Explainable machine learning for predicting women’s FP status in Ethiopia

Introduction Studying the complex and nonlinear relationships among variables and women’s FP status helps women’s reproductive health and decision-making. Studies in Ethiopia do not address this issue and have largely relied on descriptive statistics and regression models. This study applied machine-learning classification models to predict women’s FP status and identify the most influential determinants across Ethiopian zones. Methods Secondary data from the 2017–2018 Performance Monitoring and Accountability survey, comprising 7,484 women aged 15–49 years, were analyzed. Six machine learning classifiers were tested and compared to predict three FP outcomes: contraceptive use, unmet needs, and neither contraceptive use nor unmet needs. Accuracy, precision, recall, macro F1-score, Cohen’s kappa, and multiclass AUC were used to measure the performance. SHAP (SHapley Additive exPlanations) was used to interpret the model. Results The random forest achieved the highest accuracy (58.6%), Kappa (32.1%), MCC (32.3%), F1 (54.2%), and AUC (72.4%) among the multi-class models. SHAP study revealed that geographic zone (mean SHAP = 0.239) was the predominant predictor, succeeded by women's aspiration for children (0.221), socioeconomic position (0.141), educational level (0.116), religious affiliation (0.113), and knowledge of contemporary family planning techniques (0.102). The attributes of service delivery have a minimal impact on prediction. SHAP dependency plots further uncovered non-linear relationships between the age of initial sexual experience and family size for women's family planning status. Conclusion The Random Forest model had the best predictive performance among the evaluated classifiers, while SHAP provided clear interpretation of the most influential predictors. The findings showed that zone, maternal, socio-demographic, and household characteristics were significantly associated with women’s FP status in Ethiopia. The SHAP results presented the value of explainable machine learning in identifying population groups and geographic areas that may benefit from targeted FP intervention