<![CDATA[Abstrak – Prevalensi risiko komplikasi pada penderita mood disorder mengalami peningkatan yang berdampak terhadap reduksi fungsi sehari-hari dan kualitas hidup penderita gangguan depresi ini. Perluasan skirining dini dibutuhkan untuk menekan risiko komplikasi ditahap awal-awal gejala. Namun, pendekatan identifikasi konvensional melalui wawancara klinis dan laporan pasien masih bersifat subjektif dan terbatas karena ketergantungan terhadap praktisi ahli. Integrasi pengukuran aktivitas otak melalui EEG dalam skrining gejala mood disorder menjadi solusi untuk mengatasi tantangan subjektivitas dan keterbatasan tenaga ahli. Namun, kompleksitas sinyal EEG terkait dengan dimensi fitur yang tinggi dan informasi yang redundan membutuhkan teknik pengolahan sinyal yang tepat agar memberikan hasil klasifikasi yang optimal. Penelitian ini menganalisis pengaruh seleksi fitur mutual information (MI), minimum redundancy maximum relevance (mRMR), heuristic joint mutual information (HJMI), dan TOPSIS terhadap hasil klasifikasi healthy control dan mood disorder. Penelitian menggunakan dataset yang terdiri dari 266 sampel mood disorder dan 95 healthy control dengan 1.140 fitur EEG pada model klasifikasi SVM. Berdasarkan hasil pengujian, seluruh metode mampu mereduksi jumlah fitur secara signifikan, sedangkan MI dan mRMR memberikan akurasi lebih tinggi dibandingkan baseline 77,98% (model SVM tanpa feature engineering). Performa terbaik diperoleh mRMR dengan 200 fitur, menghasilkan accuracy 80,73%, precision 84,71%, recall 90,00%, dan F1-score 87,27%, serta reduksi fitur sebesar 82,46%. Hasil ini menunjukkan bahwa seleksi fitur dapat menghasilkan representasi data yang lebih ringkas sekaligus meningkatkan akurasi klasifikasi mood disorder, dengan mRMR sebagai metode terbaik pada penelitian ini. Kata kunci: EEG; Mood Disorder; Seleksi Fitur; mRMR; Mutual Information; TOPSIS; SVM; Abstract - The prevalence of complication risks among individuals with mood disorders has increased, negatively impacting their daily functionality and quality of life. Expanding early screening is necessary to mitigate these risks at the onset of symptoms. However, conventional identification approaches relying on clinical interviews and self-reports remain subjective and limited due to their dependence on expert practitioners. Integrating brain activity measurement via EEG into mood disorder symptom screening offers a potential solution to address these challenges of subjectivity and expert scarcity. Nevertheless, the complexity of EEG signals, characterized by high feature dimensionality and redundant information, requires appropriate signal processing techniques to achieve optimal classification outcomes. This study analyzes the impact of feature selection methods, namely Mutual Information (MI), Minimum Redundancy Maximum Relevance (mRMR), Heuristic Joint Mutual Information (HJMI), and TOPSIS, on the classification performance between healthy controls and mood disorder subjects. The research utilizes a dataset comprising 266 mood disorder samples and 95 healthy controls with 1,140 EEG features using a Support Vector Machine (SVM) classification model. Experimental results demonstrate that all feature selection methods significantly reduced the feature dimension, while MI and mRMR achieved higher accuracy compared to the baseline model (SVM without feature engineering at 77.98%). The best performance was achieved by mRMR using 200 features, yielding an accuracy of 80.73%, precision of 84.71%, recall of 90.00%, and F1-score of 87.27%, along with an 82.46% reduction in features. These findings indicate that feature selection can produce a more concise data representation while simultaneously improving mood disorder classification accuracy, with mRMR emerging as the optimal method in this study. Keywords: EEG; Mood Disorder; Feature Selection; mRMR; Mutual Information; TOPSIS; SVM;]]>
