Penambangan Data Berbasis Pembelajaran Mesin untuk Pengelompokan Wilayah Beresiko Stunting di Kabupaten Bandung

Abstract

Stunting merupakan masalah pertumbuhan yang bersifat multidimensional sehingga penentuan prioritas wilayah memerlukan pembacaan serentak atas faktor kesehatan, lingkungan, layanan, dan demografi. Penelitian ini bertujuan mengelompokkan 31 kecamatan di Kabupaten Bandung berdasarkan kemiripan pola kerentanan terkait stunting. Data sekunder tahun 2024 dihimpun dari Profil Kesehatan Kabupaten Bandung dan Satu Data Kabupaten Bandung, sedangkan geometri kecamatan menggunakan batas administrasi Badan Informasi Geospasial edisi 2022. Tujuh fitur yang digunakan adalah prevalensi berat badan lahir rendah, sanitasi tidak aman, air minum tidak layak, rumah tangga non-PHBS, rasio Posyandu, rasio puskesmas, dan logaritma kepadatan penduduk. Seluruh fitur distandardisasi dan diarahkan agar nilai Z lebih tinggi menunjukkan kerentanan relatif lebih besar. Algoritma K-Means diuji untuk k=2–6. Pemilihan jumlah klaster mempertimbangkan elbow, silhouette, Davies–Bouldin, Calinski–Harabasz, keseimbangan ukuran, stabilitas 100 seed, dan leave-one-out. Solusi k=2 dipilih dengan silhouette 0,222, stabilitas seed ARI 1,000, dan ukuran klaster 15 serta 16 kecamatan. Klaster A menunjukkan pola kepadatan–air minum–keterbatasan layanan, sedangkan Klaster B menunjukkan pola BBLR–perilaku kesehatan. Prevalensi stunting tertimbang masing-masing klaster adalah 7,76% dan 10,19%. Namun, perbedaan distribusi prevalensi kecamatan tidak signifikan (Mann–Whitney U=139,0; p=0,465; Cliff’s delta=0,158). Hasil ini mendukung penggunaan klaster sebagai tipologi eksploratif untuk merancang intervensi yang berbeda antar pola, bukan sebagai klasifikasi risiko tinggi dan rendah yang bersifat definitif.

Stunting is a multidimensional growth problem; therefore, determining priority areas requires a simultaneous assessment of health, environmental, service, and demographic factors. This study aims to classify 31 sub-districts in Bandung Regency based on similarities in stunting-related vulnerability patterns. Secondary data for 2024 were collected from the Bandung Regency Health Profile and Satu Data Kabupaten Bandung, while sub-district geometries were obtained from the 2022 administrative boundaries provided by the Geospatial Information Agency. Seven features were employed, consisting of low birth weight prevalence, unsafe sanitation, inadequate drinking water access, non-PHBS households, Posyandu ratio, Puskesmas ratio, and the logarithm of population density. All features were standardized and transformed so that higher Z-scores represented greater relative vulnerability. The K-Means algorithm was evaluated for k=2–6 clusters. The optimal number of clusters was determined by considering the elbow method, silhouette score, Davies–Bouldin index, Calinski–Harabasz index, cluster size balance, 100-seed stability analysis, and leave-one-out validation. The k=2 solution was selected, achieving a silhouette score of 0.222, ARI seed stability of 1.000, and balanced cluster sizes of 15 and 16 sub-districts. Cluster A represents a density–drinking water–service limitation pattern, whereas Cluster B reflects a low birth weight–health behavior pattern. The weighted stunting prevalence of each cluster was 7.76% and 10.19%, respectively. However, the distribution of sub-district-level stunting prevalence between clusters was not statistically significant (Mann–Whitney U=139.0; p=0.465; Cliff’s delta=0.158). These findings support the use of Clustering as an exploratory typology for designing differentiated interventions based on vulnerability patterns rather than as a definitive classification of high- and low-risk areas.