Summary of the dataset used in this study.

The estimation of weaning duration from incremental dentine δ15N and δ13C values offers insights into health, nutrition, and demography in past populations. In this study, we developed a novel machine learning approach to estimate weaning duration using a sequence-based Long Short-Term Memory (LSTM) neural network model. The model was trained on published serial isotopic data from 279 individuals across three tooth types (M1, dM1, dM2). We enhanced predictive performance by incorporating temporal features (differences, rolling averages, slopes) and encoding tooth type to allow the model to learn type-specific patterns. The model was evaluated with RMSE, MAE, and R2 metrics using Monte Carlo dropout to quantify predictive uncertainty. As a proof of concept, we compared our model to established methods in weaning estimation (WEAN, ChangeR) using 20 individuals from the site of Thessaloniki. Our model achieved an RMSE of 0.46 years (95% CI: 0.38–0.54), a MAE of 0.36 years (95% CI: 0.29–0.44), and an R2 of 0.82 (95% CI: 0.73–0.88), indicating reliable predictive accuracy. In contrast to previous approaches that rely primarily on δ15N, our method leverages sequential patterns in both δ15N and δ13C to infer the weaning process. This study introduces the first machine learning framework specifically designed to estimate weaning duration from serial isotopic data. By modelling complex, non-linear developmental trajectories, it offers a scalable and robust tool for reconstructing weaning practices in bioarchaeological and related research studies.