Ensemble Of Xgboost Berbasis Shap Untuk Klasifikasi Interpretable Tingkat Depresi Berdasarkan Konektivitas Eeg Pada Terapi Mindfulness

Agustin, Ela (2026) Ensemble Of Xgboost Berbasis Shap Untuk Klasifikasi Interpretable Tingkat Depresi Berdasarkan Konektivitas Eeg Pada Terapi Mindfulness. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6003241011-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Deteksi dini dan klasifikasi tingkat keparahan depresi merupakan tantangan besar dalam kesehatan mental. Penelitian ini bertujuan untuk mengembangkan metode klasifikasi tingkat depresi yang objektif, akurat, dan dapat diinterpretasikan (interpretable) berbasis data electroencephalogram (EEG) untuk mengevaluasi efektivitas terapi Mindfulness Based Cognitive Therapy (MBCT). Penelitian ini melibatkan 32 responden yang dikategorikan ke dalam kelompok normal, depresi tingkat 1, dan depresi tingkat 2 berdasarkan hasil clustering. Preprocessing sinyal EEG dilakukan melalui tahapan Butterworth Bandpass Filter, Artifact Subspace Reconstruction, dan Independent Component Analysis. Ekstraksi fitur dilakukan menggunakan pendekatan Functional Connectivity (FC) dengan metrik Pearson Correlation (PC), Mutual Information (MI), dan Phase Lag Index (PLI) pada lima subband frekuensi dan pasangan antar channel, menghasilkan total 2.565 fitur. Melalui metode seleksi fitur hibrida MI dan Grey Wolf Optimizer-Extreme Gradient Boosting (GWO XGBoost), dimensi fitur direduksi menjadi 146 fitur paling relevan. Klasifikasi dilakukan menggunakan metode Ensemble of Extreme Gradient Boosting (EoXGBoost) dengan skema nested cross-validation. Hasil penelitian menunjukkan performa model yang tinggi dengan akurasi (98,80%), sensitivitas (98,76%), presisi (98,89%), spesifisitas (99,34%), F1-score (98,82%) dan AUC (99,97%). Analisis SHAP menunjukkan bahwa fitur paling berpengaruh adalah F8_Cz_gamma_MI, dikuti konektivitas oksipital-temporal dan frontaltemporal. Selain itu, fitur berbasis MI pada subband gamma dan alpha mendominasi fitur-fitur dengan kontribusi tertinggi dalam membedakan kelas normal, depresi tingkat 1, dan depresi tingkat 2. Evaluasi terapi menunjukkan bahwa 27% responden pada kelompok perlakuan MBCT mengalami perbaikan tanpa ada yang memburuk, dibandingkan kelompok kontrol yang memiliki risiko memburuk sebesar 11%. Namun, uji statistik menunjukkan MBCT tidak berpengaruh secara signifikan dalam menurunkan tingkat depresi.
===================================================================================================================================
Early detection and classification of the severity of depression are major challenges in mental health. This study aims to develop an objective, accurate, and interpretable method for classifying depression severity based on electroencephalogram (EEG) data to evaluate the effectiveness of MindfulnessBased Cognitive Therapy (MBCT). The study involved 32 participants who were categorized into normal, level 1 depression, and level 2 depression groups based on clustering results. EEG signal preprocessing was performed through the following stages: Butterworth Bandpass Filter, Artifact Subspace Reconstruction, and Independent Component Analysis. Feature extraction was conducted using the Functional Connectivity (FC) approach with the Pearson Correlation (PC), Mutual Information (MI), and Phase Lag Index (PLI) metrics across five frequency subbands and channel pairs, resulting in a total of 2,565 features. Using a hybrid feature selection method combining MI and the Grey Wolf Optimizer-Extreme Gradient Boosting (GWO-XGBoost), the feature dimension was reduced to the 146 most relevant features. Classification was performed using the Ensemble of Extreme Gradient Boosting (EoXGBoost) method with a nested cross-validation scheme. The results demonstrate high model performance with an accuracy of 98.80%, sensitivity of 98.76%, precision of 98.89%, specificity of 99.34%, an F1-score of 98.82%, and an AUC of 99.97%. The SHAP analysis showed that the most influential feature was F8_Cz_gamma_MI, followed by occipital-temporal and frontal-temporal connectivity. In addition, MI-based features in the gamma and alpha subbands dominated the features with the highest contribution to distinguishing between the normal, Level 1 depression, and Level 2 depression classes. Therapy evaluation showed that 27% of respondents in the MBCT treatment group experienced an improvement with no one experiencing a worsening, compared to the control group, which had an 11% risk of worsening. However, statistical tests showed that MBCT did not have a significant effect on reducing depression levels.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	EoXGBoost, Functional Connectivity, GWO-XGBoost, SHAP
Subjects:	Q Science Q Science > Q Science (General) Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines.
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Statistics > 49101-(S2) Master Thesis
Depositing User:	Ela Agustin
Date Deposited:	01 Aug 2026 06:24
Last Modified:	01 Aug 2026 06:24
URI:	http://repository.its.ac.id/id/eprint/141115

Actions (login required)

View Item