Bridging the Compression-Precision Paradox: A Hybrid Architecture for Clinical EEG Report Generation with Guaranteed Measurement Accuracy

Automated EEG monitoring requires clinician-level precision for seizure detection and reporting. Clinical EEG recordings exceed LLM context windows, requiring extreme compression (400:1+ ratios) that destroys fine-grained temporal precision. A 0.5 Hz error distinguishes absence epilepsy from Lennox-Gastaut syndrome. LLMs lack inherent time-series comprehension and rely on statistical associations from compressed representations. This dual limitation causes systems to hallucinate clinically incorrect measurement values. We separate measurement extraction from text generation. Our hybrid architecture computes exact clinical values via signal processing before compression, employs a cross-modal bridge for EEG-to-language translation, and uses parameter-efficient fine-tuning with constrained decoding around frozen slots. Multirate sampling maintains long-range context while preserving event-level precision. Evaluation on TUH and CHB-MIT datasets achieves 60% fewer false alarms, 50% faster detection, and sub-clinical measurement precision. This is the first system guaranteeing clinical measurement accuracy in automated EEG reports.

💡 Research Summary

The paper tackles a fundamental obstacle to automated clinical EEG reporting: the “compression‑precision paradox.” EEG recordings span hours, thousands of channels, and millions of samples, far exceeding the token limits of contemporary large language models (LLMs). To fit these recordings into an LLM’s context window, compression ratios above 400:1 are required, but such aggressive down‑sampling destroys the fine‑grained temporal and spectral resolution needed for clinical decision‑making—e.g., a 0.5 Hz frequency error can misclassify absence epilepsy (3 Hz) as Lennox‑Gastaut syndrome (3.5 Hz). The authors formalize this dilemma mathematically (Theorem 1) and prove that any end‑to‑end neural encoder with compression >100 will inevitably map clinically distinct signals to indistinguishable embeddings, leading to systematic hallucination of measurement values.

To resolve the paradox, the authors propose a hybrid architecture that separates exact measurement extraction from narrative generation. The pipeline consists of five stages:

1. Hierarchical multirate sampling – a low‑rate stream (256–512 Hz) provides long‑range context, while a high‑rate stream (≥1 kHz) is activated only around candidate events detected via energy, kurtosis, and spectral peaks. This preserves hour‑long context without the computational burden of processing the entire signal at high resolution.

2. Measurement‑first guardrails – before any neural compression, traditional DSP routines compute precise clinical metrics: dominant frequency via Welch PSD, event duration via hysteresis thresholding, amplitude via median absolute deviation, and lateralization via graph‑based asymmetry indices. These values are stored in immutable “frozen slots” together with full provenance (algorithm, window, channels).

3. Graph‑aware neural encoder – the encoder ingests dual‑view inputs (time‑domain patches and transform‑domain band‑power) and incorporates a channel‑graph bias into multi‑head attention (L = QKᵀ/√d + βB). Linear‑time state‑space models (S4/Mamba) are interleaved to capture very long dependencies efficiently.

4. Cross‑modal bridge – EEG embeddings (768 dim) are progressively projected to the LLM’s semantic space (≈4096 dim) through intermediate layers (768→1536→2816→4096). Learnable clinical anchors, initialized from medical terminology embeddings, are aligned with EEG patterns using a contrastive InfoNCE loss, ensuring that electrophysiological signatures map to the correct clinical concepts.

5. Constrained report generation – a LoRA‑adapted LLM first emits a structured JSON schema containing the frozen measurement slots, then generates free‑form narrative conditioned on this schema. Decoding is constrained by token‑masking so that numeric fields can only be copied from the frozen slots, eliminating hallucinated numbers. Conformalized quantile regression and Earth‑Mover‑Distance (EMD)‑aware supervision provide calibrated uncertainty estimates and guarantee coverage under distribution shift.

The system is evaluated on three large public EEG corpora (TUH, TUSZ, CHB‑MIT) across three tasks: seizure detection (false alarms per 24 h and latency), value extraction (MAE for frequency, duration, amplitude), and localization (lateralization accuracy, Jaccard overlap). Compared with strong baselines (EEGNet, DeepConvNet, BENDR, EEGFormer, Chronos‑2), the proposed method achieves a 60 % reduction in false alarms (0.51 FA/24 h vs. 1.16), a 50 % reduction in detection latency (10.5 s vs. 24.2 s), and measurement errors within clinical tolerances (frequency MAE = 0.18 Hz, amplitude MAE ≈ 3 µV). Ablation studies confirm that each component—measurement guardrails, graph attention, SSM layers, hierarchical sampling, and conformal calibration—contributes substantially to performance. Robustness tests with injected artifacts (EOG/EMG/line noise) and missing channels show modest degradation (<30 % increase in false alarms, <25 % rise in measurement error), demonstrating resilience in realistic clinical settings.

The authors acknowledge limitations: the current implementation requires high‑end GPUs (A100) and substantial memory, making bedside or wearable deployment challenging without model pruning or quantization. Fixing numeric slots also reduces linguistic flexibility for expressing uncertainty or institution‑specific terminology. Future work will explore lightweight model variants, multimodal integration with imaging and electronic health records, and more expressive uncertainty phrasing.

In summary, this paper provides a theoretically grounded and empirically validated solution to the compression‑precision paradox, delivering the first EEG‑to‑text system that guarantees clinically accurate measurement values while leveraging LLMs for fluent report composition. The approach bridges signal‑processing rigor with modern language generation, paving the way for FDA‑compliant, AI‑assisted EEG interpretation in real‑world healthcare environments.