Background: Low birth weight remains a primary driver of neonatal and infant mortality in Ethiopia. Machine learning models can assist early risk identification, yet clinical adoption is often limited by black box algorithms and late pregnancy predictor variables. This study aimed to develop and validate interpretable machine learning models using early pregnancy and sociodemographic features from a national survey dataset. Methods: Secondary data from the nationwide Ethiopian Demographic and Health Survey were analyzed. Predictors were restricted to features accessible during early antenatal visits. Six machine learning algorithms were trained and evaluated on an independent holdout test set: Logistic Regression, Decision Tree, Support Vector Machine, Gradient Boosting, Random Forest and Extreme Gradient Boosting (XGBoost). Imbalance was addressed using synthetic oversampling on the training set. Model explainability was established through Shapley Additive exPlanations (SHAP). Results: Out of 12876 births, 4249 (33%) were categorized as low birth weight / small birth size. XGBoost achieved superior predictive performance with an AUC-ROC of 0.947 (95% CI: 0.910-0.938) on the test set, outperforming standard logistic regression (0.8088). Key global predictive drivers identified by SHAP values included maternal anemia status, short inter pregnancy interval (< 18 months), low maternal BMI (< 18.5 kg/m^2), rural residence, lowest household wealth quintile and delayed or non-attendance of first trimester antenatal care. Conclusion: Machine learning models trained on early pregnancy and demographic features can accurately predict low birth weight risk in Ethiopia. Integrating interpretable frameworks into primary healthcare decision support tools provides a viable strategy for early risk stratification and targeted interventions in resource-limited settings.