Bridging the Glycemic Index (GI) Data Gap in Indonesian Foods using Provenance-Aware Machine Learning

Abstract

The glycemic index (GI) is a clinically important descriptor of carbohydrate foods, but in-vivo measurement is slow and costly, motivating predictive models based on nutritional composition. Existing models are usually validated on small single-laboratory samples and report near-perfect accuracy that is rarely tested for generalizability. We assemble a purpose-built corpus of 617 Indonesian foods using an automated ETL pipeline that fuses three nutrition databases with international and literature GI references through translation and prefix-fuzzy matching. On the valid carbohydrate subset $(\mathbf{n}=\mathbf{5 4 1})$, we faithfully replicate a published Levenberg-Marquardt artificial neural network (ANN-LM) and benchmark it against gradient-boosted trees, a kernel method, and a modern neural network under repeated cross-validation with Friedman-Nemenyi and Wilcoxon testing. Tree ensembles reach $R^{2}=0.54$, significantly outperforming the ANN-LM baseline $\left(R^{2} \approx 0.37\right)$. A provenance-aware external test training on estimated GI and testing on the in vivo-measured subset drops to $R^{2}=\mathbf{0. 1 5}$ (bootstrap 95% CI includes zero), exposing a generalization gap that single-laboratory studies cannot detect. Mutual information and SHAP analysis reveal carbohydrate, the protein/carbohydrate ratio, and net carbohydrate as dominant drivers, consistent with a principled accuracy ceiling. Our contribution is a reproducible benchmark, a provenance-aware validation protocol, and an optimal boundary-limit evaluation of what composition-based GI prediction can and cannot achieve on a heterogeneous regional corpus.