腦電圖於預測疑似腦炎患者自體抗體之可行性： 先導機器學習研究

Abstract

自體免疫腦炎與副腫瘤腦炎是不常見但對患者有重要影響的疾病。其診斷之黃金標準為腦脊髓液或血清之抗體檢測，但檢查費時且價格並不便宜。因此，目前已有學者開發數種臨床量表評估抗體陽性機率，如APE2 score等等，惟這些量表均沒有腦波資訊。若能透過常規腦電圖（electroencephalography, EEG）篩選抗體陽性個案，預期可加速臨床決策。本前導初步研究以回溯性病例為基礎，收集 2017 年至 2022 年間臺大醫院疑似自體免疫腦炎病人 46 例（其中 12例抗體陽性），評估單次靜息態 EEG 結合機器學習預測抗體陽性的可行性。 整體而言，本研究顯示陰性結果，發現以腦電圖單一模態之機器學習分類效能並不統計顯著優於隨機猜測。這顯示以自體免疫腦炎之單一腦電圖，以常見的特徵擷取與傳統機器學習分析方式的框架下，可能沒有明顯可以泛化(generalize)之特異發現。本研究結果有助釐清腦波在自體免疫腦炎診斷流程中的定位，可能為輔助性而非主要診斷性的角色，並提供後續腦波於免疫腦炎的研究比較基準。

Autoimmune encephalitis and paraneoplastic encephalitis are uncommon yet clinically significant diseases. The diagnostic gold-standard is cerebrospinal-fluid autoantibody or serum antibody testing, a procedure that is time-consuming and costly. Several scoring systems—such as the APE2 score—have been proposed to estimate antibody positivity, but they do not incorporate electroencephalography (EEG) findings. If routine EEG could help identify antibody-positive patients, clinical decision-making might be accelerated. 46 patients who were admitted to National Taiwan University Hospital between 2017 and 2022 for suspected autoimmune encephalitis (AE). Among them, 12 were antibody-positive, making the prevalence 28%. The aim is to examine whether single resting state EEG, analyzed with machine-learning classifiers, could predict antibody status. Overall, this pilot study demonstrates that machine learning models trained with EEG-only data do not perform better than random guessing significantly. These findings may hint that using common feature extraction methods and conventional machine learning algorithms, there may be no apparently generalizable discriminative features in one-time routine EEG. These negative findings suggest that the role of EEG in patients with suspected antibody-related encephalitis may not be directly diagnostic, but supportive. This study also provides a reference baseline for future research on EEG and immune encephalitis.