以深度學習揭示腦電訊號之專業差異神經特徵：高爾夫新手與業餘選手之比較研究

Abstract

高爾夫推桿提供了一個理想的運動神經科學研究目標：其動作靜止、時程短暫，且可精確鎖定至客觀可驗證之事件（例如球入洞），便於腦電圖（EEG）的訊號分析。儘管既有研究已記錄到專業與初學高爾夫球員之間的群體層級頻譜差異，卻常依賴傳統的人工設計之頻帶特徵、較少構建個人層級的分類器，亦或未評估模型對未見受試者之跨受試者泛化能力。

本研究將 EEGNet——一種緊湊型卷積神經網路——應用於 41 名高爾夫球員（業餘組 18 名、初學組 23 名）之 EEG 資料集，共計 1,997 筆成功推桿試次，以30 個腦電電極、250 Hz 採樣率記錄。每筆試次為以推桿擊球瞬間為基準點之前 2 秒 EEG 片段，分類目標為受試者技能組別（初學者 vs. 業餘選手）。本研究採用兩種互補之評估方法：組內評估（試次層級隨機切分，訓練集與測試集來自相同受試者群體）以及留一受試者交叉驗證（LOSO），以系統性量化受試者身份混淆因子之影響。

組內評估結果顯示 EEGNet 達到 Cohen's κ = 0.926、準確率 96.3%、AUC = 0.995；而 LOSO 評估下，性能驟降至 κ = −0.043、準確率 46.5%、AUC = 0.502，與隨機猜測無異。兩者之差距 Δκ = 0.969，直接揭示了受試者身份混淆效應：模型所學習的是個人特有的 EEG 特徵，而非可推廣至新受試者的技能層級神經特徵。在頻率分析結果顯示，業餘選手在所有五個標準頻帶（Delta、Theta、Alpha、Beta、Gamma）之 EEG 功率均顯著高於初學者，其中 Beta 之效應最強（d = −0.341，p < 10⁻¹³）、Alpha 則次之（d = −0.303）。此結果在動作技能習得理論框架下獲得合理解釋——業餘選手處於聯結性（associative）學習階段，仍需持續的有意識認知投入，故呈現較高的整體皮質活化程度；此模式與神經效率假說（Neural Efficiency Hypothesis）在專家與新手對比下所預測之方向不完全一致，但更貼近心理動作效率假說（Psychomotor Efficiency Hypothesis）所強調之選擇性、區域性神經重組，而非整體活化程度的單純增減。模型可解釋性分析進一步顯示，EEGNet 所學習之時域濾波器自動分化為不同頻帶之偵測器，其中 Theta 頻帶（4–8 Hz）特化之濾波器與初學者認知負荷之神經機制相符；空間濾波器視覺化則顯示，各濾波器之通道加權持續集中於額葉、中央區與頂葉電極，與運動專業文獻所預測之關鍵腦區高度吻合。

本研究為首次將深度學習應用於高爾夫推桿技能 EEG 分類之研究，並論證了雙評估設計的必要性：在運動神經科學腦機介面研究中，同時呈現組內評估與跨受試者評估結果，是確保評估結論的基本要求。

Golf putting provides a well-controlled paradigm for studying neural correlates of motor expertise: the stroke is stationary, brief, and time-locked to a verifiable outcome. Prior electroencephalography (EEG) studies have documented group-level spectral differences between expert and novice golfers, yet all relied on handcrafted spectral features and reported only group-level statistics, leaving open the question of whether a data-driven classifier can generalise to unseen individuals.

This study applies EEGNet — a compact convolutional neural network with approximately 2,000 trainable parameters — to a dataset of 41 golfers (18 Amateur, 23 Novice), comprising 1,997 successful putting trials recorded across 30 EEG channels at 250 Hz. Each trial is a two-second epoch time-locked to putter-ball contact, and the classification target is expertise group (Novice vs. Amateur). Two complementary evaluation protocols are employed: a subject-dependent trial-level split, in which trials from all subjects are pooled before partitioning; and leave-one-subject-out (LOSO) cross-validation, in which each subject's trials are withheld entirely from training.

Under subject-dependent evaluation, EEGNet achieves κ = 0.926, accuracy = 96.3%, and AUC = 0.995. Under LOSO, performance collapses to κ = −0.043, accuracy = 46.5%, and AUC = 0.502 — indistinguishable from chance. The Δκ of 0.969 directly quantifies the subject-identity confound: the model learns stable, individual-specific EEG fingerprints that co-vary with the group label within the observed cohort but do not correspond to generalisable expertise-level features. Supplementary group-level frequency analysis confirms statistically significant band-power differences between groups across all five canonical EEG bands, with Amateur golfers exhibiting higher global power than Novices — most strongly in beta (d = −0.341, p < 10⁻¹³) and alpha (d = −0.303). This pattern departs from the direction predicted by the Neural Efficiency Hypothesis for expert-novice contrasts, and is interpreted instead as reflecting active motor refinement at the associative stage of skill acquisition, more consistent with the Psychomotor Efficiency Hypothesis's emphasis on selective, regionally-organised neural reallocation rather than a uniform increase or decrease in activation. Model interpretability analyses further reveal that EEGNet's learned temporal filters differentiate into band-selective detectors — including a Theta-selective filter (4–8 Hz) consistent with the novice cognitive load signature — while spatial filter visualisation shows that channel weights are consistently concentrated in frontal, central, and parietal electrodes, aligning with the brain regions implicated in the motor expertise literature.

This work represents the first deep learning application to golf expertise EEG classification and demonstrates that dual-protocol evaluation — reporting both subject-dependent and cross-subject performance — is essential for honest assessment in sports neuroscience brain-computer interface research.