NutriKEK-AI: Development and Validation of an Explainable Hybrid AI-CDSS for Chronic Energy Deficiency Prediction in Pregnancy

Abstract

Chronic Energy Deficiency (CED) in pregnant women remains a public health issue that contributes to a high risk of maternal complications, stunting, and pregnancy and childbirth risks, particularly in resource-limited areas. This study aims to develop and validate NutriKEK-AI, an Explainable Hybrid Clinical Decision Support System that integrates machine learning, Explainable Artificial Intelligence (XAI), and expert clinical knowledge to predict the risk of CED early in pregnancy. This study employs a Design and Development Research approach that adopts the CRISP-DM framework and Agile Development. Data were collected from 214 pregnant women, 32 healthcare workers, and 7 expert validators in Lebak and Pandeglang Regencies, Banten. Predictive models were developed using several machine learning algorithms and evaluated through 10-fold cross-validation, calibration analysis, expert validation, usability testing, and user acceptance. The results showed that the XGBoost model delivered the best performance with an accuracy of 88%, an F1-score of 0.85, and an AUC of 0.91. SHAP analysis identified upper arm circumference (UAC), anemia status, weight (BB), height (TB), body mass index (BMI), daily energy intake, protein intake, household income, and history of stunting. System validation demonstrated high validity (S-CVI/Ave = 0.91), strong reliability (κ = 0.84), and excellent usability (SUS = 84.6). Field implementation improved risk detection accuracy from 71.2% to 88.4%, reduced decision-making time from 14.8 to 6.3 minutes, and increased referral accuracy to 86.7%. NutriKEK-AI has proven to be accurate, transparent, and easy to use, and has the potential to serve as an effective AI-based solution to support early detection and maternal nutrition decision-making in primary health care settings with limited resources