Brain-inspired cascaded EEG speech decoding with conformal calibration and adaptive exit: simulation validation and five-stage real-data assessment

Background Online EEG-based speech decoding faces dual challenges: limited computational resources for real-time deployment, and unreliable early-exit decisions when models underestimate their own errors (silent dropping of speech segments). Existing deep decoding models lack statistically rigorous uncertainty quantification, and the exchangeability assumption of conformal prediction— commonly used for calibration—is questionable for strongly auto-correlated neural signals. Systematic evidence on when and why conformal guarantees fail on real EEG, and whether online repair is feasible, remains scarce. New method We propose a brain-inspired three-tier cascaded decoding architecture (linear baseline → MLP residual corrections), equipped with pinball quantile error heads (τ = 0.75) for conservative error estimation, split-conformal temperature calibration with a block-wise variant to mitigate exchangeability violation, and an adaptive exit controller that decouples computational budget (ρ- quantile) from precision safety (κ· ˆe ≤ ϵ). For online distribution shifts, we implement Gibbs–Candès adaptive conformal inference (ACI) to dynamically adjust κ without relying on error-head quality. The framework is validated through simulation and a five-stage real-data assessment on OpenNeuro ds004279 (3 subjects, speech-perception EEG), plus an extended three-method cross-validation (temporal TRF, CCA lag scan, word-onset ERP permutation) to adjudicate millisecond-scale speech tracking. Results Simulation confirms conformal coverage at 84.5% (target 85%), exit rate anchored at 39.5%±0.6%, and 18 non-dominated Pareto solutions all outperforming the full-run baseline. On square-wave event labels, coverage remains 85.0% across four scenarios, and adaptive exit beats both full-run and baseline in 3/3 subjects. Switching to real audio envelopes increases label information 3.4-fold (|r| 0.057→0.194) but collapses coverage to 7–29%. Block-conformal root-cause analysis identifies six-fold κ drift across subjects, strongly coupled to error-head calibration failure. ACI repairs coverage to ∼99% in 2/3 subjects (including one with negative calibration correlation), while the remaining case is diagnosed as “uncalibratable” due to test-segment error-structure collapse, necessitating data-quality gating. Extended experiment shows that crude pipelines yield all-negative results from dual technical failures (drift artifact, ERD normalization failure); after official preprocessing alignment, word-boundary β/γ ERD emerges (group p = 0.028, 3/9 individual significant) and temporal CCA shows group significance (t = 5.24, p = 0.0004) despite individual r < 0.03, while band-power CCA significance (p = 0.0065) is falsified as segment-level coupling artifact. Comparison with existing methods Unlike conventional deep decoders (CNNs, LSTMs) and recent attention-based models that provide only point estimates without uncertainty bounds, our framework delivers distribution-free coverage guarantees via conformal calibration. The adaptive exit mechanism offers flexible speed-accuracy trade-offs absent in fixed-architecture models. Critically, ACI maintains validity even when error-head ranking fails—a property not achievable by static conformal approaches. Our five-stage negative-result trajectory (including coverage collapse and rootcause diagnosis) provides an applicability boundary map for conformal prediction on real neural data, which prior studies have not systematically documented. The extended three-method cross-validation further reveals that band-power CCA, commonly used in the literature, can produce false positives due to segment-level coupling, whereas temporal CCA and word-onset ERP provide more reliable group-level speech-tracking signals. Conclusions The proposed cascaded framework with conformal calibration and adaptive exit is validated both in simulation and real-EEG scenarios, with ACI effectively repairing coverage under distribution shifts. The deployable profile defaults to TRF linear regression + ACI + quality gating, with cascaded MLP reserved for high-signal-to-noise conditions. The extended experiment establishes a methodological baseline requiring full preprocessing-chain alignment, temporal features, and threemethod cross-validation for speech-tracking studies. All code and results are released open-source for reproducible research.