📄 中文摘要
从长期脑电图（EEG）记录中生成总结异常模式、诊断结果和临床解释的临床报告，是一项劳动密集型任务。为此，我们构建了一个大规模临床EEG数据集，其中包含9,922份报告和约11,000小时的EEG记录，这些数据来源于9,048名患者。基于此数据集，我们开发了CELM（Clinical EEG-to-Language Model），这是首个能够从长时间、变长EEG记录中进行总结的临床EEG-到-语言基础模型。CELM通过深度学习技术，将复杂的神经信号模式转化为结构化且易于理解的自然语言描述，显著提高了临床报告生成的效率和准确性。
📄 English Summary
Neural Signals Generate Clinical Notes in the Wild
Generating clinical reports that summarize abnormal patterns, diagnostic findings, and clinical interpretations from long-term electroencephalogram (EEG) recordings remains a labor-intensive endeavor. To address this challenge, we have curated a large-scale clinical EEG dataset comprising 9,922 reports paired with approximately 11,000 hours of EEG recordings from 9,048 patients. Leveraging this extensive dataset, we developed CELM (Clinical EEG-to-Language Model), the first clinical EEG-to-Language foundation model capable of summarizing long-duration, variable-length EEG recordings. CELM employs advanced deep learning techniques to transform complex neural signal patterns into structured and easily understandable natural language descriptions, significantly enhancing the efficiency and accuracy of clinical report generation. At its core, the model exhibits strong multimodal learning capabilities, effectively integrating time-series EEG data with medical text information. This allows CELM to capture subtle variations within EEG signals and map them to corresponding clinical terminology and diagnostic descriptions. The architecture of CELM incorporates advanced Transformer variants and specially designed temporal feature extraction modules to handle the non-stationarity and high dimensionality inherent in EEG data. During training, we adopted a self-supervised pre-training strategy, utilizing large volumes of unlabeled EEG data to learn general neurophysiological patterns, followed by fine-tuning on annotated EEG report data to adapt to the specific clinical report generation task. Evaluation results demonstrate CELM's exceptional performance in generating clinically acceptable and information-rich summaries. The model accurately identifies and describes critical clinical features such as seizures, abnormal waveforms, and sleep stages, substantially reducing the workload of clinicians and potentially accelerating the diagnosis and treatment processes for neurological disorders.