透過可解釋的認知評估，研究轉移、少樣本和元學習在基於靜息態和聽覺穩定狀態反應腦電圖的神經退化性疾病診斷之生物標記識別

Abstract

失智症患者的靜息態和聽覺穩定狀態反應 (ASSR) 相關腦電圖 (EEG) 訊號，如腦電圖減慢、伽瑪相位鎖定和功率調變改變，通常能揭露患者的認知異常。傳統上，腦電圖的分析涉及使用功率譜密度(PSD) 地形圖、試驗間一致性(ITC) 計算和事件相關頻譜擾動(ERSP) 分析等技術將訊號轉換為空間頻率和時間頻率之影像表示。這樣的方法可以凸顯關鍵的細微差別以進行準確的診斷和解釋。隨著深度學習 (DL) 在推進診斷的關鍵解決方案展露頭角，透過可解釋的人工智慧 (xAI) 進行，它為傳統方法提供了一種可行、高效且重要的替代方案。而深度學習也揭示了腦電圖作為篩檢工具並協助找出神經退化性疾病個體生物標記的能力。 在這項研究中，我們專注於整合遷移學習、小樣本學習 (FSL) 和元學習技術，以最大化深度學習的潛力。我們收集了阿茲海默症 (AD) 、罕見疾病額顳葉失智症 (FTD) 和精神分裂症 (SZ)患者的腦電圖數據，以研究與失智症相關的差異特徵。閉眼時的靜息態腦電圖資料收集自 15 位AD患者、15 位 FTD 患者、16位SZ 患者以及 15位健康對照組 (HC)。此外，針對 ASSR 數據集我們收集了以40 Hz 刺激的數據，其中包括 10 名SZ 患者和 15 名 HC。 我們的方法採用預訓練的捲積神經網路 (CNN) 以及 miniImageNet 資料集來有效地分析不同腦電圖資料集。我們結合了FSL 和元學習，以增強模型在2-way 5-shot分類中的泛用性和適應性，其中每個episode 涉及從兩個類別中隨機採樣五個樣例，並在100 多個episode的穩健測試中對三個查詢進行測試。儘管神經退化性疾病研究中的數據可用性是個重大挑戰，而且我們的數據集規模相對較小，但相似的人類上皮細胞 2 型 (HEp-2 Cell) 數據集和血細胞計數和檢測 (BCCD）白血球數據集達到顯著的進步。 其中，HEp-2 細胞資料集包含13,596 張影像，分為六類（著絲粒、高爾基體、均質、核仁、核膜和斑點型），而BCCD 白血球資料集包括五類（嗜中性球、淋巴細胞、單核細胞、嗜酸性粒細胞和嗜鹼性粒細胞），總共12,436張影像，達到了55.1%的準確率。 相對 α 波段功率是靜息態腦電圖最顯著的特徵。使用含有殘差網路18 (ResNet-18) 的原型網路(prototypical network)和帶有ResNet-34 編碼器的匹配網路(matching network)，我們分別實現了82.83% 和76.33% 的平均準確度。其中最具辨別能力的類別對是AD 和SZ（94.44%和 92.22% 的準確率）。相較之下，我們的結果表明，P4 和 Fz 通道在 ASSR 間相干性方面實現了最高性能，平均準確度分別為 68.83% 和 70.00%。另一方面，與事件相關的光譜擾動發現T5 和 Fz 通道是有區別性的，兩個模型的平均表現為 71.17% 和 69.33%。 為了更深入研究神經動力學，我們實作了梯度加權類活化映射 (Grad-CAM)，這是 xAI 中的一個突出方法。這種方法大大增強了我們對疾病狀態和病理機制中振盪行為的理解，並可能有助於生物標記的辨識。 Grad-CAM 顯示頂葉的相對 α 功率降低，以及與靜止狀態下認知障礙的潛在關聯性。此外，基於ASSR 的Grad-CAM視覺化揭示了誘發功率和鎖相中明顯的伽馬帶振盪，表明N-甲基-D-天冬氨酸受體(NMDAR) 拮抗劑可能引起伽馬氨基丁酸(GABA) 中間神經元功能障礙。證據顯示，在 ASSR 任務期間，Fz 通道 ITC和ERSP在Grad-CAM 上減少，顯示 SZ 和 HC 受試者存在異常的神經同步和頻譜動態模式。這些神經損傷的模式會影響神經網路的完整性，而能揭示特定的大腦區域如何導致疾病。 總之，結合傳統腦電圖分析方法和新穎的深度學習技術使我們能夠解決神經退化性疾病研究中的關鍵挑戰。我們的方法不僅提高了其穩健性和有效性，還提高了複雜神經數據的可解釋性。因此，我們的方法能啟發新的臨床方案發展，以及加速新藥標的的發現。

