Stage-Wise Divergence Detection: A Diagnostic Framework for Closing Real-Time vs. Offline Accuracy Gaps in ERP-Based Brain-Computer Interfaces

Objective: Brain-computer interfaces (BCIs) based on event-related potentials (ERPs) exhibit accuracy degradation when offline-trained classifiers are deployed in real-time streaming environments. The online-offline gap is widely acknowledged but lacks a systematic diagnostic methodology. We present a stage-wise divergence detection framework that decomposes the preprocessing chain into six ordered stages and injects comparison probes at each boundary to localize signal-path mismatches. Approach: The framework instruments the ERP preprocessing pipeline at six stages (A–F: raw epoch slicing, bandpass filtering, baseline correction, Euclidean Alignment, channel reordering, trial-level scoring). Pearson correlation and L2-ratio metrics quantify divergence between offline and real-time paths at each stage. We validated the framework on a P300 row-column speller across 12 subjects from the BigP3BCI dataset (32 channels, 256 Hz) using a compact convolutional ERP classifier. Main results: The framework isolated two sequential root causes: (1) a filter phase mismatch at Stage B (causal vs. zero-phase IIR; r = 0.724), and (2) a trial-ordering error at Stage F (chronological vs. alphabetical; mean posterior r = 0.023 across 25 affected trials). Remediation recovered mean real-time accuracy from 0.5% to 68.1%, within 2.8 pp of matched offline scoring (70.9%). Stages A, C, D, and E produced negative results, confirming robustness of epoch extraction, baseline correction, Euclidean Alignment, and channel mapping. Significance: To our knowledge, this is the first reusable diagnostic protocol for systematically localizing preprocessing-induced online-offline accuracy gaps in ERP-based BCIs. The framework is paradigm-agnostic and requires no classifier modification.