Big Data-Based Stunting Detection for AI-Driven Hospital Transformation

Abstract

Stunting remains a major public health challenge in Indonesia despite ongoing national prevention programs. Pediatric hospital records offer an opportunity to identify children who require further growth assessment. This study evaluates a hybrid Self-Organizing Map (SOM) and Long Short-Term Memory (LSTM) model using 500 pediatric visit records from 200 children at RSIA Stella Maris Medan. Each child had two to three recorded visits, and observations from earlier visits were used to predict stunting status at a subsequent visit. SOM grouped health profiles into four clusters, while LSTM modeled changes across visits. The data were split by child into a development cohort of 160 children (400 visits) and an independent test cohort of 40 children (100 visits). Preprocessing parameters and the SOM were fitted on training data only and then applied to held-out data. The SOM–LSTM model achieved a reported test accuracy of 98.00%, precision of 91.50%, recall of 90.80%, and F1-score of 91.10%. Five-fold cross-validation yielded a mean accuracy of 92.40 ± 1.14% and mean F1-score of 91.52 ± 1.10%. Visits from the same child remained in the same fold, and preprocessing, SOM, and LSTM were refitted within each training fold. On the same test cohort, the SOM–LSTM model achieved an F1-score of 91.10%, compared with 86.00% for LSTM alone. Whereas the authors’ previous SOM–LSTM study classified stunting status using anthropometric data from a single visit, this study uses repeated hospital visits to predict stunting status at a subsequent visit. The combined model may help clinicians prioritize children for further growth assessment.