ARTIFICIAL INTELLIGENCE FOR BABY'S BIRTH RISK PREDICTION WITH LOW-BIRTH-WEIGHT (LBW)

• Main Author: Syaqra, Athaya
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/66545
• Institution: Institut Teknologi Bandung

Birth weight is an important indicator for baby’s health. Low-birth-weight (LBW) is a term for a baby's weight at birth below 2500 gram, where babies with LBW are at risk of stunted cognitive development, malnutrition, infection, the emergence of various morbidity, and mortality. According to the Indonesian Health Profile in 2020, the Infant Mortality Rate (IMR) for neonatal is mostly caused by LBW which reaches 35.15%. This shows that Indonesia needs to reduce infant mortality rate and prevent LBW births. Therefore, it is important to detect LBW as an early intervention, to optimize the condition of the mother during pregnancy, and to have their baby with normal birth weight. A common way to detect LBW is by using a heuristic approach, which is measuring the parameters of the mother's condition during pregnancy. However, this approach does not take into account the interaction between factors, whereas LBW can be caused by the influence of other related factors. In this case, LBW prevention efforts can utilize the use of artificial intelligence (AI), especially machine learning by processing data from various factors to create a risk class prediction model. This can improve accuracy for better detection and be implemented as an easy-to-use system for risk screening for LBW births. In this study, data exploration was conducted at a clinic of low-middle socioeconomic groups in Bandung, that the most important features on the LBW risk class prediction model training were maternal weight, uterine fundal height, and total income. The database used are antenatal care polyclinic (ANC), socioeconomic, and infant birth weight as labels. Through the distribution of the dataset randomly, using socioeconomic features, and performing hyperparameter tuning, the best parameters were obtained for training the model and produce a random forest accuracy of 94.51% for the dataset without oversampling, and 93.29% after oversampling with SMOTE. The last model is then deployed to a simple website page using a microframework called Flask and the Heroku online server so that it can be used by health providers, especially midwives at the clinic.