Vocabulary-Constrained LLM Decoding for Low-Accuracy P300 BCIs

Background: P300 matrix spellers offer a communication pathway for individuals with severe motor disabilities,but single-session letter accuracy is highly variable and often below 50%, motivating strategies that constrain the decoding alphabet to improve classification under fixed EEG signal quality. Objective: We derive and empirically validate a breakeven recall framework determining when large language model (LLM)candidate-set narrowing improves or degrades P300 speller decoding accuracy. Methods: Using the publicly available BigP3BCI dataset (n = 12 healthy adult subjects, 222 vocab-aligned rounds, 47 sessions), we compared three decoding modes: unconstrained 72-symbol Farwell-Donchin scoring, static vocabulary restriction (13 letters), and LLM-assisted narrowing via Qwen2.5 1.5B (top-5 candidates). GPT-2 (124M parameters) was evaluated on the same 222 rounds for cross-model comparison. Paired Wilcoxon tests with Bonferroni correction were used throughout. Results: Both hypotheses (H1: LLM > SV > UC; H2: SV ≈ LLM > UC) were rejected. Static vocabulary achieved 48.1%vs. 30.6% unconstrained (p = 0.001, dz = 1.24). Qwen2.5 1.5B achieved 24.4% at R = 40.1%, below both breakeventhresholds (R∗UC = 49.5%, R∗SV = 77.8%); the breakeven model predicts 24.8% exactly. GPT-2 achieved 5.0% at near-identical recall (39.6%), exposing a second necessary condition: candidate quality (phit = 12.5%) must exceed the unconstrained baseline (pUC = 30.6%). Conclusion: LLM-assisted P300 decoding requires two orthogonal conditions: candidate quality (phit > pUC) and recallsufficiency (R > R∗UC). Neither current model meets both. The framework predicts both models’ pooled accuracy exactly(Qwen2.5 1.5B: 24.8%; GPT-2: 5.0%), providing cross-model validation. Static vocabulary restriction is the robust strategy until both conditions are met; the derived thresholds provide actionable design targets, pending clinical validation (ALS, locked-in).