Pengembangan Model Machine Learning untuk Memprediksi Kematian selama Perawatan pada Pasien Stroke Hemoragik Menggunakan Data Terstruktur dari Registry of Stroke Care Quality (RES-Q)

Ika Agustin Atika Putri, Annisa Ristya Rahmanti, S.Gz, Dietisien, MS.,Ph.D ; Dr. dr. Guardian Yoki Sanjaya, MHlthInfo

2026 | Tesis | MAGISTER KEBIJAKAN DAN MANAJEMEN KESEHATAN

Latar Belakang: Stroke hemoragik merupakan kondisi neurologis akut yang memiliki tingkat mortalitas tertinggi di antara jenis stroke lainnya. Identifikasi dini terhadap risikoin-hospital mortality sangat penting untuk meningkatkan keselamatan pasien dan efektivitas intervensi klinis. Metode prediksi konvensional masih terbatas dalam akurasi dan fleksibilitas, sehingga diperlukan pendekatan machine learning (ML) untuk meningkatkan ketepatan prediksi.

Tujuan: Penelitian ini bertujuan untuk mengembangkan dan mengevaluasi model prediksi in hospital mortality pada pasien stroke hemoragik dengan menggunakan data terstruktur dari Registry of Stroke Care Quality (RES-Q) berbasis algoritma ML.

Metode: Studi ini menggunakan desain retrospektif kohort dengan mengambil data dari RSA UGM dan RSUP Dr. Sardjito (Indonesia) serta data multisenter Vietnam periode 2020–2026. Dataset Vietnam (6.189 pasien) digunakan sebagai data latih, sedangkan dataset Indonesia (144 pasien) sebagai validasi eksternal. Empat algoritma ML digunakan, yaitu Logistic Regression, Random Forest, Support Vector Machine (SVM), dan XGBoost. Evaluasi model dilakukan menggunakan accuracy, precision, recall, F1-score, dan AUC-ROC.

Hasil: Dataset Vietnam memiliki mortalitas 27,1?n dataset Indonesia 15,3%, dengan perbedaan jumlah dan distribusi outcome yang memengaruhi generalisasi model. Pada data training Vietnam, Random Forest menunjukkan performa terbaik (AUC 0,932), sedangkan pada validasi eksternal, SVM menunjukkan performa terbaik (AUC 0,747). Analisis SHAP mengidentifikasi NIHSS score sebagai prediktor mortalitas paling dominan, diikuti penggunaan ventilator, GCS score, GCS level mild injury, CT angiography (CTA), fisioterapi, arrival mode EMS, prestroke mRS, physiotherapy not required, dan usia. Temuan ini menunjukkan bahwa tingkat keparahan neurologis, status fungsional, dukungan ventilasi, dan faktor tata laksana klinis berperan penting dalam memprediksi mortalitas selama perawatan.

Kesimpulan: Model machine learning berbasis data RES-Q berpotensi menjadi alat bantu klinis dalam memprediksi mortalitas stroke hemoragik secara lebih akurat. Penelitian ini diharapkan mendukung pengembangan Clinical Decision Support System (CDSS) berbasis ML di layanan kesehatan.

Background: Hemorrhagic stroke is an acute neurological condition with the highest mortality rate among all stroke subtypes. Early identification of in-hospital mortality risk is essential to improve patient safety and clinical intervention effectiveness. Conventional prediction methods remain limited in terms of accuracy and flexibility, thus requiring a more adaptive approach using machine learning (ML) to improve predictive performance.

Objective: This study aimed to develop and evaluate a machine learning-based model for predicting in-hospital mortality in hemorrhagic stroke patients using structured data from the Registry of Stroke Care Quality (RES-Q).
Methods: This retrospective cohort study utilized data from RSA UGM and RSUP Dr. Sardjito (Indonesia), as well as a multicenter Vietnam dataset covering the period 2020–2026. The Vietnam dataset (6,189 patients) was used as the training set, while the Indonesian dataset (144 patients) was used for external validation. Four machine learning algorithms were applied, including Logistic Regression, Random Forest, Support Vector Machine (SVM), and XGBoost. Model performance was evaluated using accuracy, precision, recall, F1-score, and AUC- ROC.

Results: The Vietnam dataset had a mortality rate of 27.1%, while the Indonesian dataset had a mortality rate of 15.3%. Differences in sample size and outcome distribution affected model generalizability. In the Vietnam training dataset, Random Forest achieved the best performance (AUC 0.932), while SVM demonstrated the best performance in external validation (AUC 0.747). SHAP analysis identified NIHSS score as the most influential predictor of mortality, followed by ventilator use, GCS score, mild GCS injury category, CT angiography (CTA), physiotherapy, arrival by emergency medical services (EMS), prestroke mRS, physiotherapy not required status, and age. These findings highlight the importance of neurological severity, functional status, ventilatory support, and clinical care processes in predicting in-hospital mortality.

Conclusion: Machine learning models based on RES-Q data show potential as clinical decision support tools for more accurate prediction of hemorrhagic stroke mortality. This study supports the development of machine learning-based Clinical Decision Support Systems (CDSS) in healthcare settings.

Kata Kunci : hemorrhagic stroke, in-hospital mortality, Registry of Stroke Care Quality (RES-Q), machine learning, predictive model, algorithms.