: https://arxiv.org/abs/2605.10688