Study design and data collection

Data were accessed for research purposes from 20/11/2025 to 25/12/2025. This retrospective cohort study utilized clinical follow-up data collected from May 15, 2018, to August 15, 2024. The study population consisted of 210 postpartum women who had completed breastfeeding or had been breastfeeding for at least 11 months at the time of the survey. The primary outcome variable was the duration of breastfeeding, defined as a binary classification target based on a clinically significant threshold of 11 months (Target = 1 for duration ≥ 11 months; Target = 0 for duration < 11 months). While global guidelines recommend exclusive breastfeeding up to 6 months and continued breastfeeding up to 12 months, our choice of 11 months reflects the local epidemiological context in China, where a significant drop-off in breastfeeding occurs as maternity leave concludes and mothers return to the workforce. This threshold allows the model to accurately capture the predictive signatures of mothers who successfully navigate these socioeconomic barriers.

Data collection encompassed a multidimensional set of predictors, categorized into three domains: Sociodemographic factors, Obstetric and physiological history, and Lactation-specific metrics. To ensure data quality, participants with incomplete medical records were excluded.

Ethics Statement: The study was conducted in accordance with the Declaration of Helsinki, and approved by the Ethics Committee of Wuhan Children’s Hospital (protocol code 2025R147-E01 and date of approval 17 November 2025). Patient consent was waived due to the retrospective nature of this study, as the research involved the review of existing medical records and presented minimal risk to the subjects.

Explicit feature interaction engineering

We utilized the Yeo-Johnson power transformation to stabilize the variance of continuous variables. Subsequently, we explicitly constructed second-order interaction terms (Degree = 2) for all input variables using the polynomial expansion technique. This process expanded the original feature space to include combination features (e.g., Age × Milk Stasis Frequency). To prevent the “curse of dimensionality,” a Random Forest classifier was employed to evaluate the contribution of each generated interaction term. Only the top-ranked features were retained for the final modeling phase by utilizing the median importance score as a rigorous cutoff threshold to reduce noise.

Addressing data imbalance with SMOTE

Clinical datasets frequently exhibit class imbalance. To mitigate this bias, we employed the Synthetic Minority Over-sampling Technique (SMOTE) [22]. To strictly prevent data leakage, SMOTE was applied exclusively to the training folds after the train/test split. Specifically, the original training set contained 64 short-term and 104 long-term samples. After SMOTE application, the minority class was balanced to achieve 104 samples per class, while the independent test set remained completely uncontaminated.

XGBoost predictive framework

The core predictive engine used in this study was the eXtreme Gradient Boosting (XGBoost) algorithm. The dataset was partitioned into a training set (75%) and a stratified testing set (25%). We configured the XGBoost model with specific hyperparameters to prevent overfitting: learning rate (0.03), tree depth (), and regularization parameters (gamma and subsample).

Model interpretability with SHAP analysis

We implemented SHAP (SHapley Additive exPlanations) analysis to provide clinically actionable insights. We adopted a bidirectional attribution approach evaluating Global Importance and Directionality (Risk vs. Benefit), allowing for the visualization of both the strength and the nature of the interaction effects.

Baseline characteristics

Table 1 summarizes the baseline clinical characteristics of the 210 participants, stratified by their breastfeeding duration. Statistical analysis revealed significant differences in early feeding practices (P < 0.001) and milk stasis frequency (P = 0.020) between the short-term and long-term groups.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Baseline Characteristics of the Study Population (N = 210).

https://doi.org/10.1371/journal.pone.0354857.t001

Model performance and evaluation

Evaluated on the uncontaminated test set (N = 42), the XGBoost model achieved an overall Accuracy of 76.2% (95% CI: 64.3%−88.1%), as detailed in Table 2.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Detailed Performance Metrics of the XGBoost Model.

https://doi.org/10.1371/journal.pone.0354857.t002

A critical advantage of our model is its sensitivity to the positive class (Long-term Breastfeeding), achieving an outstanding Recall of 0.92 and a class-specific F1-Score of 0.83. This indicates exceptional effectiveness in identifying mothers with the potential to sustain breastfeeding. The model exhibited a Brier Score of 0.2106, indicating reliable probabilistic calibration for clinical assessment.

