Early identification of major depression risk among reproductive-age women using hormonal contraceptives in southern Ethiopia: A Machine Learning Approach.

Machine-learning (ML) algorithms are increasingly valuable in health sciences because they can analyze complex, high-dimensional data and detect patterns that may not be easily identified using traditional statistical methods. These models can efficiently classify individuals into low- and high-risk groups, enabling early detection of conditions such as depression, cancer, and diabetes before symptoms become severe. Thus, by revealing intricate patterns in sizable health datasets, machine-learning analysis facilitates the precise and early identification of high-risk patients, facilitating prompt diagnosis, efficient screening, and well-informed clinical decision-making. Therefore, the aim of this study is to develop a predictive model for major depression in women using hormonal contraceptives in southern Ethiopia. A cross-sectional data was collected from 1002 participants over two consecutive months from eight public hospitals in Gamo Zone southern Ethiopia from August 1 to September 30, 2025. Six machine-learning models, random forest, decision tree, logistic regression, partial decision tree (PART), Naïve Bayes, and XGBoost, were employed to predict the probability of developing major depression. The dataset was split into a training set (80% of the observations) and a test set (20% of the observations). The predictive capacities of each machine-learning model were evaluated using receiver operating characteristic curves and various measures of model performance. The SHAP and Gini information values were used in selection of important attributes for predicting major depression. The predictive performance of the six machine-learning models demonstrated Cohen's Kappa values ranging from 0.185 to 0.480, indicating slight to moderate agreement beyond chance. Among the algorithms tested, the random forest model demonstrated the highest accuracy (84.4%) and precision (93.2%). Using fivefold cross-validation and SMOTE to address data imbalance, the model achieved a high F1-score (91%) and the highest AUC (0.831) on the test data, indicating good discriminatory power. Feature-importance analysis using SHAP values identified maternal occupation and younger maternal age as the key attributes associated with an increased probability of major depression among hormonal contraceptive users, whereas longer duration of method use was identified as a protective factor associated with a 16% lower risk. The Random Forest model demonstrated the best predictive performance for major depression, achieving comparatively higher accuracy, recall, precision, F1-score and AUC values than the other models evaluated. Therefore, clinicians and healthcare researchers consider using the Random Forest model as a supportive decision-making tool for early risk prediction and targeted intervention.