Disertasi

Pengembangan Model Machine Learning dalam Memprediksi Keberhasilan Tata Laksana Anak dengan Epilepsi Resisten Obat = Development of Machine Learning Models to Predict Treatment Outcomes in Children with Drug-Resistant Epilepsy.

Epilepsi merupakan gangguan neurologis dengan sekitar 6–14% anak dapat berkembang menjadi epilepsi resisten obat. Sebanyak 164 pasien anak tercatat sebagai epilepsi resisten obat di Rumah Sakit Cipto Mangunkusumo (RSCM) pada tahun 2020–2025. Salah satu tantangan utama dalam penatalaksanaan epilepsi resisten obat adalah penentuan obat antiepilepsi (OAE) yang sesuai pada populasi anak. Penelitian ini bertujuan untuk memprediksi keberhasilan tata laksana pada anak dengan epilepsi resisten obat. Penelitian dilakukan dengan desain kohort ambispektif menggunakan data rekam medis dan wawancara langsung kepada orang tua atau pengasuh. Sebanyak 215 anak dengan epilepsi resisten obat berusia 1 bulan–18 tahun yang dirawat di RSCM dan Rumah Sakit Anak dan Bunda (RSAB) Harapan Kita pada periode November 2024–Oktober 2025 diikutsertakan sebagai subjek penelitian. Seluruh subjek menjalani pemeriksaan elektroensefalografi (EEG) dan magnetic resonance imaging (MRI) menggunakan protokol modifikasi Harmonized Neuroimaging of Epilepsy Structural Sequences (HARNESS). Pemantauan frekuensi kejang dilakukan selama tiga bulan, dengan kondisi kejang dinyatakan terkontrol apabila terjadi penurunan frekuensi kejang ≥ 75%. Paramater klinis yang digunakan dalam pengembangan model pembelajaran mesin meliputi jenis kelamin, usia saat terdiagnosis epilepsi, jenis epilepsi, jumlah dan kombinasi OAE, hasil EEG dan MRI, serta data frekuensi kejang selama tiga bulan. Lima algoritma machine learning yang digunakan dalam penelitian ini meliputi support vector machine (SVM), decision tree (DT), random forest (RF), extreme gradient boosting (XGBoost), dan regresi logistik. Hasil analisis menunjukkan variasi performa antar model, dengan tiga algoritma terbaik yaitu XGBoost yang memiliki nilai AUC (0,618) dan akurasi (0,567) tertinggi, DT dengan nilai precision (0,662) dan recall (0,540) terbaik, serta RF dengan nilai recall (0,540) dan F1 -score (0,541) tertinggi. Performa gabungan ketiga model terbaik (ensemble) juga menunjukkan nilai akurasi (0,558), precision (0,560), recall (0,636) dan Fl-score (0,596) yang lebih unggul dibandingkan performa model secara individual. Analisis feature importance pada model XGBoost dan RF menunjukkan bahwa usia saat terdiagnosis epilepsi, jenis epilepsi, serta temuan EEG dan MRI merupakan faktor yang paling berpengaruh dalam memprediksi keberhasilan pengobatan. Pendekatan pembelajaran mesin, khususnya menggunakan kombinasi algoritma XGBoost, DT, dan RF, memiliki potensi sebagai alat bantu dalam memprediksi keberhasilan tata laksana epilepsi resisten obat pada anak, sehingga diharapkan dapat dikembangkan menjadi aplikasi untuk membantu pengambilan keputusan klinis dokter spesialis anak neurologi.
Kata kunci : anak-anak, epilepsi resisten obat, Indonesia, pembelajaran mesin


Epilepsy is a neurological disorder with approximately 6–14% of children may develop drug-resistant epilepsy. There were 164 children recorded as having drugresistant epilepsy at Cipto Mangunkusumo Hospital (RSCM) in 2020–2025. One of the main challenges in the management of drug-resistant epilepsy is determining the appropriate antiepileptic drug (AED) for children. This study aims to predict the success of treatment in children with drug-resistant epilepsy. The study was conducted using an ambispective cohort design using medical records and direct interviews with parents or caregivers. As much as 215 children with drug-resistant epilepsy aged 1 month–18 years who were treated at RSCM and Harapan Kita Hospital during the period November 2024–October 2025 were participated as research subjects. All subjects underwent electroencephalography (EEG) and magnetic resonance imaging (MRI) examinations using a modified Harmonized Neuroimaging of Epilepsy Structural Sequences (HARNESS) protocol. Seizure frequency was monitored for three months, with seizures considered controlled if there was a ≥ 75% reduction in seizure frequency. Clinical parameters used in developing the machine learning model included gender, age at epilepsy diagnosis, type of epilepsy, number and combination of AEDs, EEG and MRI results, and seizure frequency data over three months. Five machine learning algorithms were used in this study, including support vector machine (SVM), decision tree (DT), random forest (RF), extreme gradient boosting (XGBoost), and logistic regression. The results of the analysis showed variations in the model's performance, with the three best algorithms were XGBoost, which had the highest AUC (0.618) and accuracy (0.567), DT with the best precision (0.662) and recall (0.540), and RF with the highest recall (0.540) and F1 -score (0.541). The combination the three best performing models (ensemble) also demonstrated superior results, achieving an accuracy (0.558), precision (0.560), recall (0.636), and an F1 -score (0.596) compared with the performance of the individual models. Feature importance analysis on the XGBoost and RF models showed that age at epilepsy diagnosis, type of epilepsy, and EEG and MRI findings were the most impactful factors in predicting treatment success. This study concluded that the machine learning approach, particularly using a combination of XGBoost, DT, and RF algorithms, had the potential to be a tool in predicting the success of drug-resistant epilepsy management in children. Therefore, this may be further developed into a software application to assist clinical decision making by pediatric neurologists.
Keywords: drug-resistant epilepsy, Indonesia, machine learning, pediatrics