Fig 1 and Fig 2 visualize the discriminative power. The Confusion Matrix (Fig 1) and ROC curve (Fig 2) confirm that the model significantly outperforms random guessing, maintaining a stable balance between sensitivity and specificity.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Confusion Matrix.

The matrix shows the classification results at the optimal threshold of 0.28.

https://doi.org/10.1371/journal.pone.0354857.g001

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Receiver Operating Characteristic (ROC) Curve.

The curve indicates robust discrimination with an AUC of 0.72.

https://doi.org/10.1371/journal.pone.0354857.g002

Identification of critical interaction features

SHAP analysis was employed to interpret these features. As shown in Fig 3 and Table 3, the model successfully identified non-linear interactions.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Directional Feature Importance.

Horizontal bars represent the feature impact magnitude. Red bars indicate Positive/Promoting Factors (positively correlated with duration months), while blue bars indicate Negative/Risk Factors (correlated with early weaning). Features are ranked by absolute importance.

https://doi.org/10.1371/journal.pone.0354857.g003

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Top 10 Key Predictors and Their Impact Direction based on SHAP Analysis.

https://doi.org/10.1371/journal.pone.0354857.t003

Principal findings and interpretation of synergistic effects

This study successfully developed an interpretable XGBoost framework that predicts breastfeeding duration with an accuracy of 76.2% and a high recall of 0.92 for the long-term breastfeeding group. These results align with recent findings [15,16], demonstrating that machine learning algorithms significantly outperform traditional regression.

The analysis revealed a critical interaction between Delivery Mode and Milk Stasis Duration (Rank 5, Positive correlation). Since vaginal delivery was coded as 1 and Cesarean as 0, this positive correlation signifies that vaginal delivery acts as a protective factor against prolonged milk stasis. Conversely, the negative impact of milk stasis is disproportionately amplified in mothers who underwent Cesarean sections. This reflects a “double burden” mechanism: post-surgical pain and reduced mobility [10] compromise the mother’s physiological resilience to cope with the additional pain of milk stasis.

Furthermore, the negative interaction between Education Level and Feeding Type (Rank 6) offers a nuanced perspective. While exclusive breastfeeding promotes long-term success, our results indicate that when highly educated mothers transition to mixed feeding, they face a significantly higher risk of early cessation compared to mothers with lower education. We hypothesize that highly educated mothers in China may face greater “return-to-work” pressures and higher opportunity costs [6,7]. Once the strict exclusive breastfeeding routine is broken, they may be more decisive in fully switching to formula feeding as a pragmatic solution. Although our dataset did not directly capture specific employment timelines, this highly plausible mechanism is well-supported by broader sociological literature regarding maternal employment constraints.

Clinical implications: From general education to precision support

The high recall (0.92) suggests the model is highly effective at identifying mothers with potential for long-term breastfeeding. Clinically, this implies a need for stratified management: Targeted Pain Management for Cesarean mothers to disrupt the “Cesarean-Stasis” negative feedback loop, and Psychosocial Support for highly educated working mothers to address practical solutions for combining work and breastfeeding.

Methodological advantages

By utilizing SMOTE for class balancing [22] and explicit feature expansion, we overcame small sample limitations [19]. To ensure a fair comparison, all baseline models were trained using the exact same preprocessed feature set (including polynomial expansion and SMOTE balancing) as the proposed XGBoost model. Furthermore, the use of SHAP values provided transparency [21].

Limitations

Despite the use of SMOTE and cross-validation, the sample size remains limited and derived from a single center. Second, the “Milk Stasis Frequency” was self-reported, potentially introducing recall bias. Future studies should aim for multi-center prospective cohorts.