Tesis

Peranan Kecerdasan Buatan yang Terintegrasi dengan Elektrokardiografi dalam Mendeteksi Penyakit Jantung Katup Stenosis Mitral Terisolasi dan Regurgitasi Mitral Terisolasi Minimal Derajat Sedang = The Role of Artificial Intelligence Integrated with Electrocardiography in Detecting Isolated Mitral Stenosis and Isolated Mitral Regurgitation at least Moderate Severity.

Latar belakang: Penyakit katup mitral merupakan salah satu penyebab utama morbiditas kardiovaskular di Indonesia, dengan dominasi etiologi penyakit jantung reumatik. Deteksi dini sangat bergantung pada ekokardiografi namun keterbatasan akses di layanan kesehatan menjadi hambatan utama. EKG lebih mudah dijangkau tetapi interpretasi manual kurang sensitif untuk mendeteksi penyakit jantung katup. Perkembangan kecerdasan buatan memungkinkan analisis pola non-linier dari sinyal EKG, dan berbagai studi internasional menunjukkan performanya menjanjikan namun bukti pada populasi Indonesia dengan profil demografi dan etiologi yang berbeda masih sangat terbatas. Tujuan: Mengetahui kemampuan model kecerdasan buatan (artificial intelligence/AI) berbasis Convolutional Neuron Network (CNN) arsitektur DenseNet-121 dalam mendeteksi stenosis mitral (SM) terisolasi dan regurgitasi mitral (RM) terisolasi minimal derajat sedang menggunakan data gambar digital 12-lead EKG. Metode: Dilakukan studi diagnostik potong lintang (Harapan Kita-UI DeepValve-ECG) terhadap 5.626 subjek dengan EKG dan ekokardiografi berpasangan ( < 30 hari) di Pusat Jantung Nasional Harapan Kita. Data dikelompokkan menjadi kontrol (3341 subjek), RM terisolasi derajat minimal–sedang (1735 subjek), dan SM terisolasi derajat minimal–sedang (550 subjek). Performa model AI dievaluasi melalui sensitivitas (recall), spesifisitas, Positive Predictive Value (PPV/precision), Negative Predictive Value (NPV), akurasi, F1-score, dan AUC-ROC. Explainable AIdengan analisis Grad-CAM digunakan untuk mengevaluasi fokus perhatian model EKG pada masing-masing kelompok. Hasil: Dilakukan analisis klasifikasi dua tahap, yaitu tahap 1 (kontrol vs RM+SM), model menunjukkan performa seimbang dengan nilai deteksi yang cukup baik yaitu sensitivitas 0,752, spesifisitas 0,752, precision 0,743, F1 -score 0,748, dan AUC-ROC 0,841. Pada tahap 2 (SM vs RM), kemampuan diskriminasi meningkat dengan AUC 0,856 namun terjadi ketimpangan performa antar kelas. Model menunjukkan precision sangat tinggi untuk RM (0,974) dengan F1 - score 0,853, sedangkan SM memiliki precision rendah (0,276) dan F1 -score 0,413. Grad-CAM menunjukkan aktivasi luas dan konsisten pada lead V3–V6 pada RM, sementara SM memperlihatkan aktivasi terbatas pada V1 –V2 dengan pola yang heterogen. Perbedaan fisiologi struktural—volume overload pada RM dan pressure overload atrial serta heterogenitas sinyal pada SM—menjelaskan variasi performa tersebut. Kesimpulan: Model AI memiliki kemampuan diskriminatif yang baik dalam mendeteksi penyakit katup mitral menggunakan data gambar EKG digital. Model cukup akurat untuk penapisan penyakit katup mitral dan mengidentifikasi RM terisolasi namun sensitivitas dan presisi pada SM masih terbatas karena heterogenitas sinyal, rendahnya prevalensi, dan pengaruh komorbid. Model ini berpotensi digunakan sebagai sistem skrining awal atau rule-in tool untuk RM pada fasilitas kesehatan tanpa akses ekokardiografi, dan dapat meningkatkan alur deteksi dini penyakit katup mitral di Indonesia.
Kata Kunci: Elektrokardiografi, Stenosis Mitral, Regurgitasi Mitral, Kecerdasan Buatan, deep learning


Background: Mitral valve disease is one of the leading causes of cardiovascular morbidity in Indonesia, predominantly driven by rheumatic heart disease. Early detection relies heavily on echocardiography; however, limited access in many healthcare settings poses a maj or challenge. While ECG is more widely available, manual interpretation lacks sensitivity for identifying structural heart disease including valve disease. Advances in artificial intelligence (AI) have enabled the extraction of non-linear patterns from ECG signals, and several international studies have demonstrated promising performance. Nevertheless, evidence in the Indonesian population, characterized by different demographic and etiological profiles, remains scarce. Objective: To evaluate the ability of an artificial intelligence (AI) model based on a Convolutional Neural Network (CNN) with DenseNet-121 architecture to detect significant (at least moderate severity) isolated mitral stenosis (MS) and significant isolated mitral regurgitation (MR) using digital 12-lead ECG images. Methods: A cross-sectional diagnostic study (Harapan Kita-UI DeepValve-ECG) was conducted involving 5,626 subjects who had paired ECG and echocardiography data ( < 30 days) at the National Cardiovascular Center Harapan Kita. Subjects were categorized into control (3,341), isolated MR (1,735), and isolated MS (550). Model performance was evaluated using sensitivity (recall), specificity, Positive Predictive Value (PPV/precision), Negative Predictive Value (NPV), accuracy, F1 -score, and AUC-ROC. Explainable AI (Grad-CAM) was applied to visualize model attention on ECG regions for each disease category. Results: A two-stage classification framework was implemented. In stage 1 (control vs MR+MS), the model demonstrated balanced and adequate performance with sensitivity 0.752, specificity 0.752, precision 0.743, F1 -score 0.748, and AUC-ROC 0.841. In stage 2 (MS vs MR), discriminative ability improved with an AUC of 0.856; however, class-wise performance was imbalanced. The model showed very high precision for MR (0.974) with an F1 -score of 0.853, whereas MS showed low precision (0.276) and an F1 -score of 0.413. Grad-CAM revealed broad and consistent activation in leads V3–V6 for MR, while MS showed limited and heterogeneous activation primarily in V1 –V2. These differences correspond to underlying structural physiology—volume overload in MR versus atrial pressure overload and greater signal heterogeneity in MS. Conclusion: The AI model demonstrates good discriminative capability for detecting mitral valve disease using digital ECG images. While sufficiently accurate for screening mitral valve disease and identifying isolated MR, sensitivity and precision for MS remain limited due to signal heterogeneity and lower prevalence. This model has potential utility as an early screening or rulein tool for MR in healthcare facilities without access to echocardiography, and may help enhance early detection pathways for mitral valve disease in Indonesia.
Keywords: Electrocardiography, Artificial Intelligence, Mitral Stenosis, Mitral Regurgitation, Deep Learning