Resting-state and auditory steady-state response (ASSR)-related electroencephalography (EEG) signals in individuals with dementia commonly reveal cognitive abnormalities, such as EEG slowing, gamma phase locking, and power modulation alterations. Traditionally, EEG analysis involves transforming signals into spatial frequency and time-frequency image representations using techniques via power spectral density (PSD) topographic maps, intertrial coherence (ITC) calculations, and event-related spectral perturbation (ERSP) analysis, highlighting crucial nuances essential for accurate diagnosis and interpretation. With the emergence of deep learning (DL) as a pivotal solution in advancing diagnostics, it offers a feasible, efficient, and crucial alternative to conventional methods for accurate differentiation, complemented by interpretation through explainable artificial intelligence (xAI). DL methods, in particular, reveal the ability of EEG to serve as a screening tool and assist in identifying biomarkers in individuals with neurodegenerative diseases. In this study, we focus on integrating transfer learning, few-shot learning (FSL), and meta-learning techniques to maximize the potential of DL. EEG data from individuals with conditions such as Alzheimer’s disease (AD), a rare disease frontotemporal dementia (FTD), and schizophrenia (SZ) were collected to investigate differential features associated with dementia. Resting-state EEG data from the eyes-closed protocol were collected from 15 patients diagnosed with AD, 15 with FTD, 16 with SZ, as well as from 15 healthy controls (HCs). In addition, 40-Hz stimulus data for the ASSR dataset, consisting of 10 individuals with SZ and 15 HCs, were obtained. Our approach adopted pretrained convolutional neural networks (CNNs) along with the miniImageNet dataset benchmark to analyze diverse EEG datasets effectively. We incorporated FSL and meta-learning to enhance the models’ generalizability and rapid adaptability in 2-way 5-shot classification, where each episode involved randomly sampling five support examples from two classes and robust testing on three queries over 100 episodes. Despite major challenges with data availability in neurodegenerative disease research and the relatively small size of our datasets, significant improvements have been observed in similar test scenarios involving the Human Epithelial Cells type 2 (HEp-2 Cell) dataset and the Blood Cell Count and Detection (BCCD) White Blood Cell dataset. The HEp-2 Cell dataset comprises 13,596 images divided into six classes (Centromere, Golgi, Homogeneous, Nucleolar, Nuclear Membrane, and Speckled). In contrast, the BCCD White Blood Cell dataset includes five classes (Neutrophil, Lymphocyte, Monocyte, Eosinophil, and Basophil), totaling 12,436 images, achieving its highest accuracy of 55.1%. Relative alpha band power is the most distinguishing feature of resting-state EEG. Using a prototypical network with residual network-18 (ResNet-18) and a matching network with ResNet-34 encoders, we achieved mean accuracies of 82.83% and 76.33%, respectively, with the most discriminative class pair being AD and SZ (94.44% and 92.22% accuracies). In contrast, our results indicate that the P4 and Fz channels achieve the highest performance in ASSR intertrial coherence, with average accuracies of 68.83% and 70.00%, respectively. On the other hand, event-related spectral perturbations identify the T5 and Fz channels as discriminative, averaging 71.17% and 69.33% for both models. To delve deeper into neural dynamics, we implemented gradient-weighted class activation mapping (Grad-CAM), a prominent feature attribution in xAI. This method has greatly enhanced our understanding of oscillatory behavior in disease states and pathological mechanisms, potentially aiding in biomarker identification. The Grad-CAMs showed decreased relative alpha power in the parietal lobes and its possible association with cognitive impairment in the resting state. Moreover, ASSR-based Grad-CAM visualizations revealed distinct gamma band oscillations in evoked power and phase locking, suggesting probable gamma-aminobutyric acid (GABA)ergic interneuron dysfunction caused by N-methyl-D-aspartate receptor (NMDAR) antagonists. The evidence showed a reduction in Fz channel ITC and ERSP Grad-CAMs during the ASSR task, indicating aberrant neural synchrony and spectral dynamics patterns across SZ and HC subjects. These patterns underlying neurological impairments affect neural network integrity, thereby revealing how specific brain regions contribute to diseases. In conclusion, combining traditional EEG analysis methods and innovative DL techniques has allowed us to address critical challenges in neurodegenerative disease research. Our methodology improves its robustness and effectiveness and the interpretability of complex neural data. Thus, our approach may inspire the development of new clinical practices and expedite the discovery of new drug targets.