| 摘要: | 臨床變數(Clinical Variables)是疾病診斷和治療效果預測的重要依據，然而臨床變數有其侷限性，像是重鬱症(Major depressive disorder, MDD)和癲癇(Epilepsy)這類常見精神或神經疾病，臨床變數無法有效反應大腦的情況。
本研究從腦電圖(Electroencephalography)中提取功能網路(Functional Connectivity)和功率(Power)，並結合臨床變數，使用留一交叉驗證(Leave-One-Out Cross-Validation, LOOCV)和機器學習的分類器進行訓練，應用於三個子題目，分別為「重鬱症之診斷」、「重鬱症之療效預測」和「頑固性癲癇之療效預測」。
「重鬱症之診斷」的目標為「透過腦電圖和臨床變數輔助診斷重鬱症」，共蒐集到77位病患和46健康者的腦電圖以及臨床變數。最好的診斷效果為絕對功率(Absolute Power)配上Medium KNN時，精確率(Precision)為0.881、召回率(Recall)為0.961、特異性(Specificity)為0.783、準確率(Accuracy)為0.894、F1 Score為0.919、曲線下面積(Area Under Curve, AUC)為0.876。
「重鬱症之療效預測」的目標為「透過腦電圖和臨床變數輔助預測重鬱症患者接受治療後4、6和8週的療效」，共蒐集到77位病患治療前的腦電圖、治療後一週的腦電圖以及臨床變數。
預測治療後4週的療效時，最好的診斷效果為治療後一週腦電圖(W1)的相位鎖定值(Phase Locking Value，PLV)和Absolute Power配上Fine Tree或Medium Tree時，Precision為0.704、Recall為0.792、Specificity為0.849、Accuracy為0.831、F1 Score為0.745、AUC為0.818。
預測治療後6週的療效時，最好的診斷效果為治療後一週腦電圖特徵減去治療前腦電圖特徵(W1-W0)的PLV、相對功率(Relative Power)和臨床變數配上Coarse Gaussian SVM時，Precision為0.724、Recall為0.724、Specificity為0.742、Accuracy為0.733、F1 Score為0.724、AUC為0.772。
預測治療後8週的療效時，最好的診斷效果為治療後一週腦電圖特徵減去治療前腦電圖特徵(W1-W0)的Absolute Power和臨床變數配上Coarse Tree時， Precision為0.675、Recall為0.849、Specificity為0.519、Accuracy為0.700、F1 Score為0.757、AUC為0.790。
「頑固性癲癇之療效預測」的目標為「過腦電圖輔助預測頑固性癲癇患者接受生酮飲食治療後3和6個月的療效」，共蒐集到90位病患的腦電圖。
預測治療後3個月的療效時，最好的診斷效果為PLV和Relative Power配上Fine Tree或Medium Tree時，Precision為0.831、Recall為0.817、Specificity為0.667、Accuracy為0.767、F1 Score為0.824、AUC為0.760。
預測治療後6個月的療效時，最好的診斷效果為PLV和Absolute Power配上Linear SVM時，Precision為0.800、Recall為0.963、Specificity為0.000、Accuracy為0.776、F1 Score為0.874、AUC為0.752。;Clinical variables are crucial for diagnosing diseases and predicting treatment outcomes. However, they have limitations, particularly for common psychiatric or neurological disorders like Major Depressive Disorder (MDD) and Epilepsy, where clinical variables cannot effectively reflect the brain′s condition.
This study extracts functional connectivity and power from electroencephalography (EEG) and combines these with clinical variables. The data are then trained using Leave-One-Out Cross-Validation (LOOCV) and machine learning classifiers, applied to three sub-topics: "Diagnosis of Major Depressive Disorder," "Prediction of Treatment Efficacy for Major Depressive Disorder," and "Prediction of Treatment Efficacy for Refractory Epilepsy.
The goal of the "Diagnosis of Major Depressive Disorder" is to "assist in diagnosing MDD through EEG and clinical variables." A total of 77 patients and 46 healthy individuals′ EEG and clinical variables were collected. The best diagnostic result was achieved using Absolute Power with Medium KNN, yielding a Precision of 0.881, Recall of 0.961, Specificity of 0.783, Accuracy of 0.894, F1 Score of 0.919, and AUC of 0.876.
The goal of the "Prediction of Treatment Efficacy for Major Depressive Disorder" is to "assist in predicting the efficacy of treatment after 4, 6, and 8 weeks using EEG and clinical variables." EEGs before treatment, EEGs one week after treatment, and clinical variables from 77 patients were collected.
For predicting treatment efficacy at 4 weeks, the best result was achieved using W1 EEG PLV and Absolute Power with Fine Tree or Medium Tree, yielding a Precision of 0.704, Recall of 0.792, Specificity of 0.849, Accuracy of 0.831, F1 Score of 0.745, and AUC of 0.818.
For 6 weeks, the best result was achieved using W1-W0 EEG PLV, Relative Power, and clinical variables with Coarse Gaussian SVM, yielding a Precision of 0.724, Recall of 0.724, Specificity of 0.742, Accuracy of 0.733, F1 Score of 0.724, and AUC of 0.772.
For 8 weeks, the best result was achieved using W1-W0 EEG Absolute Power and clinical variables with Coarse Tree, yielding a Precision of 0.675, Recall of 0.849, Specificity of 0.519, Accuracy of 0.700, F1 Score of 0.757, and AUC of 0.790.
The goal of the "Prediction of Treatment Efficacy for Refractory Epilepsy" is to "assist in predicting the efficacy of ketogenic diet treatment after 3 and 6 months using EEG." EEGs from 90 patients were collected.
For predicting treatment efficacy at 3 months, the best result was achieved using PLV and Relative Power with Fine Tree or Medium Tree, yielding a Precision of 0.831, Recall of 0.817, Specificity of 0.667, Accuracy of 0.767, F1 Score of 0.824, and AUC of 0.760.
For 6 months, the best result was achieved using PLV and Absolute Power with Linear SVM, yielding a Precision of 0.800, Recall of 0.963, Specificity of 0.000, Accuracy of 0.776, F1 Score of 0.874, and AUC of 0.752. |
